Administrative staff in universities, hospitals, banks, and government offices now spend a large part of their working day drafting letters, summarising minutes, cleaning spreadsheets, answering enquiries, and preparing reports. Public generative AI assistants make these tasks faster, so employees have started using them with or without permission. When such use happens outside official channels, it is called #Shadow_AI. The convenience is real, but so is the exposure: a single pasted paragraph can move student records, patient identifiers, salary tables, contract clauses, or unpublished policy drafts onto servers the organisation does not control and cannot audit. This article examines the problem and argues that prohibition alone fails. Bans push usage further underground, while the productivity pressure that created the behaviour remains untouched. Instead, the paper develops the case for building secure, #internal_tools that meet the same needs inside a controlled boundary. Using an integrative review of peer reviewed literature on shadow information technology, large language model security and privacy, and emerging AI #governance instruments, the study synthesises three contributions. First, it presents a driver model that explains why capable, well intentioned employees choose unsanctioned assistants. Second, it offers a taxonomy of #data_leak pathways specific to generative systems, covering prompt disclosure, vendor retention, model memorisation, indirect prompt injection, output oversharing, and shadow record creation. Third, it proposes a layered reference architecture for an internally hosted assistant, combining retrieval augmented generation over approved corpora, automated redaction, role based access, tenant isolation, complete logging, evaluation harnesses, and human review gates, wrapped in a governance shell aligned with recognised risk management frameworks. The paper closes with a staged implementation roadmap, a maturity model, limitations, and a research agenda. The core claim is simple: organisations reduce leakage not by fighting demand but by satisfying it safely. Keywords Shadow AI; generative artificial intelligence; data leakage; information governance; retrieval augmented generation; administrative work; large language models; organisational security 1. Introduction 1.1 The quiet arrival of a new risk Most technology risks arrive loudly. They come with a procurement cycle, a vendor contract, a security review, and a rollout plan. #Shadow_AI arrived quietly. It entered through the personal browser tab of an administrator who had forty emails to answer before lunch and discovered that a chat assistant could draft thirty of them in ten minutes. No one signed anything. No one told the information security team. The work simply got faster, and the habit spread through word of mouth in the way useful habits always do. This is not a story about careless employees. It is a story about a mismatch. The demand for speed in #administrative_tasks grew faster than the supply of sanctioned tools that could deliver that speed. Where such a gap exists, people close it themselves. Research on workplace technology has documented this pattern for decades under the heading of #shadow_IT, which describes hardware, software, and services adopted by business units without the knowledge or approval of the central technology function (Fürstenau, Rothe, and Sandner, 2021; Mallmann and Maçada, 2021). Generative assistants are the newest and most consequential expression of that old behaviour. What makes the current wave different is the nature of the interaction. A spreadsheet macro downloaded without approval is a tool that processes data locally. A public chat assistant is a conversation partner that requires the user to hand over the material before it can help. The act of asking for help is itself an act of disclosure. To summarise a grievance letter, you must supply the grievance letter. To fix a formula in a payroll sheet, you must show the payroll figures. The #prompt is the payload, and the payload frequently contains #sensitive_data. 1.2 Why administrative work is the sharpest edge of the problem Attention in the public debate has focused on software engineers using code assistants and on students using chatbots for essays. Administrative work has been treated as low stakes clerical activity. That assumption is wrong, and it is wrong in a specific way. Administrative staff sit at the exact points where an organisation concentrates its most regulated information. Registry offices hold #student_data, transcripts, disciplinary files, and disability accommodations. Human resources holds contracts, disciplinary records, medical certificates, and salary bands. Finance holds banking details, supplier terms, and audit findings. Clinical administration holds appointment records and referral letters. Legal and executive support holds draft settlements, board papers, and unpublished restructuring plans. These functions produce enormous volumes of routine text, and routine text is exactly what generative models handle best. The overlap between the tasks most suited to automation and the tasks that carry the highest #confidentiality burden is close to total. Add to this a second factor. Administrative roles are usually the least resourced in terms of technical support, security training, and access to sanctioned tooling. Engineering teams often receive enterprise licences with contractual protections. Administrative teams frequently do not, and so they reach for the free public version. The result is a predictable inversion: the group with the most regulated data operates with the fewest safeguards. 1.3 The failure of the prohibition reflex The first institutional response to unsanctioned assistant use is almost always a ban. A memo circulates. Domains are blocked at the firewall. Staff are reminded of their duties of #confidentiality. This response feels decisive and is largely ineffective, for three reasons. It does not remove the underlying pressure. The forty emails still have to be answered. Deadlines do not adjust to policy memos, so the incentive that produced the behaviour survives the ban intact. It is technically porous. Blocking a domain on the corporate network does nothing about the personal phone sitting next to the keyboard. Staff photograph a screen, retype a paragraph, or forward a document to a personal account. Enforcement pushes the activity onto devices with even less #monitoring, which makes the leak harder to detect rather than less likely to occur. It destroys visibility. Once use is forbidden, it is also hidden. Employees stop asking questions, stop reporting near misses, and stop telling anyone which tools they rely on. The organisation loses the informal intelligence that would let it manage the risk, and it loses it precisely because it declared the risk unmentionable. The shadow IT literature reports this dynamic consistently: prohibition converts a manageable, visible #workaround into an invisible one (Klotz and colleagues cited in Fürstenau, Rothe, and Sandner, 2021). 1.4 The argument of this paper This paper takes a different position. Unsanctioned assistant use is best understood as a signal of unmet demand, not as a discipline problem. The correct response is to meet the demand inside a boundary the organisation controls. Concretely, this means building or deploying an internal generative assistant that is at least as convenient as the public alternative, is connected to approved institutional knowledge through #retrieval_augmented_generation, applies automatic #redaction before any text reaches a model, enforces #role_based_access on both documents and features, logs every interaction for audit, and operates under a governance shell that assigns clear #accountability. The argument is not that internal tools are risk free. They introduce their own failure modes, including over broad retrieval permissions, #prompt_injection through poisoned documents, #hallucination in official correspondence, and the creation of unmanaged records. The claim is narrower and, I think, defensible: these are risks the organisation can see, measure, and control, whereas the risks of shadow usage are risks it cannot see at all. Replacing invisible exposure with governed exposure is the achievable goal. 1.5 Research questions The study addresses four questions. RQ1. What organisational, individual, and technological factors drive administrative staff towards unsanctioned generative AI use? RQ2. Through which distinct pathways does confidential information escape when such tools are used in administrative work? RQ3. What technical and architectural components are required for an internal assistant to serve administrative needs while preventing those escapes? RQ4. What governance structures, controls, and change practices allow such a tool to be adopted, trusted, and sustained? 1.6 Contribution and structure The paper contributes an integrative synthesis rather than new empirical data. It draws together three literatures that have so far developed in isolation: the organisational research on shadow IT and #policy_compliance, the technical research on large language model #privacy and security, and the regulatory literature on AI #risk_management. Section 2 reviews these strands. Section 3 sets out a conceptual framework. Section 4 explains the method. Section 5 answers RQ1 and RQ2 through a driver model and a leakage taxonomy. Section 6 answers RQ3 with a layered reference architecture. Section 7 answers RQ4 with a governance model, maturity ladder, and roadmap. Section 8 discusses implications, Section 9 states limitations, Section 10 proposes future work, and Section 11 concludes. 2. Literature Review 2.1 Shadow information technology as the parent phenomenon The concept of shadow IT predates generative models by many years and supplies the analytical vocabulary this paper needs. It refers to systems, devices, and services procured or built by business units outside the formal authority of the central IT function. Early work framed it as deviance to be eliminated. More recent scholarship treats it as an ambivalent phenomenon with both costs and benefits. Fürstenau, Rothe, and Sandner (2021) examine what happens after shadow systems are discovered and find that outcomes vary widely. Some are shut down, some are absorbed into the official landscape, and some are formally sanctioned with little change. Their configurational analysis shows that the fate of a shadow system depends on how central it has become to the work and how well it performs, not simply on whether it violated a rule. This is a direct lesson for #Shadow_AI: the useful ones will not disappear on command, and the sensible move is to plan for absorption. Mallmann and Maçada (2021) study the individual level and find that shadow IT use is often associated with perceived improvements in individual performance. The user is not sabotaging the organisation. The user is trying to do the job. Any intervention that treats these employees as adversaries will misdiagnose the situation and generate resistance. The compliance literature adds a further nuance. Jaeger and Eckhardt (2021) show that situational security awareness, meaning awareness activated at the moment of a specific decision, predicts secure behaviour far better than general awareness gained in an annual training module. This finding has sharp design implications. A yearly slide deck about #data_classification will not stop a staff member pasting a payroll table into a browser at four in the afternoon. A warning that appears in the moment, in the interface, at the point of the risky action, might. 2.2 Generative AI in administrative and knowledge work A body of empirical work now documents the productivity effects that make these tools irresistible. Noy and Zhang (2023) ran a controlled experiment on mid level professional writing tasks, including press releases, short reports, and delicate emails, and found substantial reductions in time taken together with gains in rated quality, with the largest benefits accruing to weaker writers. Brynjolfsson, Li, and Raymond (2025) studied a large customer support deployment and found similar patterns, again concentrated among less experienced staff. Dell'Acqua and colleagues (2023) reported gains for consultants on tasks inside the model's competence and, importantly, degraded performance on tasks just outside it, a result they describe as a jagged frontier. Three points follow for the present argument. The #productivity effect is real, so demand will not fade. The effect is largest for exactly the population that staffs administrative functions, namely people producing large volumes of standard professional text under time pressure. And the effect reverses when the task falls outside the model's reliable range, which means uncontrolled use produces not only leakage risk but also quality risk. 2.3 Privacy and security properties of large language models The technical literature identifies several mechanisms by which information entrusted to a model can escape. Memorisation and extraction. Carlini and colleagues (2021) demonstrated that training data can be recovered verbatim from a deployed language model by an adversary who does not have access to the training set. Follow up work quantified the relationship, showing that memorisation grows with model size, with the number of times a sequence appears, and with the amount of context available to the attacker (Carlini and colleagues, 2023). The practical meaning is direct: text submitted to a service that uses conversations for training may resurface. This is #model_memorization, and it is not hypothetical. Prompt injection. Greshake and colleagues (2023) showed that instructions hidden inside retrieved content can hijack an application built on a language model, causing it to ignore its original task and follow an attacker's instructions instead. Because the model cannot reliably distinguish data from instructions, any system that reads untrusted documents inherits this weakness. For administrative assistants that ingest incoming email, uploaded PDFs, or supplier attachments, indirect #prompt_injection is a first order threat, not an exotic one. Jailbreaks and guardrail evasion. Systematic studies of chatbot safety filters show that carefully constructed prompts can bypass them at meaningful rates, and that automated methods can generate such prompts at scale (Deng and colleagues, 2024). A #content_filter is a speed bump, not a wall. Comprehensive surveys. Yao and colleagues (2024) provide a structured account of the security and privacy landscape, covering both offensive and defensive uses. Das, Amini, and Wu (2025) offer a broader survey of security and privacy challenges. Neel and Chang (2023) review privacy issues specifically, including membership inference, extraction, and the limits of proposed mitigations. Together these confirm that no single control is sufficient and that layered defence is the only workable posture. Hallucination. Huang and colleagues (2025) survey the causes, detection, and mitigation of fabricated output. In administrative correspondence, a fabricated regulation number, a wrong deadline, or an invented precedent can create legal exposure and damage trust as effectively as a leak. Any design must therefore address accuracy alongside #privacy. 2.4 Governance frameworks and regulation Institutional responses now have formal anchors. The NIST AI Risk Management Framework (National Institute of Standards and Technology, 2023) organises practice around four functions: govern, map, measure, and manage. It is voluntary, sector neutral, and well suited to being used as the backbone of an internal control set. ISO/IEC 42001 (2023) provides a certifiable management system standard for AI, which matters for organisations that already operate an information security management system and want an auditable structure they recognise. In the European Union, the AI Act (Regulation 2024/1689) creates a risk tiered regime with obligations that scale with the potential for harm, and it introduces specific duties for general purpose models. Legal analyses have examined its structure, its definitions, and its likely effects (Veale and Zuiderveen Borgesius, 2021; Hacker, Engel, and Mauer, 2023; Novelli and colleagues, 2024). Data protection law adds a separate and immediately binding layer. Where personal data is pasted into a third party service without a lawful basis, without a processing agreement, and often without knowledge of where the servers sit, the organisation is exposed regardless of what any AI specific statute says. #Compliance failure in shadow usage is therefore not a future problem waiting on new legislation. It is a present problem under existing rules. Auditing offers a practical bridge between principle and practice. Mökander and colleagues (2023) propose a three layered approach covering governance audits, model audits, and application audits, which maps cleanly onto the internal tool proposal developed later in this paper. 2.5 The gap this paper addresses Each literature is strong on its own terms and incomplete on its own. The shadow IT work explains motivation but says nothing about vector databases or redaction pipelines. The machine learning security work characterises attacks with precision but rarely asks why an administrator in a registry office behaves as she does. The governance work sets out principles at a level of abstraction that offers little help to a team deciding whether to run an #open_weight_model on premises or buy a private endpoint. What is missing is an integrated account that moves from human motive, through leakage mechanism, to architectural control, to governance structure, in a single line of reasoning. That is what follows. 3. Conceptual Framework 3.1 Three lenses The framework combines three established perspectives. The technology, organisation, and environment perspective directs attention to adoption as a joint product of the tool's characteristics, the organisation's readiness, and external pressures such as regulation and peer behaviour. It explains why two similar institutions can face very different levels of shadow usage. Workaround theory treats deviation from prescribed process as a rational response to a blockage. The employee encounters an obstacle, perceives a way around it, weighs the perceived cost, and acts. Shadow assistant use fits this structure exactly. The blockage is time pressure and the absence of a sanctioned tool. The workaround is the public chatbot. The perceived cost is low because the risk is invisible and the sanction is improbable. Protection motivation reasoning explains the perceived cost side. People act protectively when they believe a threat is severe, that it is likely to affect them, that a protective response works, and that they are able to perform it. In shadow assistant use, every one of these beliefs is typically weak. The threat feels abstract, personal likelihood feels negligible, the protective response is unclear, and self efficacy in secure alternatives is absent because no alternative exists. 3.2 The Shadow AI Exposure Model Combining the lenses yields a simple model with four stages. Stage one, pressure. Volume of routine text work, deadline density, and staffing shortfalls create demand for acceleration. Stage two, gap. No sanctioned tool exists, or one exists but is slow, poorly integrated, restricted to senior grades, or simply unknown to the person who needs it. Stage three, substitution. The employee adopts a public assistant. Adoption is reinforced by visible time savings and by peer example, and it is unopposed by any perceived cost. Stage four, exposure. Confidential material crosses the organisational boundary through one or more of the pathways set out in Section 5.2, and the organisation loses both control and knowledge of the transfer. The intervention point is stage two. Controls applied at stage three are enforcement measures, and enforcement without substitution produces displacement rather than reduction. Closing the gap is the only move that addresses the mechanism rather than the symptom. 3.3 Design implications derived from the model The model generates four design requirements that guide the architecture in Section 6. Parity of convenience. If the internal tool is slower or more awkward than the public one, the gap remains open and substitution continues. Convenience is a security control, not a luxury. Situational guidance. Following Jaeger and Eckhardt (2021), warnings and classifications must appear at the moment of action inside the interface, not in an annual course. Default safety. Protection must not depend on the user making the right choice under time pressure. #Redaction, access limits, and logging should operate whether or not the user thinks about them. Visible accountability. Users behave more carefully when they know actions are attributable, and organisations can only learn from #audit_log data they actually collect. 4. Methodology 4.1 Design This is a conceptual paper built on an integrative literature review. The aim is theory building and design synthesis rather than hypothesis testing. No human participants were involved and no primary data were collected, so no empirical claims are made about the prevalence of shadow usage in any named institution. Where the paper refers to frequency or behaviour, it relies on published studies cited in place. An integrative review was chosen because the relevant knowledge is fragmented across disciplines that use different vocabularies and publish in different venues. The method is designed for exactly this situation, allowing the reviewer to combine diverse source types and to produce a new conceptual structure rather than a statistical summary. 4.2 Source identification Sources were sought across four bodies of work: shadow IT and information systems #policy compliance; empirical evaluations of generative AI in knowledge work; technical studies of language model security, privacy, and reliability; and AI governance instruments and legal analysis. Peer reviewed journal articles, papers from recognised conferences, standards documents, and scholarly books were considered. Preference was given to work published from 2021 onwards, reflecting the pace of change in the field, with the sole exceptions being foundational technical results that later work builds on directly. 4.3 Inclusion and exclusion Included: peer reviewed empirical studies, systematic surveys, formal standards, regulatory texts, and scholarly monographs that address at least one of the four bodies of work above. Excluded: vendor marketing material, blog posts, undated web pages, and news reporting used as primary evidence. Where widely reported industrial incidents are mentioned in the text, they are treated as illustration of a mechanism established in the peer reviewed literature, not as evidence in their own right. 4.4 Synthesis procedure Selected sources were coded thematically. Codes were grouped into three families: causes, meaning why staff adopt unsanctioned tools; mechanisms, meaning how data escapes; and controls, meaning what reduces the escape. The families were then arranged as a causal chain, and the control family was mapped against the mechanism family to identify which mechanisms are addressed by which controls and, importantly, which are addressed by none. The reference architecture in Section 6 is the direct output of that mapping. 4.5 Validity considerations Three limits on validity should be stated openly. Selection is interpretive rather than exhaustive, so a systematic review with formal protocol registration would be a valuable complement. The field moves quickly, so specific technical claims about model behaviour carry a shelf life. And the proposed architecture is argued rather than tested, which places it in the category of design proposition awaiting #evaluation. Section 9 returns to these points. 5. Findings I: Drivers and Leakage Pathways 5.1 Why administrative staff turn to unsanctioned assistants (RQ1) The synthesis identifies seven drivers, grouped into three families. 5.1.1 Task and workload drivers Volume and repetition. Administrative work is dominated by tasks with stable structure and variable content: acknowledgement letters, meeting minutes, standard enquiries, data entry checks, translation of routine notices. These are precisely the tasks where a language model performs well, and the match is obvious to anyone who tries it once. Deadline compression. Most administrative work is downstream of someone else's decision, which means the time available is set externally and is usually short. When a policy adds friction to a deadline bound task, the deadline wins. Emotional labour. A large share of administrative correspondence is difficult: refusals, complaints, disciplinary notices, chasing payments. Staff report that drafting these is disproportionately draining. An assistant that produces a serviceable first draft removes not only minutes but also dread, and that reward is powerful. 5.1.2 Organisational drivers Tool absence. The single strongest driver is the absence of a sanctioned alternative. Where nothing is provided, the public option is not a choice among several, it is the only option. This is the gap identified at stage two of the model. Tool inadequacy. Where an internal tool exists but is slow, hard to reach, disconnected from institutional documents, restricted to certain grades, or requires a separate login and a request form, staff treat it as absent. Perceived #usability determines behaviour more reliably than official status does. Unclear policy. Many institutions have issued statements that neither permit nor forbid clearly. Staff reading such statements conclude, not unreasonably, that nobody has decided, and they carry on. Ambiguity is read as permission. 5.1.3 Individual drivers Risk invisibility. The person pasting a document sees no consequence. No alarm sounds, no data visibly leaves, the answer appears, the work is done. Where feedback is absent, the perceived probability of harm collapses towards zero, which is exactly the condition under which protection motivation fails. The gap between actual and perceived risk is the psychological core of #Shadow_AI. Taken together, the drivers show that unsanctioned use is neither irrational nor malicious. It is the predictable output of a system that demands speed, supplies no safe means of achieving it, communicates ambiguously, and provides no feedback when a boundary is crossed. 5.2 A taxonomy of leakage pathways (RQ2) Leakage is not one event. It is at least seven distinct mechanisms, each requiring different controls. Confusing them is the most common analytical error in institutional policy documents, which typically address only the first. 5.2.1 Pathway one: direct prompt disclosure The user pastes confidential content into the prompt. This is the obvious pathway and the one every policy mentions. It covers pasted spreadsheets, uploaded documents, quoted emails, and typed summaries containing names, identifiers, financial figures, or health details. It is high volume and, in principle, the most tractable, because it can be intercepted at the interface. 5.2.2 Pathway two: vendor retention and secondary use Once submitted, content is stored by the provider. Depending on the plan, the jurisdiction, and the terms in force, it may be retained for a defined period, reviewed by humans for quality or safety purposes, or used to improve models. Consumer tiers historically offer weaker protections than enterprise agreements. The organisation typically has no #vendor_risk assessment on file for a service it never knew was in use, no data processing agreement, and no knowledge of #data_residency. The exposure here is contractual and regulatory as much as technical. 5.2.3 Pathway three: memorisation and extraction Where submitted content contributes to training, it may be reproduced later. Carlini and colleagues (2021, 2023) established both that extraction is feasible and that it scales with model size and repetition. Rare, distinctive strings, exactly the kind that identify a specific person, a specific case number, or a specific contract clause, are especially vulnerable. The exposure is delayed, meaning the harm may surface long after the submission that caused it, and attribution at that point is close to impossible. 5.2.4 Pathway four: account and session compromise Shadow use is typically conducted through personal accounts with personal passwords, frequently without multifactor authentication, sometimes shared between colleagues, and never covered by corporate identity management. A compromised personal account exposes a conversation history that may constitute a searchable archive of the organisation's most sensitive text. When the employee leaves, the archive leaves with them, and no offboarding process touches it. 5.2.5 Pathway five: indirect prompt injection Once an assistant is allowed to read documents, whether pasted or retrieved, it becomes vulnerable to instructions concealed within those documents. Greshake and colleagues (2023) demonstrated the mechanism. A malicious instruction in an incoming attachment, a supplier invoice, an application form, or a web page can redirect the model to reveal earlier context, alter its output, or take an action on the attacker's behalf. Administrative inboxes receive untrusted documents from the public by design, which makes this pathway structurally severe for exactly the functions this paper concerns. 5.2.6 Pathway six: output oversharing and fabrication Leakage also runs outward. An assistant with access to internal documents may surface, in an answer to one member of staff, content the person is not entitled to see, such as another employee's disciplinary file or an unreleased budget. This is not an exotic attack. It is the default outcome when retrieval permissions are set more broadly than document permissions. Alongside it sits fabrication: a confident but false statement in an official letter can create a commitment the institution did not intend to make (Huang and colleagues, 2025). 5.2.7 Pathway seven: shadow record creation Every conversation with an assistant creates a record. Draft correspondence, deliberations about a case, and summaries of confidential meetings accumulate in a chat history held by a third party. These records sit outside the organisation's #records_management regime. They will not be found by a records search, cannot be produced in response to a lawful access request, and will not be deleted under a retention schedule. In regulated sectors this is a compliance failure independent of whether any external party ever reads the content. 5.3 Mapping controls to pathways The value of the taxonomy is that it exposes which controls do what. Blocking a domain addresses pathway one weakly and pathways two through seven not at all. Staff training addresses pathway one modestly and the rest barely. A contractual enterprise agreement addresses pathways two and three and does nothing for five, six, or seven. Only an internally controlled tool, correctly designed, addresses all seven, and even then it does so only if the design is deliberate about each one. Table 1 summarises the mapping. Table 1. Leakage pathways and the controls that address them Pathway Nature of exposure Primary control Direct prompt disclosure Confidential text sent outside the boundary Internal endpoint, automated redaction, in interface classification prompts Vendor retention Third party storage and secondary use Self hosting or contracted zero retention endpoint, processing agreement, data residency guarantees Memorisation and extraction Delayed reproduction of submitted content No training on institutional prompts, isolation of tenant data, restricted fine tuning corpora Account compromise Unmanaged personal accounts and histories Corporate identity, single sign on, multifactor authentication, offboarding integration Indirect prompt injection Hidden instructions in ingested documents Content sanitisation, provenance labelling, privilege separation, restricted tool actions, human approval gates Output oversharing and fabrication Retrieval crossing entitlement lines, false statements Permission aware retrieval, citation of sources, mandatory human review before dispatch Shadow record creation Records outside the retention regime Internal logging, retention schedules, legal hold capability, records classification 6. Findings II: A Reference Architecture for a Secure Internal Assistant (RQ3) 6.1 Design objectives The architecture must satisfy the four requirements derived in Section 3.3, namely parity of convenience, situational guidance, default safety, and visible accountability, while closing all seven pathways. It is presented as seven layers. The layering matters: each layer assumes the one below it and no single layer is sufficient alone. 6.2 Layer one: the model and its hosting The first decision is where inference happens, and it determines almost everything downstream. Self hosted #open_weight_models. Weights are downloaded and run on infrastructure the organisation controls, whether #on_premise or in a private cloud subscription. Prompts never leave the boundary, which closes pathways two and three by construction. The costs are real: hardware, specialist staff, and an ongoing obligation to update. For institutions with strong internal capability or with data that cannot lawfully leave a jurisdiction, this is the strongest option. Contracted private endpoints. A commercial model is accessed through an agreement that specifies no training on submitted content, defined retention, and a stated processing region. This lowers the operational burden considerably and is a reasonable posture for many organisations, but it converts a technical guarantee into a contractual one. The protection is only as good as the agreement, the audit rights attached to it, and the willingness of the organisation to verify. The public consumer tier. Not acceptable for administrative work involving personal or confidential data, for the reasons set out in Section 5.2. A practical note on model size. #Small_language_models, in the range that runs on modest hardware, are frequently sufficient for the tasks that dominate administrative work: summarising, reformatting, drafting standard correspondence, extracting fields, and translating routine notices. Institutions often assume they must match the largest frontier system to be useful, and then abandon the project on cost grounds. For the actual task profile, that assumption is usually wrong. Matching model capability to task requirement is the single most effective way to bring an internal deployment within budget. Adaptation should be approached with care. Parameter efficient methods such as #LoRA and its quantised variants make it feasible to adapt an open model to institutional tone and format at modest cost (Hu and colleagues, 2022; Dettmers and colleagues, 2023). But #fine_tuning bakes the training corpus into the weights, and anything in that corpus can in principle be surfaced later. The rule that follows is strict: never fine tune on raw records containing personal data. Adapt on style, structure, terminology, and de-identified examples. Where the goal is to give the assistant knowledge of institutional facts, use retrieval instead, which is the subject of the next layer. 6.3 Layer two: knowledge through retrieval #Retrieval_augmented_generation attaches the model to an approved document corpus at query time. The user asks a question, the system searches the corpus, passes the most relevant passages to the model as context, and the model answers from those passages (Lewis and colleagues, 2020; Gao and colleagues, 2023). The security advantages are substantial and are often overlooked in favour of the accuracy advantages. Institutional knowledge stays in a governed store rather than inside model weights. Documents can be added, corrected, and, crucially, removed, and the removal takes effect immediately, which is not true of anything absorbed into a trained model. Answers can cite the passage they came from, which makes verification possible and reduces the space for #hallucination. And access can be enforced per document, which is the foundation of the next layer. Three design cautions apply. First, chunking and indexing decisions determine what the system can actually find, and a poorly tuned retriever produces confident answers from irrelevant passages. Second, the vector store is a copy of institutional content and must be protected accordingly, with #encryption at rest and controlled administrative access. Third, and most importantly, retrieved passages are untrusted input from the model's point of view, which brings pathway five directly into the architecture. Retrieval solves knowledge and creates injection surface at the same time. Layer five exists to handle that. 6.4 Layer three: identity, entitlement, and isolation This layer answers the question of who is allowed to see what, and it is where most internal deployments fail. Access must run through corporate identity, with single sign on and multifactor authentication, so that every interaction is attributable to a real person in a real role, and so that offboarding removes access on the day it should. Personal accounts have no place in the design. Entitlement must be enforced at retrieval, not after it. The system must filter candidate documents by the requesting user's permissions before passing anything to the model, because a passage that reaches the context window has already leaked, whatever the interface subsequently displays. This is a simple rule that is easy to violate in a hurried implementation, and violating it produces pathway six at scale. Permissions should follow the principle of #least_privilege and be expressed through #role_based_access aligned to actual job function: a finance officer sees finance documents, a registry clerk sees registry documents, and neither sees the other's. Where the assistant serves multiple departments, #tenant_isolation should be strict, so that indexes, logs, and caches do not mingle. Underlying all of this, the organisation needs a working #data_classification scheme, because entitlement rules cannot be written for documents nobody has classified. Institutions that have neglected classification for years will discover that an internal assistant project forces the issue, which is uncomfortable and, on balance, healthy. 6.5 Layer four: the protective pipeline around the prompt Between the user and the model sits a pipeline that applies protection by default, so that safety does not depend on a tired person making a good decision at four in the afternoon. Automatic #de_identification. Before text reaches the model, a detection stage identifies personal identifiers such as names, national identity numbers, student and patient numbers, contact details, bank details, and dates of birth, and replaces them with placeholders. The model works on the placeholder version. After generation, the placeholders are restored in the output shown to the authorised user. The model never sees the identifier, and the log never stores it in the clear. This technique is not perfect, since detection misses unusual formats and free text can identify a person without naming them, but it removes a very large share of the most sensitive material at near zero cost to the user, which is exactly what a default control should do. Classification prompts at the point of action. Where the pipeline detects material that appears highly sensitive, the interface should say so, in plain words, before submission, and offer the safe path. This is the operational form of Jaeger and Eckhardt's (2021) situational awareness finding. The message should be specific, brief, and actionable, and it should not be so frequent that people learn to dismiss it without reading, which is the standard failure mode of warning design. Input sanitisation. Uploaded and retrieved content should be stripped of hidden text, invisible characters, and embedded instruction patterns before it enters the context window. This is a partial defence against injection, not a complete one, but it removes the easiest attacks. Output review. The pipeline should scan generated text for identifiers that should not appear, for unsupported factual claims where citation is expected, and for content that crosses the entitlement line. Anything destined to leave the organisation, meaning any external letter, notice, or decision, should pass through explicit #human_oversight before dispatch. The assistant drafts; a person signs. This principle should be non negotiable, and it should be enforced by the workflow rather than left to individual conscience. 6.6 Layer five: adversarial resilience The threat model must assume that some documents entering the system are hostile, because in administrative work some documents entering the system are hostile. Privilege separation. The component that reads untrusted content should not be the component that holds sensitive context or that can invoke actions. Where an assistant can perform operations, such as sending mail, updating a record, or scheduling a payment, those operations must be gated behind explicit human confirmation, and the confirmation must show the actual effect in plain language rather than a summary the model wrote about itself. Provenance labelling. Content should carry a marker indicating whether it originated from a trusted internal source or an untrusted external one, and the system's instructions should treat instructions found inside external content as data, never as commands. Structured #red_teaming. Before launch and at regular intervals afterwards, a dedicated exercise should attempt injection through uploaded documents, attempt extraction of another user's data, attempt to induce fabricated authority, and attempt #jailbreak of the guardrails. Findings should be logged, tracked, and fixed like any other defect. Deng and colleagues (2024) show that automated attack generation is available to adversaries, so defenders who test only manually are testing at the wrong scale. Guardrails as one layer among many. Filters and refusal training reduce the rate of unwanted output. They do not eliminate it, and an architecture that relies on them alone is fragile. #Guardrails belong in the design; they do not constitute the design. 6.7 Layer six: logging, records, and audit Every interaction should be logged: who asked, when, which documents were retrieved, what the model produced, and what the user did with the output. The #audit_log must itself be protected, since it contains a distilled record of sensitive activity, and access to it should be restricted and monitored. The logs then serve four purposes. They allow #incident_response, because when something goes wrong the organisation can reconstruct what happened. They allow demonstration of compliance to an internal auditor or an external regulator. They provide the raw material for improvement, since patterns of failed retrieval and abandoned sessions show precisely where the tool is inadequate. And they bring the assistant inside the #records_management regime, with defined retention, deletion schedules, and the capacity to apply a legal hold. This last point is what converts pathway seven from an unmanaged liability into a managed one. 6.8 Layer seven: evaluation A tool that is not measured cannot be trusted, and an internal assistant deployed without an #evaluation harness is an act of faith. The harness should be built before launch, not after, and should cover four dimensions. Accuracy, tested against a fixed set of representative administrative questions with known correct answers drawn from the institution's own policies. Safety, tested against a suite of adversarial prompts including injection attempts and extraction attempts. Leakage, tested by having accounts with limited entitlements attempt to retrieve documents they should not see. And utility, measured through task completion time, revision effort, and user satisfaction, because a tool that is safe and useless will be abandoned, and abandonment returns the organisation to stage two of the exposure model. Results should be tracked over time, since model updates, index changes, and new document types all shift behaviour. Chen, Zaharia, and Zou (2024) document that model behaviour drifts across versions, which means an evaluation performed once is an evaluation that has already expired. 7. Findings III: Governance, Adoption, and Sustainment (RQ4) 7.1 Why architecture alone is not enough A well built assistant that nobody trusts, nobody uses, and nobody owns will fail as completely as no assistant at all, and it will fail more expensively. The technical layers in Section 6 must sit inside a governance shell that assigns responsibility, sets rules, prepares people, and maintains the system over time. 7.2 Governance structure Ownership. A single named executive owner should be accountable for the assistant, and that owner should sit close to the business, not only inside the technology function. Where AI governance is delegated entirely to security, the tool tends to be safe and unusable. Where it is delegated entirely to operations, it tends to be usable and unsafe. The owner needs both concerns in one office. A standing review body. A small cross functional group, drawn from information security, data protection, records management, legal, human resources, and the administrative functions that actually use the tool, should meet regularly to approve new use cases, review incidents, examine evaluation results, and decide what the assistant may and may not be used for. Its decisions should be written down and published internally, because ambiguity, as Section 5.1 showed, is read as permission. Alignment to a framework. Rather than inventing controls from scratch, the organisation should map its practice to an existing structure. The #NIST_AI_RMF (National Institute of Standards and Technology, 2023) provides the govern, map, measure, and manage functions that this paper's architecture already reflects. #ISO_42001 (2023) provides an auditable management system for institutions that want certification. Where the #EU_AI_Act applies, its obligations should be traced to specific controls, and the tracing should be documented in advance of enforcement rather than after an inspection. The layered audit model proposed by Mökander and colleagues (2023) provides a workable way to structure assurance across governance, model, and application levels. Documentation. Each deployed configuration should have a written record: which model, hosted where, trained or adapted on what, retrieving from which corpora, with which entitlement rules, logging what, retained for how long, evaluated against which suite, and reviewed by whom. This is unglamorous and it is the difference between a defensible position and an indefensible one. 7.3 Policy that people can actually follow Most AI policies fail because they are written as legal shields rather than as instructions. A usable policy answers four questions in language a busy administrator can act on. What may I put into the internal tool? Answer with reference to the classification scheme, with concrete examples from the reader's own work, not with abstract categories. What may I never put anywhere? A short, memorable list. Special category health data, safeguarding records, unpublished examination material, and legally privileged advice are typical entries, and the list should be short enough to remember. What must I check before I send? A brief, mandatory verification step for anything leaving the institution, covering facts, figures, names, dates, and tone. What do I do if something goes wrong? A named route, with an explicit assurance that reporting a mistake in good faith will not be punished. Without that assurance the organisation will not hear about incidents, and an organisation that does not hear about incidents cannot learn from them. 7.4 Change management and the adoption problem The behavioural evidence has a clear implication: adoption of the safe tool must be easier than adoption of the unsafe one, or the unsafe one wins. Integration over destination. The assistant should appear where the work happens, inside the mail client, the document editor, and the case management system, rather than as a separate site the user must remember to visit. Every extra click is a subsidy paid to the public alternative. Start with the painful tasks. Pilot the tool on the work staff most dislike: long minutes, repetitive acknowledgements, first drafts of difficult letters. Early value creates advocates, and advocates carry a rollout further than mandates do. Train on judgement, not on syntax. Prompt technique is learned in an afternoon. What administrative staff need is the ability to recognise which tasks the assistant is reliable for and which it is not, which is exactly the jagged frontier problem identified by Dell'Acqua and colleagues (2023). #Digital_literacy of this kind is the durable skill. Close the loop. Give users a visible way to report a bad answer, and act on the reports. A #feedback_loop that produces observable improvement converts users into participants and produces exactly the information the organisation was losing under prohibition. Speak honestly about jobs. Staff will suspect that the assistant is an argument for reducing their number, and if the institution refuses to address that suspicion, quiet resistance will follow. The honest position is usually that the tool removes the least interesting parts of the role. If that is not the position, saying so plainly is still better than allowing rumour to fill the silence. 7.5 A maturity model Institutions do not move from shadow chaos to governed practice in a single step. Five stages describe the realistic path. Stage zero, unaware. Shadow usage exists and is neither measured nor discussed. Risk is high and entirely invisible. Stage one, prohibitive. A ban is announced. Visible usage falls, actual usage does not, and reporting stops. Risk is high and now harder to see than before. Stage two, permissive with contract. An enterprise agreement is signed with an external provider. Pathways two and three are reduced. Pathways five, six, and seven remain, and internal knowledge is still not connected. Stage three, internal and governed. An internal assistant with retrieval, redaction, entitlement, and logging is deployed to a defined set of use cases under a named owner. All seven pathways are addressed, though not perfectly, and the organisation can for the first time see what it is doing. Stage four, optimised. Evaluation is continuous, adversarial testing is routine, use cases expand on the basis of measured outcomes, and the assistant is treated as ordinary infrastructure with an ordinary lifecycle. The important structural point is that stage one is not on the path to stage three. It is a detour that consumes political capital and produces no risk reduction, and organisations that pass through it usually arrive at stage three later and with less trust to spend. 7.6 A staged roadmap Phase one, discovery, roughly one to two months. Establish, through anonymous survey and amnesty rather than through disciplinary investigation, what tools staff use and for what. Anonymity and amnesty are not concessions, they are the only way to obtain true answers. Map the administrative task profile and identify the tasks with high volume and low sensitivity, which are where the pilot should begin. Phase two, foundation, two to four months. Fix what the assistant will depend on: classify the document estate, clean the permission model, and choose the hosting posture. Most of the difficulty of an internal assistant project is not the model. It is the discovery that nobody knows who is allowed to see which folder. Phase three, pilot, three months. Deploy to one department with a narrow, well chosen use case set. Build the evaluation harness before, not after. Measure time saved, error rate, and satisfaction. Fix the retrieval quality, which will be worse than expected. Phase four, expansion, six to twelve months. Add departments and use cases in order of measured value and controlled risk. Integrate into the systems where work actually happens. Formalise the review body and the incident route. Phase five, sustainment, continuous. Re evaluate after every model change. Red team on a schedule. Retire use cases that do not earn their keep. Watch for the return of shadow usage, which is the clearest signal available that the internal tool has stopped being good enough. 8. Discussion 8.1 The central reframing The argument of this paper reduces to a single reframing. #Shadow_AI is not primarily a compliance failure by employees. It is primarily a service failure by institutions. The employee who pastes a letter into a public assistant has correctly identified that a tool exists which would make the work better and faster, and has found that the institution does not provide it. Punishing that judgement does not improve it. Providing the tool does. This framing has a cost, and it should be acknowledged. It asks institutions to spend money and effort in response to behaviour they consider a violation, which feels to many managers like rewarding the offence. The answer is that the alternative is not compliance, it is concealment, and concealment is the more expensive outcome. The choice is not between risk and no risk. It is between risk you can see and risk you cannot. 8.2 What internal tools do and do not solve Honesty about limits matters, because overselling produces the disappointment that kills programmes. Internal deployment substantially reduces exposure through pathways one, two, three, four, and seven. Prompts stay inside the boundary, retention is under institutional control, submitted content does not train a foreign model, identity is managed, and conversations become governed records. Internal deployment does not by itself solve pathways five and six, and in some respects it sharpens them. Connecting an assistant to institutional documents creates injection surface where none existed, and it creates the possibility of retrieval crossing entitlement lines at machine speed. An internal assistant with sloppy permissions can leak more internal information more efficiently than any public chatbot ever could, because the public chatbot never had access to the file share. This is the honest counterweight to the paper's main recommendation, and it is the reason Sections 6.4 and 6.6 insist on permission aware retrieval and adversarial testing as non optional. Nor does internal deployment resolve accuracy. A locally hosted model fabricates as readily as a remote one. #Human_oversight before dispatch remains necessary, and any design that quietly removes it in the name of efficiency has traded a visible risk for an invisible one. 8.3 Implications for practice For institutional leaders. Treat unsanctioned usage as intelligence about unmet need. Fund the alternative before, or at the same time as, announcing the restriction. A restriction with no alternative is an instruction to hide. For information security teams. Move from gatekeeper to enabler. The security function's leverage lies in shaping the safe path, not in policing the unsafe one, because policing at this scale does not work. For data protection officers. The exposure created by shadow usage is live under existing law and does not wait for AI specific statutes. A processing activity that nobody documented, using a processor nobody assessed, in a location nobody identified, is a present failure. For administrative staff and their managers. The demand for speed is legitimate. The right response to it is to ask for a safe tool loudly and repeatedly, and to report near misses without fear, since a reported near miss is worth more to the institution than a hundred pages of policy. 8.4 Sector considerations In higher education, the concentration of #student_data in registry and admissions functions, combined with chronic administrative understaffing, produces a high risk profile with low control maturity. Assessment related material adds a second, distinct exposure, since leaked examination content damages academic integrity in a way that no data protection remedy repairs. In healthcare administration, appointment letters, referrals, and correspondence carry special category data, and the legal consequences of exposure are correspondingly severe. The redaction layer is not an optimisation here. It is a precondition. In finance and the public sector, the difficulty is often less about personal data and more about market sensitive or politically sensitive drafts, where premature disclosure of a document that is merely under consideration can cause harm even though it contains no personal identifiers at all. Classification schemes built purely around personal data will miss this category entirely, which is a common and costly design error. 8.5 The economics Institutions frequently abandon internal assistant projects on cost grounds after benchmarking against the largest available frontier model. This is a category error. The relevant comparison is not the best model available, it is the smallest model that performs the actual administrative task profile to an acceptable standard, and that is usually a far cheaper proposition. Set against that cost sits the expected cost of the alternative: regulatory penalty, incident response, remediation, notification, litigation, and reputational damage, none of which appear in the technology budget but all of which are paid from the same institutional purse. The #procurement conversation improves considerably when both sides of that ledger are on the table. 9. Limitations Four limitations should be stated plainly. The paper is conceptual. It synthesises existing evidence and proposes a design; it does not test that design. The architecture in Section 6 is a set of design propositions, and its effectiveness in reducing measured leakage remains an empirical question. The review is integrative rather than systematic. Sources were selected interpretively across four literatures, and no formal protocol was registered. A systematic review with explicit search strings and screening criteria would strengthen the evidence base and might surface findings this synthesis missed. Technical claims have a shelf life. Model capabilities, attack techniques, and defensive methods change quickly, and specific statements about what current systems can and cannot do should be treated as provisional. The structural analysis, meaning the drivers, the pathways, and the layered response, is expected to be more durable than any particular technical detail. Context varies. The recommendations assume an institution with an identifiable technology function, a document estate that can in principle be classified, and enough resource to run a pilot. Small organisations lacking these will need a lighter path, most plausibly a contracted private endpoint with strong terms rather than self hosting, and the paper does not develop that path in detail. 10. Future Research Six directions follow directly. Prevalence. Rigorous, anonymous measurement of unsanctioned assistant use in administrative functions, across sectors and jurisdictions, is almost entirely absent from the peer reviewed record. The behaviour is widely asserted and rarely measured. Intervention effects. Comparative studies of what actually changes behaviour, contrasting prohibition, training, in interface warnings, and provision of a safe alternative, would put the argument of this paper to a real test. The prediction it makes is falsifiable: provision should outperform prohibition on measured leakage. Redaction under real conditions. The performance of automated de-identification on genuine administrative text, which is messy, multilingual, abbreviated, and full of local conventions, is not well characterised. Published benchmarks rarely resemble a real registry inbox. Injection defence in document heavy workflows. Administrative assistants ingest untrusted documents by design. Defences evaluated in this specific setting, rather than in general benchmarks, are needed. Trust and adoption. What makes administrative staff trust an internal assistant enough to abandon a public one they already like? Citation of sources, visible logging, and demonstrated accuracy are plausible mechanisms and remain untested. Governance transfer. Whether existing frameworks such as the #NIST_AI_RMF and #ISO_42001 translate effectively to resource constrained institutions, including small universities, local authorities, and community health providers, is an open and practically urgent question. 11. Conclusion The staff member who pastes a confidential letter into a public chatbot is not the problem. She is the symptom. The problem is an institution that asked for speed, provided no safe way to achieve it, communicated its rules ambiguously, and then expressed surprise when people found their own way. This paper has argued that combating #Shadow_AI requires substitution rather than suppression. It set out why administrative work sits at the sharpest edge of the risk, because the tasks best suited to automation are the same tasks that carry the most regulated information. It explained why prohibition fails, since it leaves demand intact, is technically porous, and destroys the visibility on which any real control depends. It identified seven distinct leakage pathways and showed that the usual institutional responses address only one of them. It proposed a seven layer architecture for an internal assistant, combining controlled hosting, retrieval over approved corpora, permission aware access, a protective pipeline with automatic redaction, adversarial resilience, complete logging, and continuous evaluation. And it placed that architecture inside a governance shell with a named owner, a usable policy, a maturity path, and a staged roadmap. The proposal is not a claim of safety. An internal assistant with careless permissions can leak faster than any public tool, and a locally hosted model fabricates as confidently as a remote one. The claim is that these are failures an institution can see, measure, and fix, while the failures of shadow usage are, by definition, ones it cannot. Every organisation now faces a choice it cannot avoid by inaction. It can build the safe path, or it can watch its staff build the unsafe one. They are already building. The only open question is whether anyone is watching, and whether what they find is something the institution designed or something it merely allowed. Hashtags #Shadow_AI #Generative_AI_Governance #Data_Leak_Prevention #Secure_Internal_AI #Administrative_Automation #LLM_Security #Enterprise_AI #Information_Governance #Retrieval_Augmented_Generation #Privacy_By_Design #AI_Risk_Management #Shadow_IT #Responsible_AI #Data_Protection #Digital_Transformation References Brynjolfsson, E., Li, D., and Raymond, L. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2), 889 to 942. Carlini, N., Ippolito, D., Jagielski, M., Lee, K., Tramer, F., and Zhang, C. (2023). Quantifying memorization across neural language models. Proceedings of the Eleventh International Conference on Learning Representations. Carlini, N., Tramer, F., Wallace, E., Jagielski, M., Herbert-Voss, A., Lee, K., Roberts, A., Brown, T., Song, D., Erlingsson, U., Oprea, A., and Raffel, C. (2021). Extracting training data from large language models. Proceedings of the 30th USENIX Security Symposium, 2633 to 2650. Chen, L., Zaharia, M., and Zou, J. (2024). How is ChatGPT's behavior changing over time? Harvard Data Science Review, 6(2). Das, B. C., Amini, M. H., and Wu, Y. (2025). Security and privacy challenges of large language models: A survey. ACM Computing Surveys, 57(6), 1 to 39. https://doi.org/10.1145/3712001 Dell'Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., and Lakhani, K. R. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Technology and Operations Management Unit Working Paper 24-013. Deng, G., Liu, Y., Li, Y., Wang, K., Zhang, Y., Li, Z., Wang, H., Zhang, T., and Liu, Y. (2024). MASTERKEY: Automated jailbreaking of large language model chatbots. Proceedings of the Network and Distributed System Security Symposium. Dettmers, T., Pagnoni, A., Holtzman, A., and Zettlemoyer, L. (2023). QLoRA: Efficient finetuning of quantized LLMs. Advances in Neural Information Processing Systems, 36. European Union (2024). Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence. Official Journal of the European Union. Floridi, L. (2023). The ethics of artificial intelligence: Principles, challenges, and opportunities. Oxford University Press. Fürstenau, D., Rothe, H., and Sandner, M. (2021). Leaving the shadow: A configurational approach to explain post-identification outcomes of shadow IT systems. Business and Information Systems Engineering, 63(2), 97 to 111. https://doi.org/10.1007/s12599-020-00635-2 Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., Dai, Y., Sun, J., Wang, M., and Wang, H. (2023). Retrieval-augmented generation for large language models: A survey. arXiv preprint. Greshake, K., Abdelnabi, S., Mishra, S., Endres, C., Holz, T., and Fritz, M. (2023). Not what you have signed up for: Compromising real-world LLM-integrated applications with indirect prompt injection. Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security, 79 to 90. https://doi.org/10.1145/3605764.3623985 Hacker, P., Engel, A., and Mauer, M. (2023). Regulating ChatGPT and other large generative AI models. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 1112 to 1123. https://doi.org/10.1145/3593013.3594067 Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W. (2022). LoRA: Low-rank adaptation of large language models. Proceedings of the Tenth International Conference on Learning Representations. Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., and Liu, T. (2025). A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2), 1 to 55. https://doi.org/10.1145/3703155 International Organization for Standardization and International Electrotechnical Commission (2023). ISO/IEC 42001:2023 Information technology, artificial intelligence, management system. ISO. Jaeger, L., and Eckhardt, A. (2021). Eyes wide open: The role of situational information security awareness for security-related behaviour. Information Systems Journal, 31(3), 429 to 472. https://doi.org/10.1111/isj.12317 Mallmann, G. L., and Maçada, A. C. G. (2021). The mediating role of social presence in the relationship between shadow IT usage and individual performance: A social presence theory perspective. Behaviour and Information Technology, 40(4), 427 to 441. https://doi.org/10.1080/0144929X.2019.1702101 Marcus, G. (2024). Taming Silicon Valley: How we can ensure that AI works for us. MIT Press. Mökander, J., Schuett, J., Kirk, H. R., and Floridi, L. (2023). Auditing large language models: A three-layered approach. AI and Ethics, 4, 1085 to 1115. https://doi.org/10.1007/s43681-023-00289-2 Narayanan, A., and Kapoor, S. (2024). AI snake oil: What artificial intelligence can do, what it cannot, and how to tell the difference. Princeton University Press. National Institute of Standards and Technology (2023). Artificial intelligence risk management framework, AI RMF 1.0. NIST AI 100-1. U.S. Department of Commerce. Neel, S., and Chang, P. (2023). Privacy issues in large language models: A survey. arXiv preprint. Novelli, C., Casolari, F., Rotolo, A., Taddeo, M., and Floridi, L. (2024). Taking AI risks seriously: A new assessment model for the AI Act. AI and Society, 39, 2493 to 2497. https://doi.org/10.1007/s00146-023-01723-z Noy, S., and Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187 to 192. https://doi.org/10.1126/science.adh2586 Schneier, B. (2023). A hacker's mind: How the powerful bend society's rules, and how to bend them back. W. W. Norton and Company. Schuett, J. (2024). Risk management in the Artificial Intelligence Act. European Journal of Risk Regulation, 15(2), 367 to 385. https://doi.org/10.1017/err.2023.1 Shneiderman, B. (2022). Human-centered AI. Oxford University Press. Solaiman, I. (2023). The gradient of generative AI release: Methods and considerations. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 111 to 122. https://doi.org/10.1145/3593013.3593981 Suleyman, M., and Bhaskar, M. (2023). The coming wave: Technology, power, and the twenty-first century's greatest dilemma. Crown. Touvron, H., Martin, L., Stone, K., and colleagues (2023). Llama 2: Open foundation and fine-tuned chat models. arXiv preprint. Veale, M., and Zuiderveen Borgesius, F. (2021). Demystifying the draft EU Artificial Intelligence Act. Computer Law Review International, 22(4), 97 to 112. https://doi.org/10.9785/cri-2021-220402 Weidinger, L., Uesato, J., Rauh, M., Griffin, C., Huang, P., Mellor, J., Glaese, A., Cheng, M., Balle, B., Kasirzadeh, A., Biles, C., Brown, S., Kenton, Z., Hawkins, W., Stepleton, T., Birhane, A., Hendricks, L. A., Rimell, L., Isaac, W., Haas, J., Legassick, S., Irving, G., and Gabriel, I. (2022). Taxonomy of risks posed by language models. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 214 to 229. https://doi.org/10.1145/3531146.3533088 Wu, X., Duan, R., and Ni, J. (2024). Unveiling security, privacy, and ethical concerns of ChatGPT. Journal of Information and Intelligence, 2(2), 102 to 115. https://doi.org/10.1016/j.jiixd.2023.10.007 Yao, Y., Duan, J., Xu, K., Cai, Y., Sun, Z., and Zhang, Y. (2024). A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly. High-Confidence Computing, 4(2), 100211. https://doi.org/10.1016/j.hcc.2024.100211 Zhao, W. X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., Du, Y., Yang, C., Chen, Y., Chen, Z., Jiang, J., Ren, R., Li, Y., Tang, X., Liu, Z., Liu, P., Nie, J., and Wen, J. (2023). A survey of large language models. arXiv preprint.

Offices in Nairobi, Warsaw, Manila, Dubai and Toronto increasingly look alike. They share the same open floor plates, the same glass meeting boxes, the same barista counter near reception, the same green wall behind the sofa, and the same signage telling staff to book a desk through an app. This article asks a simple question with complicated answers: why does #workplace_design converge across firms that differ in size, sector, national culture and climate? Drawing on #institutional_isomorphism as developed in organizational sociology, the paper builds an integrative conceptual framework that connects three well known pressures, namely coercive, mimetic and normative pressures, to the concrete material choices that shape modern offices. The article synthesises literature on organizational space, corporate real estate, activity based working, coworking and post pandemic hybrid arrangements, and it uses that synthesis to explain why standardization spreads even when local evidence about outcomes is weak or contradictory. The analysis identifies the carriers of convergence, including global design and consulting firms, furniture manufacturers, certification bodies, professional associations, industry benchmarks and business media. It also identifies the consequences, including #decoupling between the stated purpose of a workplace and its everyday use, the exporting of design assumptions that suit some climates and cultures but not others, and the tendency of firms to treat visible design features as proof of a modern #organizational_culture. The article closes with propositions for future research, methodological suggestions for students, and practical guidance for managers who want workplaces that fit their people rather than workplaces that merely fit the trend. The central argument is that the modern office is not only a functional container for work. It is also a claim about #legitimacy, and that claim is increasingly written in a global vocabulary that few organizations author themselves. Keywords: institutional theory, organizational space, corporate real estate, office design, hybrid work, legitimacy, decoupling, standardization 1. Introduction Walk into the head office of a mid sized insurance company in one country, then walk into a software firm in another country on a different continent, and you may struggle to tell the two apart from photographs alone. The reception desk is low and pale. There is a plant wall. There is a coffee bar with high stools. Behind a glass partition there are small rooms named after cities, mountains or jazz musicians. There is a long bench of shared desks with monitors on arms, a set of soundproof phone booths, and a lounge area with a rug and a bookshelf that holds design magazines nobody reads. There is a sign asking staff to clear their desk at the end of the day. This #visual_similarity is not an accident, and it is not fully explained by cost, technology or efficiency. Two firms in the same city, with different budgets, different work processes and different workforces, will often build offices that resemble each other far more than either resembles the office it replaced. Firms that operate in very different regulatory and cultural settings will still adopt the same #open_plan_office layouts, the same booking software, the same #wellness_features and the same language of collaboration. Organizational theory has a name for this pattern. DiMaggio and Powell described the tendency of organizations in a shared field to grow more alike over time, not necessarily because similarity makes them more efficient, but because similarity makes them more legitimate, more fundable, more employable and more comprehensible to the outside world (DiMaggio and Powell, 1983). They called this process #institutional_isomorphism, and they argued that it operates through three broad mechanisms. Coercive pressure comes from regulation, from law, and from powerful actors on whom an organization depends. Mimetic pressure comes from uncertainty, which pushes organizations to copy peers they consider successful. Normative pressure comes from professions, whose training, associations and shared standards spread a common definition of what good practice looks like. Workplace design is an unusually clear field in which to observe these mechanisms, for three reasons. First, the output is physical and visible. Unlike a strategy or a culture, a floor plan can be photographed, measured, benchmarked and copied. Design is the most public part of an organization's internal life, and it circulates freely through award schemes, magazines, real estate marketing, social media and recruitment pages. Visibility accelerates #diffusion. Second, evidence about what actually works is unusually contested. Decades of research on open plan environments report mixed and often negative findings for concentration, satisfaction and even face to face interaction, yet the open plan remains the global default (Bernstein and Turban, 2018; Haapakangas et al., 2018). Where causal evidence is weak, imitation fills the gap. Uncertainty is precisely the condition under which #mimetic_isomorphism thrives. Third, workplace decisions are made by a dense professional network of architects, designers, corporate real estate managers, facility managers, human resource leaders and consultants, most of whom are trained in similar programmes, attend the same conferences, read the same benchmarks and use the same vocabulary. This is a textbook setting for #normative_isomorphism. The result is what this article calls the standardized global office. It is a recognisable package of features, including reduced #desk_ownership, a variety of settings for different tasks, generous social and food related space, visible wellbeing and sustainability signals, digital booking and #space_utilization sensors, and a strong emphasis on collaboration as the office's core justification. This package travels. It travels into countries with different #labour_laws, into climates where large glazed facades are thermally punishing, into cultures where hierarchy is expressed through private space, and into sectors where most work is individual and concentration heavy rather than collaborative. 1.1 Research problem The problem is not that global standards exist. #global_standards can raise the floor of quality, protect health and safety, reduce energy use and make buildings comparable. The problem is that the mechanisms which spread workplace standards are only loosely connected to the mechanisms that would test whether those standards serve a particular organization's people and tasks. When adoption is driven mainly by #legitimacy rather than by fit, the organization may end up with an office that photographs well, satisfies investors and reassures graduate recruits, while quietly frustrating the people who work in it every day. Institutional theory calls the gap between adopted form and actual practice #decoupling (Meyer and Rowan, 1977; Bromley and Powell, 2012). 1.2 Research questions This article addresses four questions. Through which specific mechanisms do coercive, mimetic and normative pressures operate in the field of workplace design? Who are the carriers that move design templates across borders, sectors and organizational sizes? What outcomes follow from convergence, in terms of employee experience, organizational performance claims, and the relationship between stated purpose and actual use? Under what conditions do organizations resist, edit or translate global templates rather than simply importing them? 1.3 Contribution The article makes three contributions. Theoretically, it extends #institutional_isomorphism into the material and spatial domain, where the artefacts of imitation are unusually durable and expensive. Space, unlike a policy document, cannot be quietly revised at the next committee meeting. A twelve year lease locks in an institutional decision long after the trend that inspired it has faded. Empirically, it organises a scattered literature on offices, coworking, activity based working and hybrid work into a single institutional frame. Practically, it offers a diagnostic vocabulary that allows students, researchers and managers to distinguish between design choices that are grounded in evidence about their own work and design choices that are grounded in what everybody else appears to be doing. 1.4 Structure of the article Section 2 reviews the theoretical foundations, covering institutional theory, the three isomorphic mechanisms, and the sociology of organizational space. Section 3 develops the conceptual framework and states testable propositions. Section 4 explains the review method and its limits. Section 5 presents the analysis, mechanism by mechanism, and then examines five design templates in detail. Section 6 discusses outcomes, including decoupling, inequality and environmental cost. Section 7 examines resistance, translation and divergence. Section 8 offers implications, and Section 9 concludes. 2. Theoretical Foundations 2.1 The institutional turn in organizational analysis Classical organizational theory assumed that structures follow function. An organization designs itself to be efficient, and if it fails, competition eliminates it. #neo_institutionalism challenged this. Meyer and Rowan argued that many formal structures in modern organizations are adopted because they are rationalised myths circulating in the environment, not because they demonstrably improve output (Meyer and Rowan, 1977). Organizations adopt them to signal that they are proper, modern and trustworthy through #ceremonial_adoption. Conformity buys survival, resources and social support, sometimes at the cost of technical efficiency. To protect the technical core from these ceremonial structures, organizations may decouple, meaning they display the structure while continuing to work as before. DiMaggio and Powell extended this by asking why organizational fields become homogeneous. Their answer, the #iron_cage of isomorphism, remains one of the most cited arguments in the social sciences (DiMaggio and Powell, 1983). Their key move was to shift attention from the single organization to the #organizational_field, defined as the set of organizations that constitute a recognised area of institutional life, including suppliers, consumers, regulators and competitors. Once a field is structured, and once its members interact frequently, share information, and are exposed to the same professional discourse, homogeneity follows. Later work refined this picture. Organizations are not passive. They respond strategically, ranging from acquiescence through compromise, avoidance and defiance to manipulation (Oliver, 1991). Ideas do not simply diffuse like a virus; they are edited and translated through #translation as they move, gaining local features and losing others (Sahlin and Wedlin, 2008; Ansari, Fiss and Zajac, 2010). Fields are also arenas of competing #institutional_logics, so what counts as appropriate may be contested rather than settled (Thornton, Ocasio and Lounsbury, 2012; Lounsbury et al., 2021). Contemporary reviews continue to debate how far institutional theory has drifted from its critical roots, and whether it now explains everything and therefore nothing (Alvesson and Spicer, 2019; Munir, 2020; Ocasio and Gai, 2020). These debates matter here, because a purely descriptive account of convergence in offices would miss the question of who benefits from it. 2.2 The three mechanisms #coercive_isomorphism arises from formal and informal pressure exerted by organizations on which a focal organization depends, and by cultural expectations in the wider society. In workplace design, coercion appears through building codes, fire regulations, accessibility law, occupational health and safety rules, #energy_performance requirements, and increasingly through disclosure regimes tied to environmental and social reporting. It also appears through dependence. A subsidiary that depends on a global headquarters for budget approval will build what the group real estate standard specifies. A tenant that depends on a landlord's fit out guidelines will build within them. A supplier that wants a contract with a large client may need to demonstrate the client's preferred certification. Coercion is not only about the state. #mimetic_isomorphism arises from uncertainty. When goals are ambiguous and cause and effect relationships are poorly understood, organizations model themselves on others they perceive as legitimate or successful. Few managers can state with confidence how many square metres of collaborative space produce how much innovation. Under this ambiguity, copying a visible and admired peer is a rational shortcut and a defensible one. If the design fails, the manager can point out that it is what leading firms do. #normative_isomorphism arises from professionalization. Two aspects matter. First, formal education creates a cognitive base and a shared vocabulary. Architects, interior designers, facility managers and corporate real estate professionals learn similar canons and similar precedents. Second, professional networks span organizations and diffuse models rapidly. Membership associations, accreditation schemes, conferences, awards and trade journals define what a competent professional recommends. A designer who proposes a floor of enclosed private offices to a client in the current climate is not merely making an unusual choice; they are risking their professional standing. Boxenbaum and Jonsson note that these three mechanisms are analytically distinct but empirically entangled (Boxenbaum and Jonsson, 2017). A certification scheme, for example, may begin as a normative professional project, become a mimetic marker of prestige, and end as a coercive requirement written into a lease or a lending condition. Workplace design offers exactly this trajectory, and the article returns to it in Section 5. 2.3 Space as an organizational and institutional phenomenon Research on organizational space has grown considerably. Space is not a neutral container. It is produced through practice, it carries meaning, and it enacts power (Dale and Burrell, 2008; van Marrewijk and Yanow, 2010). Reviews have mapped the field and argued for treating space as a process rather than a fixed backdrop, attending to how spatial arrangements are used, contested and reinterpreted over time (Weinfurtner and Seidl, 2019; Stephenson et al., 2020). Space is also an active participant in organizing, shaping who meets whom, what is visible, and what is possible to say (Cnossen and Bencherki, 2019). This literature has largely developed in parallel with institutional theory rather than in dialogue with it. That separation is a missed opportunity. Institutional theory explains why particular spatial forms become taken for granted across a field. Spatial theory explains what those forms do once they are built. Bringing them together allows us to ask a sharper question. If the standardized office is an institutional artefact, what work does it perform for the organization that adopts it, and what work does it perform on the people inside it? The applied literature on offices adds the empirical texture. Handbooks now systematically link theories from psychology, economics and management to office environments and the services around them (Appel-Meulenbroek and Danivska, 2021; Danivska and Appel-Meulenbroek, 2021). Research on #activity_based_working has produced a large and uneven body of findings, with systematic reviews reporting inconsistent effects on health, satisfaction and performance, and highlighting the importance of fit between the person, the task and the setting (Engelen et al., 2019; Hoendervanger et al., 2019; Marzban et al., 2023). Studies of the physical office as a resource for mental health point to the importance of factors such as noise, privacy, light and control, which are precisely the factors that dense open layouts tend to compromise (Bergefurt et al., 2022). Work on new ways of working shows that spatial change is usually bundled with cultural and technological change, which makes clean evaluation difficult and makes rhetorical claims easy (Kingma, 2019; Aroles, Mitev and de Vaujany, 2019). 2.4 Legitimacy, decoupling and the visible office Legitimacy is a generalised perception that an organization's actions are desirable and appropriate within a socially constructed system of norms and beliefs (Suddaby, Bitektine and Haack, 2017). Offices are legitimacy machines. They are visited by clients, investors, journalists, regulators and candidates. They appear in annual reports and recruitment videos. A workplace therefore communicates before anyone inside it has done any work at all. It says: we are modern, we are collaborative, we care about wellbeing, we are sustainable, we are the kind of employer a talented person joins. Because the signal can be sent through visible features alone, the temptation to decouple is strong. An organization can install a wellbeing room while maintaining workloads that make wellbeing impossible. It can install a green wall while running a facade that leaks energy. It can call a floor collaborative while its teams sit in silence with headphones on. Bromley and Powell distinguish between policy practice decoupling, where the stated policy is not implemented, and means ends decoupling, where the policy is implemented but has no demonstrated connection to the outcome it claims to produce (Bromley and Powell, 2012). Workplace design is unusually vulnerable to the second kind. Firms genuinely do build the collaboration zones. They genuinely do remove assigned desks. What often remains unproven is whether these features produce the outcomes used to justify them. 3. Conceptual Framework and Propositions 3.1 The framework in outline The framework proposed here has four elements: pressures, carriers, templates and outcomes. Pressures are the three isomorphic mechanisms already described. They do not act directly on a floor plan. They act on the decision makers who commission, design, approve and occupy it. Carriers are the actors and artefacts that transport design templates across organizational boundaries. They include global architecture and design practices, real estate advisory firms, management consultancies, furniture manufacturers, certification bodies, professional associations, industry benchmarking services, award schemes, business media, and the mobile professionals who carry expectations from one employer to the next. Templates are recognisable packages of spatial and managerial features. This article examines five: the open plan, activity based working, the coworking aesthetic, the wellbeing and sustainability certified office, and the hybrid or #hot_desking office of the post pandemic period. Outcomes include structural convergence across the field, employee level experiences that may or may not match the promises attached to the template, decoupling between form and function, and the redistribution of costs and benefits among groups inside and outside the organization. The framework treats these as a cycle rather than a chain. Outcomes feed back into pressures, because once enough organizations adopt a template it becomes the taken for granted baseline, and departing from it requires justification. This is the mechanism by which a novelty becomes a norm and then becomes an obligation. 3.2 Propositions The following propositions are offered for empirical testing. They are stated in a form that a student could operationalise in a dissertation. P1. The more visible a workplace feature is to external audiences, the more rapidly it diffuses across an organizational field, independent of evidence of its effect on task performance. Reception areas, cafes and social spaces should diffuse faster than acoustic treatment or storage. P2. Under conditions of high uncertainty about how work will be organised, such as during and immediately after a major disruption, mimetic pressure increases and the diversity of workplace designs in a field decreases. The period following the coronavirus pandemic should therefore show convergence rather than experimentation, despite the rhetoric of experimentation. P3. Organizations that are more dependent on external audiences for resources, including listed companies, professional service firms, and firms competing for scarce graduate talent, will adopt legitimacy oriented design features earlier and more completely than organizations that are less exposed. P4. The greater the proportion of workplace decisions made by professionals whose reference group lies outside the organization, the stronger the normative convergence of its design. P5. The probability of means ends decoupling increases with the abstractness of the outcome claimed for a design. Claims about innovation, culture and collaboration are more prone to decoupling than claims about desk cost per person or energy use, because the latter are measured routinely and the former are not. P6. Templates are more likely to be translated rather than copied where local coercive pressures, such as labour codes, works council rights, climate and building regulation, contradict the imported form, and where the local professional field has independent status and vocabulary. 4. Method 4.1 Design of the review This article is a conceptual and integrative review rather than a primary empirical study. That choice is deliberate and should be stated plainly, because students often encounter review articles that present themselves as if they had discovered new data. This article has not. Its purpose is to reorganise existing knowledge under a theoretical frame that has not been consistently applied to it, and to generate propositions that can then be tested. Literature was gathered from three streams. The first is institutional theory in organization studies, including the foundational statements and the recent critical reappraisals. The second is research on organizational space and the sociology of work environments. The third is applied workplace research published in built environment, corporate real estate, facility management, environmental psychology and human resource outlets. Sources were selected for relevance to the question of convergence rather than for exhaustive coverage of any single stream, which is the standard limitation of an integrative approach. 4.2 Analytical procedure Each source was read for two things. First, does it describe or explain a pressure toward similarity in workplace form? Second, does it identify a carrier that transmits that pressure? Findings were then grouped by mechanism, and the five design templates were reconstructed as ideal types, meaning simplified models that no single office matches exactly but that many offices approximate. Illustrative examples are drawn from publicly reported practice and from published studies. They are used to make mechanisms concrete, not to prove them. Where the article describes patterns without a citation, it should be read as an argument to be tested rather than as an established fact, and this distinction is signalled in the text. 4.3 Limitations of the method Three limits should be acknowledged at the outset rather than buried in the conclusion. The published workplace literature is heavily weighted toward North America, Western Europe, Australia and a small number of East Asian settings, so claims about global convergence rest on evidence that is itself geographically skewed. Much of the practitioner evidence comes from organizations with a commercial interest in the templates they evaluate, including design firms, furniture manufacturers and benchmarking services, and this should be treated as data about the field rather than as neutral evidence about outcomes. Finally, a conceptual review cannot establish causality; it can only sharpen the questions that a causal study would need to answer. 5. Analysis: How the Standardized Office Is Produced 5.1 Coercive pressures and the office Coercion is the most obvious mechanism and the most often underestimated one, because people associate it only with government. Regulation. Building codes, fire safety rules, minimum ventilation rates, accessibility standards and occupational safety law set hard boundaries. These already produce a floor of similarity across buildings within a jurisdiction. Energy performance requirements have tightened in many countries, and where disclosure of building energy ratings is mandatory, the market prices it, which converts a public rule into a private incentive. Regulation therefore standardises not only what is permitted but what is valuable. #corporate_real_estate standards. Multinational groups typically publish internal workplace standards that specify the ratio of desks to people, the mix of settings, the finishes, the technology, the signage and sometimes the coffee. A country manager who wishes to deviate must build a business case against a group standard that already carries the authority of the centre. This is a private form of law, and it is one of the most powerful engines of global convergence. It also travels quickly, because a group standard written for a temperate headquarters can be applied within a year to an office in a hot, humid or extremely cold location, where the same glazing ratio and the same open ceiling produce entirely different thermal and acoustic results. Leases and landlords. Commercial #leases increasingly include fit out guidelines, sustainability clauses and requirements about how the space may be altered. Landlords providing category A fit out effectively pre determine ceiling heights, lighting grids, floor boxes and partition lines. Tenants then design within those constraints and, unsurprisingly, arrive at similar answers. Finance and disclosure. As sustainability reporting has moved from voluntary to mandatory in a growing number of jurisdictions, the real estate portfolio has become a reportable item. Emissions from buildings appear in corporate disclosures. Lenders and investors ask about the certification status of assets. What began as a reputational nicety becomes a financing condition, which is coercion in its purest institutional sense. #ESG reporting has therefore become one of the most effective standardising forces in the built environment, even though its object is energy and carbon rather than the shape of a desk. Health after the pandemic. The coronavirus period generated a wave of coercive pressure in the form of occupancy limits, distancing rules, ventilation guidance and cleaning protocols. Many of these were temporary. Their institutional residue was not. Ventilation and air quality moved from an engineering detail to a board level topic, and features that signalled health became part of the expected package. Proposition link. These pressures support P3. Organizations that are exposed to regulators, investors and landlords adopt the standard package earliest, because for them the cost of non conformity is immediate and financial rather than merely aesthetic. 5.2 Mimetic pressures and the appeal of the visible peer Mimicry requires two conditions: uncertainty about what works, and a visible model that appears to be succeeding. Workplace design supplies both in abundance. The uncertainty is genuine. Nobody has a reliable production function for knowledge work. Research on the relationship between physical environment and outcomes is real and growing, but it is contingent, contextual and frequently contradictory. Studies of transitions to open and activity based environments report improvements in some outcomes and deterioration in others, often within the same organization (Haapakangas et al., 2018; Engelen et al., 2019; Marzban et al., 2023). One influential study using wearable sensors found that moving to a more open layout reduced face to face interaction and increased electronic communication, which is the opposite of the stated intention (Bernstein and Turban, 2018). Faced with such findings, a manager could reasonably conclude that the evidence is unclear. Under unclear evidence, imitation is the least embarrassing option. The models are extremely visible. The workplaces of a small number of technology firms, the #tech_campus model, have functioned as the field's reference architecture for two decades. Their campuses were photographed, toured, written about and imitated with astonishing speed. Features designed for a particular business model, a young workforce, a specific city, and an unusual cost structure were exported to banks, ministries, universities and hospitals. The slide, the beanbag and the table football entered organizations that had no cultural use for them, and left again quietly. What survived was the underlying template: low partitions, abundant social space, visible informality, and the interpretation of these features as evidence of innovation. Benchmarking industrialises imitation. #benchmarking services collect data on utilisation, satisfaction and cost across thousands of buildings and hundreds of organizations, and then report where a client sits relative to peers. This is genuinely useful information. It is also, institutionally, a machine for producing conformity, because a benchmark defines a normal distribution and invites every organization to move toward the successful end of it. The measure becomes the target. When a benchmark reports that leading organizations provide seven desks for every ten staff, that ratio becomes a goal rather than an observation, and the reasons why it worked in the original setting are rarely transported with the number. Awards, media and the photographable office. #design_awards, magazine features and social media create a visual canon. Photographs favour certain conditions: strong daylight, long sightlines, clean surfaces, empty desks, an interesting ceiling, one plant, and no people. An office designed to photograph well is not necessarily an office designed to work well. This is an important and underexamined mechanism. The circulation of images, rather than the circulation of evidence, may be the single most powerful mimetic force in the field. It supports P1 directly: the most visible features diffuse fastest. Crisis and convergence. P2 proposed that disruption increases mimicry. The pandemic period is a natural test. Organizations were confronted with radical uncertainty about attendance, purpose and cost. Rather than producing a flowering of diverse local solutions, the period produced a remarkably uniform vocabulary within roughly two years: the office as a destination, the office for collaboration, neighbourhoods for teams, reduced desk ratios, booking apps, and an emphasis on culture. The uniformity of the language across sectors and countries is itself evidence of mimetic pressure operating at speed. 5.3 Normative pressures and the profession of workplace Normative pressure is quieter than the other two, and more durable, because it works through what professionals believe rather than through what they are forced to do. Education and shared canon. Architects and interior designers are trained through precedent. Studios teach by reference to admired buildings. Corporate real estate and facility management professionals are trained through certification programmes and bodies of knowledge that codify best practice. Human resource professionals learn a vocabulary of engagement, employer brand and employee experience that maps easily onto spatial features. When three professions with different training converge on the same solution, it is not because they independently verified it. It is because the solution has become the shared professional common sense. Associations and standards. Professional bodies in facility management, corporate real estate and workplace strategy hold conferences, publish guidance, run accreditation and give awards. International management system standards for facility management provide a common structure for how the function should be organised. Wellbeing and sustainability rating systems provide a common checklist of what a good building contains. Each of these is defensible in isolation. Together they form a normative infrastructure that makes the standard office the professionally correct answer, and any deviation a matter requiring explanation. Consultants as translators and homogenisers. Management and real estate consultancies occupy a structural position between organizations, and they are paid, in part, for their knowledge of what other organizations are doing (Sturdy, Wright and Wylie, 2016). Their business model rewards portable frameworks. A method that must be rebuilt from scratch for every client is unprofitable; a method that can be repackaged is highly profitable. #consultants therefore do not merely transmit templates. They have an economic interest in the existence of templates. The mobility of staff. Professionals move between employers and carry expectations with them. A workplace lead who delivered an activity based project at one firm will propose it at the next, because it is what they know, what they can execute and what their reputation rests on. This is #professionalization acting through careers, and it supports P4. 5.4 Carriers: the infrastructure of convergence It is worth naming the carriers explicitly, because students often treat isomorphism as an abstract force. It is not. It is carried by identifiable organizations with identifiable interests. Global design and architecture practices operate the same delivery model across continents, use the same software, the same detail libraries and often the same project teams flying between regions. Their product is partly the reassurance that a client is getting the same quality as its international peers, which is a promise of similarity. Real estate advisory and brokerage firms publish research, advise on portfolios, and set the terms in which occupancy is discussed. They define the categories, such as grade A, prime, flexible and core, in which the market thinks. #furniture_manufacturers are among the most influential and least analysed carriers. They fund research, publish workplace insight, sponsor conferences, and design products that presuppose a way of working. A product range built around unassigned seating, mobile storage and modular soft seating is not a neutral object. It embeds a theory of work, and once a firm has bought it, that theory is installed for a decade. #certification_bodies convert contested design ideas into checkable criteria. This is an enormously powerful act of institutionalisation, because it turns a debate into a score. Once a score exists, it can be required, compared and marketed. Media, awards and social platforms provide the visual circulation described earlier. Flexible space operators industrialised a particular aesthetic and a particular density and then sold it, at scale, to corporate occupiers, which meant that even organizations that never leased a coworking desk absorbed its design language. 5.5 Five templates 5.5.1 The open plan The open plan is the oldest and most consequential template. Its justification has shifted repeatedly over a century, from efficiency and supervision, through communication and hierarchy flattening, to collaboration and cost. The persistence of the form across changing justifications is itself a strong indicator of institutional rather than technical logic. When the reason changes but the answer stays the same, the answer came first. The evidence base is not kind to the strong version of the claim. Research consistently associates high density open environments with noise distraction, reduced perceived privacy and lower satisfaction, and at least one influential study found reduced rather than increased face to face interaction after an office opened up (Bernstein and Turban, 2018; Haapakangas et al., 2018; Bergefurt et al., 2022). The counterargument, that open plans are cheaper per person, is true and is rarely stated openly, because #cost_reduction is a less inspiring justification than collaboration. The institutional function of the collaboration narrative is precisely to make a cost decision sound like a culture decision. This is #myth_and_ceremony in the original sense. 5.5.2 Activity based working #activity_based_working proposed a sophisticated correction. Rather than one setting for everyone, provide many settings suited to different tasks, and let people choose. In principle this respects the diversity of knowledge work. In practice, outcomes depend heavily on implementation. Systematic reviews report inconsistent effects on satisfaction, health and performance, and the most consistent finding is that outcomes depend on perceived fit between the person, the task and the chosen setting, and on whether people actually switch settings as intended (Engelen et al., 2019; Hoendervanger et al., 2019; Marzban et al., 2023). The institutional story is that the second half of the model, meaning the behavioural change, the etiquette, the management support and the sufficient supply of enclosed settings, is expensive and invisible, while the first half, meaning the varied furniture and the removal of assigned desks, is cheap, visible and immediately produces a saving. Organizations therefore adopt the visible half. The result is often not activity based working at all but desk reduction with a nicer palette. The template has been adopted; the practice has been decoupled. This is a direct illustration of P5. 5.5.3 The coworking aesthetic #coworking began as a genuinely different institutional form, associated with independent workers, community and the possibility of economic diversity outside the corporation (Bouncken and Reuschl, 2018; Vidaillet and Bousalham, 2020). Within a decade its visual language, meaning exposed services, industrial materials, neon, plants, communal tables, coffee and a curated informality, had been absorbed almost completely by conventional corporate interiors. The community was left behind; the aesthetic travelled. This is one of the clearest examples of translation in the field, but a translation that strips out the political content of the original. What corporations imported was not the sharing of resources between independent workers, but the appearance of a startup, which functions as a claim to be agile and youthful. The form was decoupled from its purpose at the moment of adoption. 5.5.4 The certified wellbeing and sustainability office #WELL_certification and #LEED style rating systems for environmental performance and occupant health have become close to obligatory in premium markets. They deliver real benefits. Better ventilation, daylight, thermal comfort and materials are not marketing; they matter for health and cognition, and the research on wellbeing in the built environment supports attention to these factors (Altomonte et al., 2020; Bergefurt et al., 2022). Studies comparing occupant outcomes before and after relocation into certified buildings report improvements in perceived environmental quality and related outcomes, though isolating the effect of certification itself from the effect of simply moving into a newer building remains methodologically difficult (Licina and Yildirim, 2021). The institutional dynamic is nevertheless visible. Certification converts a complex, contextual design problem into a points system. Points can be pursued efficiently, which means design effort flows toward the credits that are easiest to earn rather than toward the improvements that matter most in a given climate or for a given occupant group. A green wall is highly visible, photographs beautifully and signals commitment. Improving the acoustic performance of a ceiling is invisible and signals nothing, although it may do far more for the people below it. #biophilic_design is genuinely supported by evidence, but the version that spreads fastest is the version that can be seen from the lobby. 5.5.5 The hybrid and shared desk office The most recent template is the post pandemic office. Its logic runs as follows. Remote work is now feasible and desired. Attendance falls. Therefore desks are underused. Therefore reduce the desk count and reallocate the space to collaboration. Therefore introduce booking, sensors and neighbourhoods. Therefore justify the office as a place for connection, culture and mentoring, and require #return_to_office attendance on set days to make connection happen. Each step in this chain is plausible, and the underlying observations are well documented. #remote_work has stabilised at levels far above the pre pandemic baseline in many economies (Barrero, Bloom and Davis, 2023). Well designed hybrid arrangements have been shown, in at least one large randomised study, to improve retention without harming performance (Bloom, Han and Liang, 2024). At the same time, large scale evidence indicates that fully remote work can make collaboration networks more static and siloed, which gives a genuine reason to value co presence (Yang et al., 2022). Practitioner oriented research on hybrid design stresses that outcomes depend on deliberate choices about which activities benefit from the office and which do not (Gratton, 2022; Hopkins and Bardoel, 2023). Studies of what would actually attract people back point to concrete attributes such as travel time, autonomy over where to sit, and the quality of the setting rather than to slogans about culture (Appel-Meulenbroek et al., 2022). The institutional point is that this reasoning arrived everywhere at once, in nearly identical language, and produced nearly identical designs, across organizations whose work, workforces and cities differ enormously. The convergence is far greater than the underlying variation in circumstances would predict, which is the signature of #mimetic_isomorphism rather than of independent optimisation. And the most contested element, the mandated attendance day, spread through the field with striking speed once a few prominent firms adopted it, which is a textbook diffusion pattern through a visible and admired reference group. The #hot_desking element also produces a predictable friction. If the office is justified as a place for team connection, but desks are unassigned and booked individually, teams may not sit together, and the stated purpose is undermined by the operating model. The #neighbourhood_model is the field's attempt to patch this contradiction. The contradiction exists because two templates, one about cost and one about culture, were adopted simultaneously without being reconciled. 6. Outcomes of Convergence 6.1 Structural convergence and the narrowing of the design imagination The first outcome is simply that the range of thinkable options narrows. Once a template becomes the field norm, alternatives are not merely rejected; they stop being proposed. A design team that suggests assigned desks with high partitions for a team of analysts who spend eighty percent of their day in concentrated individual work will be asked whether they are behind the times. The burden of proof has inverted. Evidence is demanded of the deviant option and waived for the standard one. This inversion is the practical meaning of DiMaggio and Powell's iron cage, and it is the point at which #convergence becomes costly rather than merely tidy. 6.2 Decoupling and the gap between promise and practice The second outcome is decoupling. Three varieties appear in workplace design. Rhetorical decoupling occurs when the language used to justify a design bears no relation to its operation. A floor is described as a collaboration hub and is used, in practice, as a quiet library by people who came in to escape their homes. Operational decoupling occurs when the supporting practices required by a design are never implemented. An activity based environment requires clear protocols, adequate enclosed space, leaders who model the behaviour, and enough settings that switching is realistic. Where these are absent, people revert to territorial behaviour, informally claiming desks and defending them, and the model fails while the furniture remains. Evaluative decoupling occurs when the outcomes claimed for a design are never measured. Firms that would never approve a capital project without a financial case will approve a workplace project justified by innovation, culture and collaboration, and then never measure innovation, culture or collaboration. If the claim is never tested, it can never be falsified, and the template survives indefinitely. This is means ends decoupling in Bromley and Powell's sense, and it is arguably the central pathology of the field. 6.3 Uneven distribution of costs and benefits The third outcome concerns who wins and who loses. Convergence is not neutral in its effects. Senior staff typically retain more control over where they sit and when they attend, and often retain enclosed space even in offices that are formally open. #knowledge_workers in junior roles bear more of the noise, the desk booking, the commuting and the visibility. Employees with caring responsibilities, disabilities, sensory sensitivities, or health conditions are affected disproportionately by unassigned seating, because the loss of a stable, adjusted, personal workstation is not a minor inconvenience but a removal of an accommodation. A standard that treats all workers as interchangeable occupants of interchangeable settings is not inclusive simply because it is uniform. Uniformity and equity are different things, and the literature on perceived fit suggests that variance in individual needs, rather than an average, is what determines whether an environment works (Hoendervanger et al., 2019; Bergefurt et al., 2022). There is also a labour dimension. Desk booking systems, occupancy sensors and badge data generate detailed records of where people are and when. Even where the stated purpose is space planning, the infrastructure of #surveillance is created, and the question of who may access it and for what purpose is often settled after installation rather than before. Research on control in new modes of organizing warns that visibility and freedom are entangled rather than opposed, and that the relaxed office can be a more, not less, controlled environment (de Vaujany et al., 2021; Hafermalz, 2021). 6.4 Environmental and material cost Convergence has a physical price. Refitting an office to the current template typically means stripping out an interior that was itself installed to meet the previous template. Ceilings, partitions, carpets, joinery and furniture are removed and replaced on cycles far shorter than their technical life. The #embodied_carbon of a fit out is significant, and doing it repeatedly in pursuit of a design language undermines the sustainability claims that the new design is meant to demonstrate. There is a genuine institutional irony here: the pursuit of visible #sustainability can generate invisible waste. A serious response would treat retention, adaptation and long life loose fit design as the sustainable default, but these choices produce no photograph and no launch event. 6.5 The global asymmetry Finally, convergence is directional. Templates travel outward from a small number of design centres, mostly in high income economies, and are received elsewhere. A template developed for a temperate climate, a car free downtown, a particular labour law regime and a particular set of cultural expectations about hierarchy and privacy is exported as universal best practice. In settings with extreme heat, unreliable public transport, different family and prayer obligations, different norms about status and enclosure, and different levels of investment in acoustic and mechanical systems, the same design produces different results. This is not an argument that design ideas should stay where they were born. It is an argument that transfer without translation is a failure of professional judgement, and that the field's incentive structures reward transfer and punish translation, because transfer is fast, cheap, defensible and looks like the reference project in the brochure. Sklair's analysis of iconic architecture as an instrument of capitalist globalization is instructive here: the built form is a carrier of ideology as well as of function, and its universality is a claim rather than a fact (Sklair, 2017). The #global_south dimension of workplace research remains strikingly thin, and this is one of the clearest gaps in the literature. 7. Resistance, Translation and Divergence #isomorphism is a tendency, not a law. Organizations are not helpless, and the theory itself has always allowed for strategic response (Oliver, 1991). Four patterns of departure are worth identifying, because they are where the interesting research lies. Editing. Ideas change as they move. Organizations rarely adopt a template whole; they select the elements that fit, rename them, and attach them to local narratives (Sahlin and Wedlin, 2008; Ansari, Fiss and Zajac, 2010). A firm may adopt varied settings but keep assigned desks, or adopt unassigned seating for mobile roles only. This partial adoption is often described as failure by consultants and as pragmatism by employees. It is usually the latter. Countervailing coercion. Where labour law, collective agreements or works councils give employees a formal voice in workplace change, the imported template is negotiated rather than installed. In several European jurisdictions, workplace change of this kind is a matter for co determination, which slows adoption, forces evidence and often preserves individual workstations. This is a strong test of P6, and a comparative study of adoption rates across jurisdictions with different consultation rights would be a valuable contribution. Professional counter movements. Within the design and research community there is a visible reaction against uncritical adoption, expressed in the growing attention to acoustics, privacy, neurodiversity, individual fit and the right to concentrate. The literature on perceived fit and on the office as a resource for mental health is, in effect, a professional counter argument to density (Hoendervanger et al., 2019; Bergefurt et al., 2022). Whether it becomes a new orthodoxy or a genuine pluralism is an open question. Genuine #divergence. Some organizations have work that simply cannot be fitted into the template. Laboratories, trading floors, control rooms, clinical settings, secure environments and manufacturing offices retain distinctive forms because the technical core resists. Instructively, these are precisely the settings where the technical requirements are unambiguous. Where the task speaks clearly, institutional pressure is weaker. This observation is theoretically important: it suggests that isomorphism colonises the spaces where measurement is hardest, which is most of white collar work. 8. Discussion and Implications 8.1 Theoretical implications Three theoretical points follow. First, materiality changes the dynamics of isomorphism. Most institutional research studies practices, policies and structures, which can be adopted cheaply and abandoned quietly. A workplace cannot. It requires capital, a lease and years of commitment. This means that in workplace design the cost of ceremonial adoption is unusually high, and the lag between the fading of a trend and the disappearance of its physical residue is unusually long. Organizations are therefore governed for years by the fashion that prevailed at the moment of their last fit out. This is a distinctive contribution to the isomorphism literature, which has paid limited attention to the durability of the artefacts through which conformity is expressed. Second, visibility should be treated as a variable in diffusion research, not a constant. The article proposes, in P1, that the visibility of a feature predicts its rate of adoption independent of its efficacy. If supported, this would refine mimetic theory by specifying what gets copied, and not merely that copying occurs. Third, the field illustrates the entanglement of the three mechanisms with unusual clarity. A rating system starts as a normative professional project, becomes a mimetic status marker, and ends as a coercive financing condition. Tracing that trajectory empirically, using the history of a single certification scheme, would be an excellent doctoral project and would speak directly to the concept evolution debates in the institutional literature (Boxenbaum and Jonsson, 2017). 8.2 Implications for managers and workplace professionals Practical guidance follows from the analysis rather than from opinion. Separate the cost case from the culture case. If desk reduction is being pursued for savings, say so. Bundling it with a story about collaboration invites cynicism and guarantees that the collaboration claim will never be tested. Employees are rarely deceived by the bundling, and the credibility lost is greater than the money saved. Measure what you claim, and treat #productivity claims as testable. If a design is justified by collaboration, define what collaboration means in your organization, measure it before and after, and publish the result internally. This single discipline would eliminate most evaluative decoupling. It is also, in most organizations, entirely feasible. Design for the distribution, not the average. The evidence on perceived fit indicates that the same setting works very differently for different people and tasks. An environment that is tolerable on average may be unusable for a substantial minority. Provide genuine choice, including the choice of enclosure and stability, and treat that provision as a core requirement rather than an exception granted through a medical process. Ask where the template came from. Before adopting a peer's design, establish the conditions that made it work there, and check whether those conditions exist here. This is the discipline of translation, and it is the professional skill that the field most badly needs. Consider the lease as an institutional commitment. The question is not only whether a design is right today but whether it will still be defensible when the current trend has passed. Adaptable, loose fit, lightly serviced space is the hedge against your own future embarrassment. 8.3 Implications for students and researchers For students, this topic is unusually rich, because the phenomenon is physically accessible. A dissertation does not require a large budget. Photographic and floor plan comparison of offices within one city and one sector can document convergence directly. Interviews with the designers, real estate managers and human resource leaders who made the decisions can reconstruct the reasoning and identify the reference organizations they named. Analysis of award submissions, design magazine features and corporate recruitment pages can map the circulation of images and claims. Occupancy and satisfaction data, where accessible, can test whether the promised outcomes materialised. Four concrete study designs are worth naming. A comparative case study of two offices built from the same group standard in different climates. A content analysis of the justifications given for workplace change in a sample of annual reports over a decade, tracing the shift from cost to culture to wellbeing to hybrid. A network study of which organizations are cited as models in design briefs. A quasi experimental before and after study of a single relocation, measuring the specific outcome the organization itself claimed. 8.4 Limitations This article is conceptual. Its propositions are untested. Its evidence base is skewed toward the settings where workplace research is published, and it therefore risks reproducing the very asymmetry it criticises. It treats the templates as ideal types, which underplays the substantial variation within each. It also cannot resolve the causal question at the heart of the matter, namely how much of the observed similarity is institutional and how much reflects genuine convergence in the nature of knowledge work itself, which has, after all, become more digital, more mobile and more collaborative almost everywhere. A fair reading must concede that some convergence is functional. The argument here is narrower: that the mechanisms of the field are structurally unable to tell the functional convergence from the ceremonial kind, because they do not test. 9. Conclusion Offices look alike because the pressures that produce them are alike. Regulation and corporate standards set boundaries. Uncertainty about what actually makes knowledge work productive drives organizations to copy the peers they admire. A dense and mobile professional community, with shared training, shared conferences, shared benchmarks and shared suppliers, carries a common definition of good practice across borders. Certification schemes convert contested judgements into scores, and scores travel faster than arguments. Images travel fastest of all. The result is a global template that arrives with a strong narrative and weak evidence, and that is adopted in its visible half far more reliably than in its demanding half. The cost is borne unevenly, mostly by the people who have the least say in the design and who spend the most time inside it. The environmental cost is borne by everyone, and is quietly increased by the cycle of refitting that convergence requires. None of this means that global ideas should be refused. It means they should be interrogated. The question a workplace professional should ask is not what leading organizations are doing, but what the work in front of them actually requires, and whether the imported answer can survive contact with local climate, local law, local culture and local tasks. That is a harder question, and it produces less impressive photographs. It also produces offices that people can work in. Institutional theory teaches that organizations rarely escape their fields. But it also teaches that fields are made by their members, and that the standards which feel inevitable today were argued into existence within living memory. The office is not a law of nature. It is a set of decisions, most of which were taken elsewhere, by people with different problems. Recovering the ability to take those decisions locally, on evidence, is the practical project that this analysis recommends. References Altomonte, S., Allen, J., Bluyssen, P.M., Brager, G., Heschong, L., Loder, A., Schiavon, S., Veitch, J.A., Wang, L. and Wargocki, P. (2020) Ten questions concerning well being in the built environment. Building and Environment, 180, 106949. Alvesson, M. and Spicer, A. (2019) Neo institutional theory and organization studies: a mid life crisis? Organization Studies, 40(2), pp. 199 to 218. Ansari, S.M., Fiss, P.C. and Zajac, E.J. (2010) Made to fit: how practices vary as they diffuse. Academy of Management Review, 35(1), pp. 67 to 92. Appel-Meulenbroek, R. and Danivska, V. (eds.) (2021) A Handbook of Theories on Designing Alignment Between People and the Office Environment. London: Routledge. Appel-Meulenbroek, R., Kemperman, A., van de Water, A., Weijs-Perree, M. and Verhaegh, J. (2022) How to attract employees back to the office? A stated choice study on hybrid working preferences. Journal of Environmental Psychology, 81, 101784. Aroles, J., Mitev, N. and de Vaujany, F.X. (2019) Mapping themes in the study of new work practices. New Technology, Work and Employment, 34(3), pp. 285 to 299. Barrero, J.M., Bloom, N. and Davis, S.J. (2023) The evolution of working from home. Journal of Economic Perspectives, 37(4), pp. 23 to 49. Bergefurt, L., Weijs-Perree, M., Appel-Meulenbroek, R. and Arentze, T. (2022) The physical office workplace as a resource for mental health: a systematic scoping review. Building and Environment, 207, 108505. Bernstein, E.S. and Turban, S. (2018) The impact of the open workspace on human collaboration. Philosophical Transactions of the Royal Society B, 373(1753), 20170239. Bloom, N., Han, R. and Liang, J. (2024) Hybrid working from home improves retention without damaging performance. Nature, 630, pp. 920 to 925. Bouncken, R.B. and Reuschl, A.J. (2018) Coworking spaces: how a phenomenon of the sharing economy builds a novel trend for the workplace and for entrepreneurship. Review of Managerial Science, 12(1), pp. 317 to 334. Boxenbaum, E. and Jonsson, S. (2017) Isomorphism, diffusion and decoupling: concept evolution and theoretical challenges. In: Greenwood, R., Oliver, C., Lawrence, T.B. and Meyer, R.E. (eds.) The SAGE Handbook of Organizational Institutionalism, 2nd edition. London: Sage, pp. 79 to 104. Bromley, P. and Powell, W.W. (2012) From smoke and mirrors to walking the talk: decoupling in the contemporary world. Academy of Management Annals, 6(1), pp. 483 to 530. Cnossen, B. and Bencherki, N. (2019) The role of space in the emergence and endurance of organizing: how independent workers and material assemblages constitute organizations. Human Relations, 72(6), pp. 1057 to 1080. Dale, K. and Burrell, G. (2008) The Spaces of Organisation and the Organisation of Space: Power, Identity and Materiality at Work. Basingstoke: Palgrave Macmillan. Danivska, V. and Appel-Meulenbroek, R. (eds.) (2021) A Handbook of Management Theories and Models for Office Environments and Services. London: Routledge. de Vaujany, F.X., Leclercq-Vandelannoitte, A., Munro, I., Nama, Y. and Holt, R. (2021) Control and surveillance in work practice: cultivating paradox in new modes of organizing. Organization Studies, 42(5), pp. 675 to 695. DiMaggio, P.J. and Powell, W.W. (1983) The iron cage revisited: institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), pp. 147 to 160. Engelen, L., Chau, J., Young, S., Mackey, M., Jeyapalan, D. and Bauman, A. (2019) Is activity based working impacting health, work performance and perceptions? A systematic review. Building Research and Information, 47(4), pp. 468 to 479. Gratton, L. (2022) Redesigning Work: How to Transform Your Organisation and Make Hybrid Work for Everyone. London: Penguin Business. Haapakangas, A., Hallman, D.M., Mathiassen, S.E. and Jahncke, H. (2018) Self rated productivity and employee well being in activity based offices: the role of environmental perceptions and workspace use. Building and Environment, 145, pp. 115 to 124. Hafermalz, E. (2021) Out of the panopticon and into exile: visibility and control in distributed new culture organizations. Organization Studies, 42(5), pp. 697 to 717. Hoendervanger, J.G., Van Yperen, N.W., Mobach, M.P. and Albers, C.J. (2019) Perceived fit in activity based work environments and its impact on satisfaction and performance. Journal of Environmental Psychology, 65, 101339. Hopkins, J. and Bardoel, A. (2023) The future is hybrid: how organisations are designing and supporting sustainable hybrid work models in post pandemic Australia. Sustainability, 15(4), 3086. Kingma, S. (2019) New ways of working: work space and cultural change in virtualizing organizations. Culture and Organization, 25(5), pp. 383 to 406. Licina, D. and Yildirim, S. (2021) Occupant satisfaction with indoor environmental quality, sick building syndrome and self reported productivity before and after relocation into WELL certified office buildings. Building and Environment, 204, 108183. Lounsbury, M., Steele, C.W.J., Wang, M.S. and Toubiana, M. (2021) New directions in the study of institutional logics: from tools to phenomena. Annual Review of Sociology, 47, pp. 261 to 280. Marzban, S., Candido, C., Mackey, M., Engelen, L., Zhang, F. and Tjondronegoro, D. (2023) A review of research in activity based working over the last ten years: lessons for the post COVID workplace. Journal of Facilities Management, 21(3), pp. 313 to 333. Meyer, J.W. and Rowan, B. (1977) Institutionalized organizations: formal structure as myth and ceremony. American Journal of Sociology, 83(2), pp. 340 to 363. Munir, K.A. (2020) Challenging institutional theory's critical credentials. Organization Theory, 1(1), pp. 1 to 10. Ocasio, W. and Gai, S.L. (2020) Institutions: everywhere but not everything. Journal of Management Inquiry, 29(3), pp. 262 to 271. Oliver, C. (1991) Strategic responses to institutional processes. Academy of Management Review, 16(1), pp. 145 to 179. Sahlin, K. and Wedlin, L. (2008) Circulating ideas: imitation, translation and editing. In: Greenwood, R., Oliver, C., Suddaby, R. and Sahlin, K. (eds.) The SAGE Handbook of Organizational Institutionalism. London: Sage, pp. 218 to 242. Scott, W.R. (2014) Institutions and Organizations: Ideas, Interests and Identities, 4th edition. Thousand Oaks: Sage. Sklair, L. (2017) The Icon Project: Architecture, Cities, and Capitalist Globalization. New York: Oxford University Press. Stephenson, K.A., Kuismin, A., Putnam, L.L. and Sivunen, A. (2020) Process studies of organizational space. Academy of Management Annals, 14(2), pp. 797 to 827. Sturdy, A., Wright, C. and Wylie, N. (2016) Management as Consultancy: Neo Bureaucracy and the Consultant Manager. Cambridge: Cambridge University Press. Suddaby, R., Bitektine, A. and Haack, P. (2017) Legitimacy. Academy of Management Annals, 11(1), pp. 451 to 478. Thornton, P.H., Ocasio, W. and Lounsbury, M. (2012) The Institutional Logics Perspective: A New Approach to Culture, Structure and Process. Oxford: Oxford University Press. van Marrewijk, A. and Yanow, D. (eds.) (2010) Organizational Spaces: Rematerializing the Workaday World. Cheltenham: Edward Elgar. Vidaillet, B. and Bousalham, Y. (2020) Coworking spaces as places where economic diversity can be articulated: towards a theory of syntopia. Organization, 27(1), pp. 60 to 79. Weinfurtner, T. and Seidl, D. (2019) Towards a spatial perspective: an integrative review of research on organisational space. Scandinavian Journal of Management, 35(2), 101009. Yang, L., Holtz, D., Jaffe, S., Suri, S., Sinha, S., Weston, J., Joyce, C., Shah, N., Sherman, K., Hecht, B. and Teevan, J. (2022) The effects of remote work on collaboration among information workers. Nature Human Behaviour, 6(1), pp. 43 to 54. Hashtags #institutional_isomorphism #workplace_design #organizational_space #corporate_real_estate #office_standardization #global_corporate_trends #hybrid_workplace #activity_based_working #open_plan_debate #facility_management #organizational_theory #legitimacy_and_decoupling #employee_wellbeing #design_convergence #future_of_work

Office work has quietly become a job about managing information rather than producing it. Professionals spend a large share of the working week on scheduling, drafting, filing, reconciling, approving, chasing, and reporting. These tasks are necessary, but they rarely require the judgment that organizations actually pay knowledge workers for. The arrival of #agentic_workflows, in which software agents built on large language models plan, use tools, remember context, and act across several systems with limited supervision, promises to shift that balance. This article asks a focused question: how can agentic systems absorb complex administrative work in a way that raises, rather than hollows out, the quality of #human_strategic_oversight? The study uses a structured narrative review of peer reviewed research published between 2021 and 2025 across information systems, organization science, human computer interaction, and machine learning. It synthesizes this literature into three contributions. First, it offers a working definition of an agentic workflow that separates it from earlier #robotic_process_automation and from simple prompt based assistance. Second, it maps the empirical evidence on productivity, quality, and error, showing that gains are real but uneven, that they concentrate among less experienced workers, and that they collapse at the edges of a system's competence. Third, it proposes a five layer reference architecture and an accompanying #oversight_maturity_ladder that treats supervision as designed work with its own budget, staffing, and evidence trail, rather than as a checkbox at the end of a pipeline. The article argues that the value of agentic automation in the office is not measured by how many tasks disappear, but by how much reliable attention it returns to the humans who remain accountable. It closes with implications for managers, system designers, regulators, and students preparing for careers in an automated back office. Keywords: agentic artificial intelligence; workflow automation; administrative work; human oversight; large language model agents; organizational design; delegation; accountability 1. Introduction The modern office runs on a hidden layer of coordination. Before a contract is signed, dozens of small acts must happen: a request is logged, a template is retrieved, a clause is compared against a policy, a stakeholder is chased for approval, a number is copied from one system into another, and a record is filed where an auditor can find it later. None of these acts is difficult in isolation. Together, they consume an enormous share of #professional_time and attention. Researchers of knowledge work have described this as an overhead problem rather than a skills problem. Communication and coordination tools that were meant to make collaboration easier ended up creating a constant stream of interruptions and small obligations that fragment the day (Newport, 2021; Mark, 2023). The result is a familiar complaint in almost every organization: the people hired to think spend most of their time on #administrative_friction. Two waves of automation have tried to solve this. The first was #robotic_process_automation, which recorded and replayed keystrokes across legacy interfaces. It worked well for stable, high volume, rule based steps, and it failed whenever a screen layout changed or a document arrived in an unexpected format (Herm et al., 2023). The second wave was the generative assistant, which could write, summarize, and answer questions on demand. It was flexible, but it sat outside the workflow. A human still had to decide what to ask, paste the result somewhere useful, and take responsibility for whatever happened next. The current wave is different in kind, and #autonomy is the reason. An agentic system does not just answer. It plans a sequence of steps, calls tools and application programming interfaces, reads and writes to real systems of record, checks its own output, and adapts when something goes wrong (Yao et al., 2023; Schick et al., 2023; Sumers et al., 2024). It can be given a #goal rather than an instruction. This is what makes the concept of #delegation, long studied in organizations as a relationship between people, newly relevant as a relationship between a person and a piece of software (Baird & Maruping, 2021). Delegation changes the question that managers must answer. The old question was "which tasks can we automate?" The new question is "what are we still responsible for once the machine is doing the work?" That is a question about oversight, and it is harder than it looks. Decades of automation research warn that when systems handle the routine and hand back only the exceptions, human supervisors lose the practice, the situational awareness, and sometimes the motivation needed to catch the exceptions well. Recent empirical work in human computer interaction has confirmed a modern version of this problem: people tend to accept confident machine output without checking it, especially when checking is effortful (Buçinca et al., 2021; Passi & Vorvoreanu, 2022; Vasconcelos et al., 2023). So there is a tension at the center of this topic. Agentic workflows can free professionals from clerical drudgery. They can also quietly transfer #judgment out of human hands while leaving humans holding the legal and moral bag. The phrase #human_in_the_loop is often used as if it settles the matter. It does not. A person who clicks approve on eighty items an hour is in the loop in name only (Green, 2022; Sterz et al., 2024). This article takes the optimistic case seriously without taking it on faith. Its guiding claim is that agentic automation in the office is worth pursuing precisely because it can raise the quality of human attention, but that this outcome is a design choice, not a natural consequence of the technology. 1.1 Research questions The study is organized around four questions. RQ1. What technical and organizational features distinguish an #agentic_workflow from earlier forms of office automation? RQ2. What does credible evidence say about the effects of these systems on productivity, quality, and error in administrative work? RQ3. Which failure modes threaten these systems, and how do they differ from the failure modes of rule based automation? RQ4. What architecture and governance arrangements allow automation to strengthen rather than weaken #strategic_oversight? 1.2 Contribution and scope The article contributes a synthesis, a framework, and a set of practical instruments. The synthesis pulls together three literatures that rarely speak to each other: the machine learning work on agent architectures and benchmarks, the organizational work on augmentation and delegation, and the human factors work on trust, reliance, and oversight. The framework, presented in Section 6, describes five layers that every serious deployment needs and a maturity ladder that describes how oversight capability grows. The instruments include a task triage matrix, an oversight budget concept, and a checklist of evidence that an agentic deployment should be able to produce on demand. The scope is deliberately limited to #administrative_work in professional settings: finance operations, human resources, legal and compliance support, procurement, sales operations, project administration, and general office coordination. The article does not address consumer applications, physical robotics, safety critical control systems, or the specific case of autonomous software engineering, although evidence from software engineering is used where it illuminates general patterns. 1.3 Structure Section 2 defines the terms and reviews the shift from scripted automation to agentic systems. Section 3 sets out the theoretical framing. Section 4 explains the review method. Section 5 presents findings under six themes. Section 6 develops the proposed architecture and maturity ladder. Section 7 discusses what the findings mean. Section 8 draws out implications. Section 9 states limitations and a research agenda. Section 10 concludes. 2. Background: From Scripted Automation to Agentic Workflows 2.1 The administrative burden in knowledge work Administrative work is often treated as a residual category, the leftover that is not "real" work. That framing hides its scale and its cost. Coordination tasks are numerous, low in individual value, high in aggregate cost, and unusually corrosive to concentration because they arrive unpredictably. Empirical studies of attention in office settings show that the average professional switches tasks very frequently and takes a substantial recovery period to return to focused work after an interruption (Mark, 2023). Communication overhead is not a personal failing but a structural feature of how modern organizations coordinate (Newport, 2021). Business process management has long studied this territory, and its central insight remains valid: most administrative pain is caused by #badly_designed_processes, not by lazy people. Process mining research shows that real processes are messier than the diagrams that describe them, with rework loops, undocumented shortcuts, and long waiting times that nobody owns (van der Aalst, 2022). Any automation program that skips process discovery tends to automate the mess and make it faster. This is the first reason to be careful with #agentic_automation. A capable agent applied to a broken process produces broken outcomes at scale. 2.2 What robotic process automation could and could not do Robotic process automation deserves credit. It gave organizations a practical way to bridge systems that had no shared interface, and it delivered measurable savings in transaction heavy functions such as accounts payable, claims handling, and payroll. Its limits were #structural. The scripts were brittle, they carried a heavy #maintenance_burden, and they could only handle inputs that fell within a narrow expected pattern. Research on implementation experience found that many programs stalled not because the technology failed but because the organizational groundwork, including process standardization, governance, and change management, was missing (Herm et al., 2023). Crucially, robotic process automation could not deal with #unstructured_information. An invoice in an unusual layout, an email that contains a request buried in a paragraph of pleasantries, or a contract clause phrased in an unfamiliar way would all break the script. Since most administrative friction is precisely about unstructured information, the technology reached a ceiling. 2.3 The generative turn and its limits Large language models removed that ceiling. They could read a messy email, extract the request, summarize a contract, and draft a reply. Early field experiments demonstrated real gains on writing and support tasks (Noy & Zhang, 2023; Brynjolfsson et al., 2025). But a model that only produces text is a tool, not a worker. It does not know what happened yesterday, cannot log into the procurement system, and cannot notice that its own draft contradicts a policy document stored elsewhere. The gap between a capable #language_model and a useful colleague is a gap of #agency: the ability to plan, act, observe the result, and adjust. 2.4 Defining the agentic workflow Drawing on the technical literature (Yao et al., 2023; Schick et al., 2023; Shinn et al., 2023; Wang et al., 2024; Xi et al., 2025; Sumers et al., 2024), this article defines an agentic workflow as follows. An agentic workflow is a goal directed sequence of actions carried out by one or more software agents that reason over a task, decompose it into steps, invoke external tools and systems of record, retain and retrieve relevant context across those steps, evaluate intermediate results, and adapt the plan in response to what they observe, while operating under an explicit authority boundary set by a human principal. Five properties follow from this definition and separate agentic systems from their predecessors. Goal orientation. The system is given an outcome, not a script. "Close the month end reconciliation for the marketing cost center" rather than "click here, then here." Planning and decomposition. The system breaks the goal into sub tasks and can revise that plan mid execution. This capacity is what allows it to survive surprises (Yao et al., 2023). Tool use. The system calls external functions, databases, search interfaces, and enterprise applications, and treats their output as evidence (Schick et al., 2023). Memory. The system stores and retrieves working context, episodic records of what it did, and semantic knowledge about policies and entities (Sumers et al., 2024). Without #persistent_memory an agent cannot maintain a case across days, which is exactly what administrative work requires. Reflection and self correction. The system evaluates its own output against a criterion and retries or escalates (Shinn et al., 2023). #Reflection is the closest thing an agent has to professional self doubt, and it is imperfect. To this technical definition, the organizational literature adds a sixth property that the machine learning literature often ignores. Bounded authority. An agentic workflow in a real organization carries an explicit statement of what it may decide alone, what it must propose for approval, and what it must refuse to touch. Authority is the organizational form of autonomy, and it is the only version of autonomy that a board, an auditor, or a regulator can actually inspect (Berente et al., 2021; Chan et al., 2024). 2.5 Single agent and multi agent designs Two design families dominate. In the single agent design, one reasoning loop holds the plan and calls tools as needed. It is simpler to trace and easier to govern. In the #multi_agent_design, several specialized agents take on distinct roles, such as researcher, drafter, checker, and coordinator, and communicate with each other. Research prototypes have shown that role differentiation can improve performance on complex, decomposable tasks (Wu et al., 2024; Hong et al., 2024; Qian et al., 2024). Multi agent designs carry a governance cost that is easy to underestimate. Errors introduced by one agent can be accepted as fact by another, communication loops can consume resources without converging, and #responsibility becomes diffuse in a way that mirrors the worst pathologies of human committees. The literature on evaluation of agents warns that impressive demonstrations often do not survive rigorous, cost aware measurement (Liu et al., 2024; Kapoor et al., 2024). 2.6 A short vocabulary Because terms are used loosely in practice, this article uses the following distinctions consistently. Assistant: responds to a human request, produces content, takes no action in external systems. Script: executes a fixed sequence of deterministic steps, breaks on deviation. Agent: plans, acts through tools, and adapts within a bounded authority. #Orchestration: the coordination layer that routes work among agents, tools, and humans, and enforces the authority boundary. Oversight: the human activity of setting goals, monitoring execution, judging exceptions, and remaining accountable for outcomes. 3. Theoretical Framing 3.1 The automation and augmentation paradox The most useful starting point is the argument that automation and augmentation are not two separate strategies but two ends of a loop that organizations move through over time. Raisch and Krakowski (2021) show that firms that pursue pure automation tend to lose the human expertise that would allow them to improve the system later, while firms that pursue pure augmentation fail to capture the efficiency that would justify the investment. The paradox is that each choice undermines the conditions for the other. Applied to office work, this means that stripping humans out of an administrative process is not free even when the process runs well. The people who used to handle exceptions were also the people who noticed that the policy was out of date, that a supplier was behaving oddly, or that a form was asking the wrong question. When they leave, that #tacit_sensing capacity leaves with them. Agentic systems do not currently replace it. 3.2 Delegation to agentic artifacts Baird and Maruping (2021) provide the sharpest theoretical instrument for this topic. They argue that when software becomes agentic, the relationship between user and system stops being "use" and becomes "delegation," a transfer of rights and obligations that can flow in both directions. A human can delegate a task to an artifact, and an artifact can delegate back to a human by escalating. This framing has three consequences that recur throughout this article. First, delegation requires a clear statement of what is being delegated. Vague delegation is the root of most automation failure. Second, delegation creates a #residual_accountability that never transfers. The delegating human remains answerable, which is why oversight cannot be an afterthought. Third, delegation is dynamic. As trust develops, the boundary should move, and it should be able to move back. Static permission models are a poor fit for systems whose reliability varies by task. 3.3 Conjoined agency Murray et al. (2021) describe the ways in which human and machine agency combine, ranging from arrangements where the machine augments a human decision to arrangements where human and machine actions are so intertwined that the outcome cannot be attributed to either alone. Their concept of #conjoined_agency matters for oversight because it predicts a specific hazard: when agency is deeply intertwined, the ability to assign responsibility after a failure degrades, and so does the ability to learn from it. In practical terms, an agentic workflow that produces a decision through fifteen internal reasoning steps, three tool calls, and one human approval click has produced a conjoined outcome. If the organization cannot reconstruct who or what contributed what, it cannot correct the process. This is the argument for treating traceability as a first class requirement rather than a compliance chore (Chan et al., 2024; Novelli et al., 2024). 3.4 Oversight as work, not as posture The final theoretical move in this article is to insist that oversight is labor. It takes time, requires skill, degrades under load, and must be resourced. Green (2022) shows that policies requiring a human to supervise an algorithmic system often fail because they assume the human has the information, the time, and the incentive to disagree. Sterz et al. (2024) develop the conditions under which oversight can be effective, including the supervisor's causal power to intervene and their genuine understanding of the system's behavior. Laux (2024) argues that meaningful oversight depends on institutionalized distrust, meaning structures that expect and look for failure rather than structures that hope for success. If oversight is work, then it has a cost, and that cost must appear in the business case. An automation program that eliminates one hundred hours of clerical work and creates forty hours of unfunded supervision has not delivered what it claimed. This article calls that gap the #oversight_deficit, and it is the central practical risk identified in the review. 4. Methodology 4.1 Design This study is a structured narrative review. It follows the discipline of a systematic search, with explicit sources, search terms, and inclusion criteria, but it uses narrative synthesis rather than statistical aggregation because the underlying studies are too heterogeneous in method, unit of analysis, and outcome measure to pool. Some are randomized field experiments, some are benchmark evaluations of software, and some are qualitative organizational studies. Forcing them into a single effect size would misrepresent all of them. 4.2 Search strategy Literature was gathered from Scopus, Web of Science, the ACM Digital Library, IEEE Xplore, AIS Electronic Library, and the proceedings of major machine learning conferences. Searches combined three concept clusters. Cluster one, technology: agentic artificial intelligence, large language model agent, autonomous agent, tool use, orchestration, robotic process automation, intelligent process automation. Cluster two, work: knowledge work, administrative work, back office, business process, productivity, task allocation. Cluster three, governance: human oversight, human in the loop, accountability, delegation, trust, reliance, auditing. 4.3 Inclusion and exclusion criteria Included: peer reviewed journal articles, peer reviewed conference papers, and scholarly books published between January 2021 and the end of 2025, written in English, that address at least one of the four research questions with either empirical evidence or a substantive conceptual contribution. A small number of widely cited preprints and technical reports were included where they define an architecture or a benchmark that the peer reviewed literature builds on directly, since excluding them would have left visible gaps in the technical account. Excluded: vendor white papers, marketing material, opinion pieces without argumentation, studies of purely physical robotics, and work published before 2021 except where a concept is historically necessary, in which case the idea is described in the text rather than treated as current evidence. 4.4 Screening and coding Titles and abstracts were screened against the criteria, then full texts of the surviving items were read. Each retained source was coded against a common template covering: the phenomenon studied, the method, the setting, the reported effects, the failure modes described, the oversight mechanisms proposed, and the strength of the evidence. Coding categories were developed inductively from an initial reading of a subset of sources and then applied to the whole corpus, with categories revised twice as new themes emerged. Evidence strength was recorded on a simple three level scale. Strong evidence means a controlled or field experiment with a defined counterfactual. Moderate evidence means systematic observation, benchmark evaluation with a credible baseline, or a well constructed case study. Weak evidence means demonstration, prototype, or argument without empirical test. This grading is applied explicitly in Section 5 so that readers can see where the ground is firm and where it is not. 4.5 Synthesis Themes were built by grouping coded segments and then testing each grouping against disconfirming sources. Where the literature disagrees, the disagreement is reported rather than resolved by preference. The framework in Section 6 is a synthetic product: it is derived from the reviewed material but it is a proposal, not a finding, and it is labeled as such. 4.6 Limitations of the method Three limitations should be kept in mind while reading Section 5. The field is moving faster than the publication cycle, so peer reviewed evidence lags the systems actually deployed. This creates a bias toward the capabilities of slightly older systems. Most of the strong evidence comes from tasks that are convenient to measure, such as writing, coding, and customer support. Complex multi step administrative processes are harder to study and are therefore underrepresented in the strong evidence category. Publication and demonstration bias is likely. Failed deployments rarely produce papers, and agent benchmarks are frequently reported by the same teams that build the agents. The evaluation literature identifies this as a live problem (Kapoor et al., 2024). 5. Findings The review produced six themes. Each is presented with an explicit note on evidence strength. 5.1 Theme one: the anatomy of an agentic workflow Evidence strength: moderate for architecture, weak for comparative performance of design choices in office settings. Across the technical literature a consistent architecture appears, even where the vocabulary differs. It has six functional parts. The goal interface. This is where a human states an outcome and its constraints. It is the most neglected part of most implementations and the most consequential. An #underspecified_goal is the leading cause of confident, well executed, wrong work. Research on delegation predicts exactly this: delegation without a clear specification of rights and obligations produces misaligned effort (Baird & Maruping, 2021). The planner. The planner decomposes the goal into steps and sequences them. Interleaving reasoning with action, rather than planning everything up front, has proved more robust because the agent can incorporate what it learns from each step (Yao et al., 2023). Chain of thought style decomposition improves performance on multi step problems (Wei et al., 2022), though it also produces plausible reasoning traces that can be wrong, which has implications for oversight discussed below. The tool layer. #Tools connect the agent to reality: the enterprise resource planning system, the document store, the calendar, the ticketing system, the search index. Models can be taught to decide when to call a tool and how to use its output (Schick et al., 2023). In an office setting, the tool layer is where authority is actually enforced, because an agent can only do what its tools permit. This is why #permission_scoping at the tool boundary is more reliable than instructing the model to behave. Memory. Working memory holds the current case. Episodic memory records what was done and why. Semantic memory holds policies, definitions, and entity knowledge, often through retrieval over an organizational corpus (Sumers et al., 2024). Memory is what allows an agent to handle a purchase order that takes eleven days and four approvals, rather than only tasks that fit inside one conversation. The critic. A separate evaluation step, whether implemented as self reflection or as a distinct checking agent, catches a meaningful share of errors before they escape (Shinn et al., 2023). Its weakness is systematic: a critic that shares the same model and the same blind spots as the actor will miss correlated errors. Independent verification, ideally through a deterministic rule check or an external data source, is stronger than #self_critique. The orchestrator. This layer routes work, enforces the authority boundary, handles escalation to humans, manages retries, and writes the audit trail. In research prototypes it is often thin. In production it is the most important component in the system. Multi agent frameworks such as conversational orchestration (Wu et al., 2024), role based software teams (Hong et al., 2024; Qian et al., 2024), and simulated social agents (Park et al., 2023) demonstrate that specialization can work. What the literature does not yet establish is that multi agent designs beat well engineered single agent designs on ordinary administrative tasks once cost, latency, and traceability are counted. The honest position is that the design space is open and that complexity should be added only when a specific task demands it. 5.2 Theme two: which administrative tasks are actually suitable Evidence strength: moderate, drawn from process research, deployment case studies, and benchmark evaluation. Not all administrative work is equally automatable, and the dividing line is not difficulty. It is a combination of #verifiability, #reversibility, and #consequence. The review supports a triage of office tasks into four classes. Class A: high volume, verifiable, reversible. Examples: extracting fields from invoices and matching them to purchase orders; classifying and routing inbound requests; scheduling across constraints; drafting first versions of routine correspondence; assembling a monthly report from known data sources; reconciling two ledgers and flagging differences. These tasks have a checkable ground truth and a cheap undo. They are the natural first territory for #agentic_delegation. Class B: complex, verifiable, consequential. Examples: preparing a compliance filing; assembling a due diligence pack; drafting a contract from a clause library against a counterparty request; onboarding a new employee across six systems. Agents can do most of the assembly, but a human must certify the result because the cost of error is high and the error may not be visible in the artifact itself. These tasks should be automated to the point of proposal, never to the point of commitment. Class C: judgment bearing. Examples: deciding who to promote, whether to terminate a supplier, how to respond to a regulator, how to weigh a policy exception for a distressed customer. The administrative shell of these decisions can be automated, meaning the gathering, the summarizing, and the option laying out. The decision cannot be, and the literature on high stakes algorithmic decision support suggests that even providing a recommendation can distort the human judgment it was meant to support (Lebovitz et al., 2022; Green, 2022). Class D: unsuitable. Tasks where the input is unverifiable, the action is irreversible, and the consequence is severe, or where legal duty attaches personally to a named individual. Irreversible external actions, such as sending money, publishing a public statement, deleting records, or making a binding commitment, belong here unless a hard gate with independent verification sits in front of them. The practical rule that emerges from the review is blunt: automate the #assembly, not the #commitment. The value of an agent in the office is that it can walk the ninety percent of a process that consists of gathering, checking, formatting, and chasing, and then present a human with a decision that is properly prepared. That is not a small prize. It is the difference between a professional who spends the day preparing decisions and one who spends the day making them. 5.3 Theme three: what the evidence says about performance Evidence strength: strong for a small number of task types, weak for full agentic workflows in production. This is the theme where discipline matters most, because enthusiasm has outrun measurement. The strongest evidence concerns generative assistance rather than fully agentic execution. In a randomized experiment on professional writing tasks, access to a generative model reduced time taken by roughly forty percent and improved rated output quality, with the largest gains going to participants who had scored lower at baseline (Noy & Zhang, 2023). In a large field study of customer support, access to a model based assistant increased issues resolved per hour by roughly fourteen percent on average, and the effect was concentrated among novice and lower skilled workers while experienced workers gained little (Brynjolfsson et al., 2025). A field experiment with management consultants found significant gains inside the boundary of what the model does well, and a measurable increase in wrong answers on a task designed to sit just outside that boundary, where consultants using the model were more likely to accept a misleading conclusion (Dell'Acqua et al., 2023). Three findings recur across these studies and should be treated as the empirical backbone of this article. Finding one: gains are real and often large on tasks the system handles well. This is not #hype. Time savings in the range of a third to a half on drafting and structured output tasks are well documented. Finding two: gains compress skill differences. #Novices improve more than experts. This has implications for training pipelines that Section 5.6 develops, because the same effect that lifts a novice's output may prevent that novice from ever becoming an expert. Finding three: performance falls off a cliff at the edge of competence, and humans do not see the cliff. The #jagged_frontier finding, that a system's ability is uneven in ways that are invisible from the outside, is the single most important practical result in this literature (Dell'Acqua et al., 2023). Users cannot tell from the fluency of an output whether they are inside or outside the region of competence. At the task exposure level, analysis of occupational task data suggests that a very large share of the workforce has at least some tasks that could be meaningfully affected, with administrative and clerical roles among the most exposed (Eloundou et al., 2024). Macroeconomic assessments urge caution about translating task level exposure into rapid aggregate productivity growth, pointing out that many affected tasks are hard to automate reliably and that measured effects at the level of the firm and the economy are likely to be far smaller and slower than the task counts suggest (Acemoglu, 2025). For agentic systems specifically, the evidence base is thinner and more sobering. Benchmarks designed to test agents in realistic computer environments, including web navigation and end to end operating system tasks, show that leading systems complete only a minority of tasks that humans complete easily (Zhou et al., 2024; Xie et al., 2024; Liu et al., 2024). Even on software engineering tasks drawn from real repositories, resolution rates were low when the benchmark was introduced, and although they have since improved substantially, the pattern of uneven, hard to predict competence persists (Jimenez et al., 2024). Critically, methodological work shows that many reported agent results are inflated by weak baselines, absent cost accounting, and evaluation setups that leak information (Kapoor et al., 2024). The responsible summary is therefore narrow. Agentic workflows are demonstrably useful for structured, verifiable, bounded administrative tasks. They are not yet reliable autonomous performers of long, consequential, multi system processes without supervision, and any business case that assumes otherwise is a bet rather than a plan. 5.4 Theme four: failure modes Evidence strength: strong for individual failure modes, moderate for their interaction. Agentic systems fail differently from scripts. A script fails loudly and stops. An agent fails quietly and continues. Fabrication. Language models produce fluent, confident content that is not grounded in fact (Ji et al., 2023; Huang et al., 2025). In an office context this looks like an invented policy reference, a plausible but wrong invoice number, or a summary that asserts something the source document never said. Retrieval reduces but does not eliminate the problem, particularly when the retrieved evidence is ambiguous. Error compounding. In a multi step workflow, a small error at step two becomes an input at step three. Because downstream steps treat upstream output as evidence, the system's own mistakes are laundered into facts. This is the failure mode most specific to agentic designs, and it is why step level verification matters more than final output review. Reasoning that looks better than it is. A readable reasoning trace increases human trust whether or not it reflects the actual computation. Explanations can help people calibrate, but poorly designed explanations increase overreliance by giving the appearance of justification (Bansal et al., 2021; Vasconcelos et al., 2023). An audit trail that only records the agent's stated reasoning is therefore weaker evidence than an audit trail that records the tool calls, the inputs, and the outputs. Prompt injection and data poisoning. An #agent that reads external content, such as an inbound email, a supplier PDF, or a web page, can be given instructions by that content. Indirect prompt injection has been demonstrated against realistic application setups and remains an open problem (Greshake et al., 2023; Perez & Ribeiro, 2022). This is a category of risk that scripted automation simply did not have. An accounts payable agent that reads attachments is, in security terms, executing untrusted input. Excess permission. Agents accumulate credentials because credentials make them useful. The result is a piece of software with broad write access, weak identity, and no personnel file. Practitioners and researchers converge on the same recommendation: give agents narrow, task scoped, time limited, individually attributable access (Shavit et al., 2023; Chan et al., 2024). Silent drift. Underlying models are updated, prompts are edited, tools change their responses, and the process the agent was built for changes shape. Performance degrades without any alarm sounding. Without continuous evaluation against a fixed test set, an organization will discover drift through a complaint rather than a metric. Automation complacency. The human failure mode, and the oldest one: #complacency. When a system is right most of the time, supervisors stop checking, and their ability to check decays. Cognitive forcing interventions, which require the person to engage before seeing the machine's answer, reduce overreliance but impose effort that users tend to dislike, which is itself a finding worth taking seriously (Buçinca et al., 2021; Passi & Vorvoreanu, 2022). Diffused responsibility. In a multi agent design, or in a workflow where a human approves an output assembled from several machine steps, no single actor experiences themselves as the author of the decision. Conjoined agency makes attribution genuinely hard rather than merely inconvenient (Murray et al., 2021; Novelli et al., 2024). 5.5 Theme five: mechanisms that make oversight real Evidence strength: moderate. The mechanisms are well argued and partly tested, but few have been evaluated in production office settings. The review identifies eight mechanisms that distinguish oversight in substance from oversight in name. One: authority tiers. Define, per task type, whether the agent may act, must propose, or must refuse. Write this down as policy, enforce it in the tool layer, and review it on a schedule. Tiering is the operational form of bounded autonomy (Berente et al., 2021). Two: hard gates on irreversible actions. Any action that moves money, sends an external communication, deletes data, or creates a binding obligation passes through a gate that a human must open, and the gate presents the evidence needed to judge, not merely a summary. Three: verification that is independent of the model. Deterministic checks, rule engines, database lookups, and reconciliations catch errors that a language based critic will miss because they do not share its priors. The strongest workflows pair a probabilistic generator with a deterministic verifier. Four: sampling based assurance. Total review is impossible at scale and #total_trust is reckless. The alternative is statistical: review a random sample of completed items, review a targeted sample of high risk items, track the error rate, and adjust the authority tier when the rate moves. This treats the agent the way a quality function treats a production line. Five: cognitive forcing at the point of judgment. Where a human must decide, design the interface so the human forms a view before seeing the machine's recommendation, or so that the machine presents evidence and options rather than a conclusion (Buçinca et al., 2021). This costs time and it is unpopular. It is also one of the few interventions with experimental support against overreliance. Six: traceability by construction. Log the goal, the plan, every tool call with its inputs and outputs, every retry, every escalation, the model version, and the human decisions. #Visibility into what agents are doing is a precondition for governance, not a byproduct of it (Chan et al., 2024). Seven: continuous evaluation. Maintain a held out set of representative cases with known correct outcomes and run it on a schedule. Publish the results internally. An agent without a live scorecard is an unmanaged risk (Kapoor et al., 2024). Eight: an intervention path that works. Effective oversight requires that the supervisor can actually stop the system, that stopping is not career punishing, and that the supervisor understands the system well enough to know when to stop it (Sterz et al., 2024; Laux, 2024). If overriding the agent is slower and more embarrassing than going along with it, nobody will override it. These mechanisms cost money and time. That is the point. An organization that will not fund oversight has not automated a process; it has abandoned one. 5.6 Theme six: people, skills, and the shape of the office Evidence strength: moderate, mostly from organizational studies and field experiments. The last theme concerns what happens to the humans. The #residual_work is harder. If agents absorb the routine and hand back the exceptions, the human job becomes a stream of difficult, ambiguous cases with the easy ones removed. This is not a reduction in workload. It is an intensification of it, and it should be staffed accordingly. #Expertise_formation is threatened. Professionals become experts by doing large numbers of ordinary cases. The routine work that agents remove is also the training ground. If a junior analyst never reconciles a ledger by hand, they may never develop the intuition that tells them when a reconciled ledger is lying. Because assistance compresses skill gaps at the level of output (Noy & Zhang, 2023; Brynjolfsson et al., 2025), organizations may see good work from juniors while the pipeline of future experts quietly empties. The review found no strong evidence resolving this concern, and it should be treated as the most important open empirical question in the field. Occupational identity resists. Qualitative organizational research shows that professionals do not simply accept machine outputs into their practice; they interpret them through the standards and status structures of their occupation, sometimes engaging deeply and sometimes disengaging entirely (Anthony et al., 2023; Lebovitz et al., 2022). Deployment strategies that ignore this produce either quiet non use or uncritical acceptance, and both are failures. #New_roles appear. The review identifies four that recur in the literature and in practice: the process owner who defines goals and authority; the #agent_supervisor who monitors quality and handles escalation; the evaluation engineer who maintains the test sets and watches for drift; and the assurance function that audits the whole arrangement independently. Organizations that automate without creating these roles create the oversight deficit by default. Managers gain time and inherit ambiguity. The strategic promise of this technology is that a manager who is freed from status reporting can spend the recovered hours on judgment, relationships, and direction. That promise is conditional. Recovered time flows to whatever the organization measures and rewards. If the reward system still rewards responsiveness and volume, the recovered hours will be consumed by more email, not more thinking. 6. A Framework for Oversight Centered Agentic Workflows This section presents the article's synthetic contribution. It is a proposal derived from the review, not an empirical result, and it is offered as a structure that organizations and researchers can test. 6.1 The five layer reference architecture Layer one: the process layer. Before any agent is built, the process is discovered as it actually runs, not as it is documented. Process mining and observation reveal the rework loops, the shadow steps, and the waiting times (van der Aalst, 2022). The process is then simplified. This layer exists to prevent the most common failure of automation programs, which is the expensive automation of waste. Agentic capability makes this discipline more important, not less, because an adaptable agent can survive a bad process and hide its badness. Layer two: the delegation layer. Here the organization writes down, for each task type, the goal, the success criteria, the data the agent may touch, the actions it may take alone, the actions it must propose, the actions it must never take, and the conditions under which it must escalate. This is a delegation contract in the sense of Baird and Maruping (2021), and it should be readable by a manager, an auditor, and a machine. If it cannot be written, the task is not ready for delegation. Layer three: the execution layer. The agent itself: planner, tools, memory, critic, orchestrator. Design guidance from the review is conservative. Prefer the simplest architecture that solves the task. Prefer deterministic tools over model reasoning wherever a deterministic tool exists. Scope permissions at the tool boundary. Give the agent its own identity so its actions are attributable. Log everything. Layer four: the assurance layer. Independent verification, sampling based review, a live evaluation suite, drift monitoring, and a security posture that treats all ingested content as untrusted. This layer answers the question an auditor will eventually ask: how do you know it is working? Layer five: the accountability layer. A named human owner for each workflow. A defined intervention path. A record that survives staff turnover. A periodic review that can move the authority boundary in either direction. #Accountability is not an abstraction here; it is a person whose name is on the process and who has the power and the information to stop it. The layers are ordered deliberately. Most failed programs invert the order, starting with layer three because it is the fun one, and retrofitting layers one, two, four, and five under regulatory pressure later. 6.2 The oversight maturity ladder Organizations do not move from manual work to trusted autonomy in one step, and they should not try. The review supports a five rung ladder, in which each rung requires evidence produced on the rung below. Rung one: assisted. The human does the work; the agent drafts, summarizes, and retrieves. Oversight is inherent because the human is the actor. Value is modest but risk is low. This rung is where an organization learns what the technology is actually good at in its own context. Rung two: supervised execution. The agent performs the workflow end to end and presents the completed result with its evidence. Every item is reviewed before it takes effect. The purpose of this rung is not efficiency; it is #evidence_collection. The organization is building the error rate statistics that justify the next rung. Rung three: sampled autonomy. The agent acts within a defined class of tasks. A statistically meaningful sample is reviewed after the fact. Escalation rules handle low confidence and anomalous cases. The authority boundary is narrow and explicit. This is where genuine efficiency arrives. Rung four: bounded autonomy with hard gates. The agent acts broadly within a domain, but every irreversible or consequential action passes a gate. Assurance is continuous. Humans concentrate on exceptions, on the gates, and on the health of the system rather than on individual items. Rung five: strategic oversight. The human role is redefined. Individuals no longer supervise items; they supervise the workflow as a system. They set goals, tune the authority boundaries, interpret the error statistics, decide when the process itself should change, and remain accountable for outcomes. This is the rung the title of this article points toward, and it is reachable only if rungs two, three, and four have generated the evidence and the skills to support it. Two rules govern movement on the ladder. First, promotion requires evidence: an agent moves up only when its measured error rate on real cases, over a meaningful period, is below a threshold set in advance. Second, demotion must be easy: any material incident returns the workflow to a lower rung automatically, without a negotiation. #Reversibility of trust is what separates a governed system from a hopeful one. 6.3 The oversight budget The final instrument is a simple accounting discipline. For any proposed automation, estimate three quantities. The hours removed from human hands. The hours of new human work created: review, exception handling, escalation, evaluation maintenance, incident response, and the periodic audit. The residual risk, expressed as the expected cost of the errors that will get through, given a realistic error rate and the consequence of each error type. The net benefit is the first quantity minus the second minus the third. Business cases that report only the first quantity are not business cases. They are advertisements. Making the #oversight_budget explicit forces the conversation that most automation programs avoid, and it is the practical antidote to the oversight deficit identified in Section 3.4. 7. Discussion 7.1 The core argument restated The findings support a claim that is more modest than the marketing and more demanding than the skepticism. Agentic workflows can absorb a large volume of complex administrative work. They cannot yet be trusted to do it unsupervised at consequence. The value they create in the office is therefore not primarily the elimination of labor. It is the relocation of labor from assembly to judgment. That relocation only produces strategic benefit if the organization deliberately invests the recovered capacity in oversight and direction, and deliberately protects the conditions under which good judgment is formed. Put plainly: the technology gives back time. Whether that time becomes strategy or becomes more email is an organizational decision, not a technical one. 7.2 Why the jagged frontier is the central management problem Traditional automation had a legible boundary. Everyone knew what the script could and could not do, because it broke at the boundary. Agentic systems do not break at the boundary. They produce fluent, confident, well formatted output on the wrong side of it (Dell'Acqua et al., 2023). This is a new kind of managerial problem, because the usual signal of failure, which is visible failure, has been removed. Every recommendation in this article follows from that single fact. Verification must be independent because the system's own confidence is uninformative. Gates must be hard because a plausible output at the wrong moment can cause irreversible harm. Evaluation must be continuous because the boundary moves when the model or the process changes. Human skill must be maintained because a supervisor who cannot do the work cannot recognize when it has been done wrongly. 7.3 The paradox at the heart of the promise Section 3 introduced the automation and augmentation paradox. The findings sharpen it into an uncomfortable circularity. Effective oversight of an agentic workflow requires deep understanding of the underlying work. Deep understanding of the work is acquired by doing it. Agentic workflows remove the doing. Three responses appear in the literature and in practice, and none of them is fully satisfying. The first is deliberate practice: require professionals to perform samples of the automated work manually, as pilots maintain manual flying hours. It preserves skill but costs efficiency, and it will be the first thing cut under budget pressure. The second is redesigned expertise: accept that the skill required shifts from performing the task to evaluating the task, and train people accordingly. This is coherent but unproven, and it may underestimate how much evaluation depends on prior performance. The third is structural: retain a smaller cadre of deep experts and let the rest of the workforce operate at a higher level of abstraction. This is efficient and fragile, because it concentrates the organization's ability to detect failure in a few people. The review cannot adjudicate between these. It can insist that organizations choose consciously rather than drifting into the third by accident. 7.4 Comparison with the previous automation wave It is worth stating clearly what is genuinely new. Robotic process automation changed the cost of executing rules. Agentic automation changes the cost of exercising #discretion. That is a different thing, and it moves automation from the periphery of professional work into its center. The governance implication is proportionate. Rule execution could be governed by testing the rules. Discretion cannot be governed by testing alone, because the space of possible situations is not enumerable. It must be governed the way organizations govern human discretion: through bounded authority, sampling, audit, escalation, and named accountability. The most useful mental model for an agentic system in the office is not a faster script. It is a fast, tireless, capable, occasionally confidently wrong new joiner who has read every policy document and has no common sense, no stake in the outcome, and no fear of consequences. That framing explains the entire governance apparatus recommended here. Nobody would give such a person unsupervised authority over payments. 7.5 #Security is not a footnote The review found a persistent gap between how the technical security literature discusses agentic risk and how organizational deployments treat it. Indirect prompt injection means that any agent processing external content is executing instructions from parties outside the organization (Greshake et al., 2023). In an office context, that includes every inbound email, every supplier document, every scraped web page, and every uploaded attachment. The #mitigations are architectural rather than clever. Separate the channel that carries instructions from the channel that carries data. Never let content read by an agent expand that agent's permissions. Require human confirmation for actions triggered by externally sourced content. Assume that any agent with write access and an inbox will eventually be targeted. 7.6 What this means for the meaning of strategic oversight The phrase in the title deserves scrutiny. Strategic oversight is not simply oversight performed by a senior person. The review suggests it has four components. Direction: deciding what outcomes the workflow should pursue and what it must never do. Calibration: knowing, from evidence rather than impression, where the system is reliable and where it is not, and adjusting authority accordingly. Exception judgment: handling the cases the system escalates, which are by construction the hard ones. Institutional learning: reading the pattern of errors and escalations as information about the process, the policy, or the environment, and changing them. The fourth component is the one that justifies the word strategic. An organization whose humans only handle exceptions is running a repair shop. An organization whose humans read exceptions as signals about the world is learning. #Institutional_learning is the highest return use of the attention that automation returns, and it is the one most likely to be neglected because it does not show up in a productivity dashboard. 8. Implications 8.1 For managers and process owners Start with the #process, not the agent. Map how the work actually flows, remove the waste, and only then decide what to delegate. Use the task triage from Section 5.2. Automate assembly aggressively. Automate commitment never, without a gate. Write the delegation contract before writing the prompt. If the goal, the boundaries, and the escalation conditions cannot be stated in plain language, the workflow is not ready. Fund the oversight budget explicitly. Count the review hours, the evaluation hours, and the expected cost of errors in the business case. Climb the maturity ladder deliberately. Earn each rung with evidence and make demotion automatic after incidents. Protect the #training_pipeline. Decide consciously how junior professionals will acquire expertise when the routine work that used to teach them has gone. Reinvest the recovered time on purpose. Change what is measured and rewarded, or the time will disappear into low value responsiveness. 8.2 For system designers Design for #traceability from the first line of code, not as a compliance retrofit. Enforce authority at the tool boundary, not in the prompt. A model instructed not to do something is a request; a tool that cannot do something is a control. Prefer deterministic verification wherever it exists. Reserve model based judgment for what only it can do. Treat all ingested content as untrusted input. Build the interface for calibration, not for confidence. Show evidence, uncertainty, and provenance. Design against the reflex to accept a fluent answer (Bansal et al., 2021; Vasconcelos et al., 2023). Make escalation cheap and fast. If asking for help is expensive, the agent will not do it and neither will the human. 8.3 For policymakers and auditors Human oversight requirements should specify the conditions that make oversight effective, including the supervisor's information, time, competence, and power to intervene, rather than merely requiring that a human be present (Green, 2022; Sterz et al., 2024). Auditability should be defined in terms of reconstructable action logs, not model explanations, because stated reasoning is not the same as actual computation. Agent identity and permission scoping should be treated as basic controls, comparable to access management for human staff (Chan et al., 2024; Shavit et al., 2023). Regulatory frameworks that classify systems by risk, including the European approach to high risk applications, will need supplementary guidance for systems whose risk arises from composition and autonomy rather than from a single model's output. 8.4 For students and future professionals This journal serves students, so the implications for them deserve explicit treatment. The value of a professional will shift from producing outputs to specifying goals, judging evidence, and taking responsibility. All three are learnable, and none of them is taught well by watching a machine work. Learn the underlying craft even when a machine can do it. Reconcile the accounts by hand at least once. Read the whole contract at least once. The purpose is not nostalgia. It is that only someone who has done the work can reliably spot when the work has been done badly. Learn to write a #specification. The ability to state a goal, its constraints, its success criteria, and its forbidden actions precisely is becoming a core professional skill, and it is close to what lawyers and engineers have always called drafting. Learn #evaluation. Being able to design a test set, measure an error rate, and interpret it is a rarer and more valuable skill than being able to write a clever prompt. Learn to be usefully suspicious. #Calibrated_skepticism, the habit of asking what evidence supports a claim and what would show it to be wrong, is the scarce human input in a world of fluent machine output. 9. Limitations and Future Research 9.1 Limitations This is a review, not a primary study, and it inherits the weaknesses of the literature it reviews. The strong empirical evidence concerns assistive generative tools, not fully agentic workflows in production office settings, so the article's confident claims are narrower than its topic. The technical literature moves faster than peer review, so the picture of capability is necessarily slightly dated. Demonstration bias in agent research is documented and probably not fully corrected here. Most of the evidence is drawn from large organizations in high income economies, and the applicability of the framework to small firms and to public administration in resource constrained settings is untested. Finally, the framework in Section 6 is a proposal built from the literature; it has not been validated in the field, and readers should treat it as a hypothesis to be tested rather than a finding to be applied uncritically. 9.2 A research agenda Six questions follow directly from the gaps identified in this review. First, and most urgently, what happens to expertise formation? A longitudinal study following cohorts of junior professionals in heavily automated back offices, measuring not their output but their ability to detect errors over time, would answer the most important open question in the field. Second, what does effective oversight actually cost? Field studies that measure the labor of supervision, including its time, its accuracy, and its degradation under load, would allow oversight budgets to be built on data rather than guesswork. Third, do multi agent designs outperform single agent designs on real administrative processes when cost, latency, error rate, and traceability are all counted? The current evidence is dominated by demonstrations. Fourth, which interface designs produce calibrated reliance in professional settings, as opposed to laboratory tasks? Cognitive forcing works in experiments but is unpopular; the field needs designs that people will actually use. Fifth, how should authority boundaries be adjusted over time, and can the promotion and demotion rules proposed here be formalized and tested? Sixth, how are the benefits and costs distributed? Task exposure analyses suggest clerical and administrative roles are among the most affected (Eloundou et al., 2024). Whether the people in those roles are elevated to oversight work or displaced from it is not a technical question, and it will not be settled by technical means. 10. Conclusion The office has always contained two kinds of work. There is the work of preparing decisions, which is enormous, repetitive, and invisible, and there is the work of making them, which is small, difficult, and consequential. For most of the history of professional work, the first has crowded out the second. Agentic workflows offer a genuine chance to change that ratio. They can read the messy document, gather the scattered evidence, chase the missing approval, assemble the file, check it against policy, and put a prepared decision in front of a human being. The evidence reviewed here shows that they do parts of this well already, that the gains are real, that they help the inexperienced most, and that they fail in ways that are quiet, confident, and hard to see. That last property is what makes this a management problem rather than a procurement decision. A system that fails visibly can be supervised casually. A system that fails invisibly must be supervised by design: through bounded authority, hard gates on irreversible actions, verification that does not share the machine's blind spots, sampling based assurance, complete traceability, continuous evaluation, and a named person with the information and the power to stop it. The promise in the title of this article is conditional. Automating complex administrative tasks elevates human strategic oversight only if the organization treats oversight as real work, funds it, staffs it, measures it, and protects the expertise that makes it possible. Where that is done, the professional's day changes shape in the way the technology's advocates describe: less assembly, more judgment, more attention available for direction and for learning. Where it is not done, the same technology produces a thinner version of the same office, with faster output, weaker checking, and a slow erosion of the human capability that everyone assumed would still be there when it was needed. The technology, in the end, decides very little. It returns #time and it returns #risk in the same package. What an organization does with each is a choice, and it is still ours to make. References Acemoglu, D. (2025). The simple macroeconomics of AI. Economic Policy, 40(121). Anthony, C., Bechky, B. A., & Fayard, A. L. (2023). Collaborating with AI: Taking occupations seriously in the age of algorithmic technologies. Organization Science, 34(5). Baird, A., & Maruping, L. M. (2021). The next generation of research on IS use: A theoretical framework of delegation to and from agentic IS artifacts. MIS Quarterly, 45(1), 315-341. https://doi.org/10.25300/MISQ/2021/15882 Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M. T., & Weld, D. (2021). Does the whole exceed its parts? The effect of AI explanations on complementary team performance. Proceedings of the CHI Conference on Human Factors in Computing Systems. Berente, N., Gu, B., Recker, J., & Santhanam, R. (2021). Managing artificial intelligence. MIS Quarterly, 45(3), 1433-1450. https://doi.org/10.25300/MISQ/2021/16274 Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2). Bucinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1). https://doi.org/10.1145/3449287 Chan, A., Ezell, C., Kaufmann, M., Wei, K., Hammond, L., Bradley, H., Bluemke, E., Rajkumar, N., Krueger, D., Kolt, N., Heim, L., & Anderljung, M. (2024). Visibility into AI agents. Proceedings of the ACM Conference on Fairness, Accountability, and Transparency. Chan, A., Salganik, R., Markelius, A., Pang, C., Rajkumar, N., Krasheninnikov, D., Langosco, L., He, Z., Duan, Y., Carroll, M., Lin, M., Mayhew, A., Collins, K., Molamohammadi, M., Burden, J., Zhao, W., Rismani, S., Voudouris, K., Bhatt, U., ... Maharaj, T. (2023). Harms from increasingly agentic algorithmic systems. Proceedings of the ACM Conference on Fairness, Accountability, and Transparency. Dell'Acqua, F., McFowland, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Working Paper 24-013. Dumas, M., Fournier, F., Limonad, L., Marrella, A., Montali, M., Rehse, J. R., Accorsi, R., Calvanese, D., De Giacomo, G., Fahland, D., Gal, A., La Rosa, M., Volzer, H., & Weber, I. (2023). AI-augmented business process management systems: A research manifesto. ACM Transactions on Management Information Systems, 14(1), 1-19. https://doi.org/10.1145/3576047 Dwivedi, Y. K., Kshetri, N., Hughes, L., Slade, E. L., Jeyaraj, A., Kar, A. K., Baabdullah, A. M., Koohang, A., Raghavan, V., Ahuja, M., et al. (2023). Opinion paper: So what if ChatGPT wrote it? Multidisciplinary perspectives on opportunities, challenges and implications of generative conversational AI for research, practice and policy. International Journal of Information Management, 71, 102642. Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306-1308. Feuerriegel, S., Hartmann, J., Janiesch, C., & Zschech, P. (2024). Generative AI. Business and Information Systems Engineering, 66, 111-126. https://doi.org/10.1007/s12599-023-00834-7 Green, B. (2022). The flaws of policies requiring human oversight of government algorithms. Computer Law and Security Review, 45, 105681. Greshake, K., Abdelnabi, S., Mishra, S., Endres, C., Holz, T., & Fritz, M. (2023). Not what you've signed up for: Compromising real-world LLM-integrated applications with indirect prompt injection. Proceedings of the ACM Workshop on Artificial Intelligence and Security. Herm, L. V., Janiesch, C., Helm, A., Imgrund, F., Hofmann, A., & Winkelmann, A. (2023). A framework for implementing robotic process automation projects. Information Systems and e-Business Management, 21, 1-35. Hong, S., Zhuge, M., Chen, J., Zheng, X., Cheng, Y., Zhang, C., Wang, J., Wang, Z., Yau, S. K. S., Lin, Z., Zhou, L., Ran, C., Xiao, L., Wu, C., & Schmidhuber, J. (2024). MetaGPT: Meta programming for a multi-agent collaborative framework. Proceedings of the International Conference on Learning Representations. Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., & Liu, T. (2025). A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2). Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1-38. https://doi.org/10.1145/3571730 Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., & Narasimhan, K. (2024). SWE-bench: Can language models resolve real-world GitHub issues? Proceedings of the International Conference on Learning Representations. Kapoor, S., Stroebl, B., Siegel, Z. S., Nadgir, N., & Narayanan, A. (2024). AI agents that matter. Transactions on Machine Learning Research. Krakowski, S., Luger, J., & Raisch, S. (2023). Artificial intelligence and the changing sources of competitive advantage. Strategic Management Journal, 44(6), 1425-1452. Laux, J. (2024). Institutionalised distrust and human oversight of artificial intelligence: Towards a democratic design of AI governance under the European Union AI Act. AI and Society. Lebovitz, S., Lifshitz-Assaf, H., & Levina, N. (2022). To engage or not to engage with AI for critical judgments: How professionals deal with opacity when using AI for medical diagnosis. Organization Science, 33(1), 126-148. https://doi.org/10.1287/orsc.2021.1549 Liu, X., Yu, H., Zhang, H., Xu, Y., Lei, X., Lai, H., Gu, Y., Ding, H., Men, K., Yang, K., Zhang, S., Deng, X., Zeng, A., Du, Z., Zhang, C., Shen, S., Zhang, T., Su, Y., Sun, H., ... Tang, J. (2024). AgentBench: Evaluating LLMs as agents. Proceedings of the International Conference on Learning Representations. Mark, G. (2023). Attention span: A groundbreaking way to restore balance, happiness and productivity. Hanover Square Press. Mokander, J. (2023). Auditing of AI: Legal, ethical and technical approaches. Digital Society, 2(3). Mollick, E. (2024). Co-intelligence: Living and working with AI. Portfolio. Murray, A., Rhymer, J., & Sirmon, D. G. (2021). Humans and technology: Forms of conjoined agency in organizations. Academy of Management Review, 46(3), 552-571. https://doi.org/10.5465/amr.2019.0186 Newport, C. (2021). A world without email: Reimagining work in an age of communication overload. Portfolio. Novelli, C., Taddeo, M., & Floridi, L. (2024). Accountability in artificial intelligence: What it is and how it works. AI and Society, 39, 1871-1882. Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187-192. https://doi.org/10.1126/science.adh2586 Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. Proceedings of the ACM Symposium on User Interface Software and Technology. Passi, S., & Vorvoreanu, M. (2022). Overreliance on AI: Literature review. Microsoft Research Technical Report. Perez, F., & Ribeiro, I. (2022). Ignore previous prompt: Attack techniques for language models. NeurIPS Workshop on Machine Learning Safety. Qian, C., Liu, W., Liu, H., Chen, N., Dang, Y., Li, J., Yang, C., Chen, W., Su, Y., Cong, X., Xu, J., Li, D., Liu, Z., & Sun, M. (2024). ChatDev: Communicative agents for software development. Proceedings of the Annual Meeting of the Association for Computational Linguistics. Raisch, S., & Krakowski, S. (2021). Artificial intelligence and management: The automation-augmentation paradox. Academy of Management Review, 46(1), 192-210. https://doi.org/10.5465/amr.2018.0072 Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., & Scialom, T. (2023). Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems. Shavit, Y., Agarwal, S., Brundage, M., Adler, S., O'Keefe, C., Campbell, R., Lee, T., Mishkin, P., Eloundou, T., Hickey, A., Slama, K., Ahmad, L., McMillan, P., Vallone, A., Beutel, A., Passos, A. T., & Robinson, D. G. (2023). Practices for governing agentic AI systems. OpenAI Technical Report. Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., & Yao, S. (2023). Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems. Sterz, S., Baum, K., Biewer, S., Hermanns, H., Lauber-Ronsberg, A., Meinel, P., & Langer, M. (2024). On the quest for effectiveness in human oversight: Interdisciplinary perspectives. Proceedings of the ACM Conference on Fairness, Accountability, and Transparency. Sumers, T. R., Yao, S., Narasimhan, K., & Griffiths, T. L. (2024). Cognitive architectures for language agents. Transactions on Machine Learning Research. Tschang, F. T., & Almirall, E. (2021). Artificial intelligence as augmenting automation: Implications for employment. Academy of Management Perspectives, 35(4), 642-659. van der Aalst, W. M. P. (2022). Process mining: A 360 degree overview. In Process Mining Handbook, Lecture Notes in Business Information Processing, 448. Springer. Vasconcelos, H., Jorke, M., Grunde-McLaughlin, M., Gerstenberg, T., Bernstein, M. S., & Krishna, R. (2023). Explanations can reduce overreliance on AI systems during decision-making. Proceedings of the ACM on Human-Computer Interaction, 7(CSCW1). Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., Chen, Z., Tang, J., Chen, X., Lin, Y., Zhao, W. X., Wei, Z., & Wen, J. (2024). A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6). Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems. Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., Jiang, L., Zhang, X., Zhang, S., Liu, J., Awadallah, A. H., White, R. W., Burger, D., & Wang, C. (2024). AutoGen: Enabling next-generation LLM applications via multi-agent conversation. Proceedings of the Conference on Language Modeling. Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., Zheng, R., Fan, X., Wang, X., Xiong, L., Zhou, Y., Wang, W., Jiang, C., Zou, Y., Liu, X., ... Gui, T. (2025). The rise and potential of large language model based agents: A survey. Science China Information Sciences, 68(2). Xie, T., Zhang, D., Chen, J., Li, X., Zhao, S., Cao, R., Hua, T. J., Cheng, Z., Shin, D., Lei, F., Liu, Y., Xu, Y., Zhou, S., Savarese, S., Xiong, C., Zhong, V., & Yu, T. (2024). OSWorld: Benchmarking multimodal agents for open-ended tasks in real computer environments. Advances in Neural Information Processing Systems. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2023). ReAct: Synergizing reasoning and acting in language models. Proceedings of the International Conference on Learning Representations. Zhou, S., Xu, F. F., Zhu, H., Zhou, X., Lo, R., Sridhar, A., Cheng, X., Ou, T., Bisk, Y., Fried, D., Alon, U., & Neubig, G. (2024). WebArena: A realistic web environment for building autonomous agents. Proceedings of the International Conference on Learning Representations. Hashtags #AgenticWorkflows #AgenticAI #OfficeAutomation #AdministrativeAutomation #HumanOversight #StrategicOversight #FutureOfWork #KnowledgeWork #IntelligentProcessAutomation #AIGovernance #HumanAICollaboration #BusinessProcessManagement #DigitalTransformation #AIAccountability #WorkplaceAI

This article examines how entrepreneurs who live, work, and build ventures outside their country of birth use two under-priced resources to grow across borders: the relationships they hold in more than one country, and the ability to operate in more than one language. Drawing on an integrative review of scholarship in international business, migration studies, and entrepreneurship published mainly between 2019 and 2025, the paper builds a conceptual framework that links #transnational_networks and #multilingual_fluency to three outcomes that matter for #global_expansion: faster #opportunity_recognition, lower #liability_of_foreignness, and cheaper access to resources in distant markets. The framework rests on four theoretical pillars: #social_capital theory, the #mixed_embeddedness perspective, the network view of internationalisation, and the resource-based view of the firm. The analysis argues that the advantage of the #expatriate_entrepreneur is not simply that they know people in two places or speak two languages. The advantage comes from the way these two assets combine. Language turns a contact into a confidant, and a network turns a language skill into a market. The paper also examines the limits and costs of this position, including network closure, reputational spillovers, the emotional labour of constant #code_switching, and the growing role of #machine_translation in flattening the value of ordinary language skills. It closes with implications for students, founders, universities, and policymakers, and sets out an agenda for future research. The paper is conceptual and integrative rather than empirical, and its propositions are offered as testable claims rather than settled findings. Keywords: transnational entrepreneurship, migrant entrepreneurship, diaspora networks, language in international business, internationalisation, social capital, mixed embeddedness 1. Introduction A founder in Dubai sells software to buyers in Nairobi and Karachi. A Vietnamese-born engineer in Berlin builds a hardware company whose suppliers are all in Shenzhen. A Nigerian graduate of a Canadian university runs a logistics firm whose customers are Nigerian importers and whose bankers are Canadian. None of these people fit neatly into the old categories of business research. They are not classic exporters, because they are not sending goods out from a home base into a foreign market. They are not classic multinationals, because they have no headquarters and no subsidiary structure. They are something else: people whose lives are stretched across at least two countries, and whose businesses are built directly on that stretch. Scholarship has given this person several names. Some call them the migrant entrepreneur. Others prefer the #transnational_entrepreneur, or the diaspora entrepreneur, or, in this article, the #expatriate_entrepreneur. The labels differ in emphasis, but they point to the same underlying figure: a founder who is simultaneously embedded in more than one national context and who converts that double position into commercial advantage. The scale of the phenomenon is no longer marginal. Research on the United States has shown that immigrants are substantially more likely to start firms than native-born citizens, and that immigrant-founded firms account for a disproportionate share of new employment (Azoulay et al., 2022; Kerr & Kerr, 2020). Similar patterns appear in Canada, Australia, the United Kingdom, and the Gulf states, where large expatriate populations sit at the centre of trade and services. Yet business school curricula still teach internationalisation mostly as something a domestic firm does after it matures. The expatriate founder inverts that sequence. For them, the international element is not a later stage. It is the starting condition. This article asks a focused question. How do transnational networks and multilingual fluency, acting together, enable an expatriate entrepreneur to expand a business across borders more effectively than a comparable domestic founder? Three sub-questions follow from it. First, what exactly is a transnational network, and which of its features carry commercial value? Second, what does language fluency actually do inside a business process, beyond the obvious act of translating words? Third, under what conditions do these advantages fail, reverse, or become liabilities? The article makes four contributions. It brings together two literatures that have grown separately. Research on migrant and #diaspora_networks has been dominated by sociology and economic geography. Research on language in international business has been dominated by management and organisational studies. The two rarely speak to each other, even though any practising expatriate founder experiences them as one thing. It proposes a mechanism-based framework rather than a list of correlates. Instead of asserting that networks and languages "help," the paper specifies the pathways through which they help, and the boundary conditions on each pathway. It treats the disadvantages seriously. A large part of the writing on migrant business is either celebratory or pitying. Neither posture is analytically useful. This paper treats the expatriate position as a portfolio of assets and liabilities that must be actively managed. It is written for students. The language is deliberately plain. Technical terms are defined at first use. The structure follows the conventions of an indexed journal article so that readers can see how such an article is built, while the prose stays readable. The rest of the paper proceeds as follows. Section 2 defines the key concepts. Section 3 sets out the theoretical framework. Sections 4 and 5 review the evidence on networks and on language respectively. Section 6 describes the review method. Section 7 presents the integrated framework and its propositions. Section 8 illustrates the framework using documented patterns from the literature. Section 9 examines the dark side. Sections 10 and 11 draw out practical and policy implications. Section 12 states limitations and a research agenda. Section 13 concludes. 2. Conceptual Background: Who Is the Expatriate Entrepreneur? 2.1 Untangling the labels Words matter here, because the labels carry political weight that can distort analysis. In everyday speech, a wealthy Briton in Dubai is an expatriate, while a Bangladeshi construction worker in the same city is a migrant. The distinction is social, not analytical. For the purposes of this paper, the term #expatriate_entrepreneur is used in a neutral and technical sense: any person who founds or leads a venture while residing outside their country of origin, and whose venture draws in some material way on that cross-border position. This definition covers the wealthy consultant and the corner-shop owner alike. What separates them is the scale and structure of their resources, not their category of being. Within the academic literature, several overlapping terms are used. Immigrant or ethnic entrepreneurship typically studies businesses founded by migrants that serve a local co-ethnic market in the host country. The classic case is the neighbourhood grocery, restaurant, or salon. Here the cross-border element may be thin. The founder migrated once, and the business is domestic. Transnational entrepreneurship describes ventures that operate in and depend on two or more countries at once. The founder maintains active, ongoing ties in both, and the business model actually requires those ties. Sinkovics and Reuber (2021) show that this strand has grown quickly but remains fragmented across disciplinary silos. Diaspora entrepreneurship emphasises the collective dimension. It looks at how a dispersed community with a shared origin sustains commercial infrastructure, including trade routes, credit, information, and trust. Elo, Taube, and Servais (2022) note the definitional confusion in this space and argue for greater precision about who counts as a diaspora actor and on what basis. Returnee entrepreneurship covers founders who go abroad, accumulate education or experience, and then return home to start a firm, carrying foreign knowledge and contacts with them. This is sometimes described as #brain_circulation rather than brain drain. This article treats these as variants of one underlying figure rather than as rival categories. The common thread is #dual_embeddedness: the founder is simultaneously anchored in the institutional, social, and linguistic fabric of more than one place. 2.2 What makes the position distinctive A domestic entrepreneur who wants to sell abroad faces a well-documented set of costs. Zaheer's (1995) concept of #liability_of_foreignness captures them: unfamiliarity with local practice, lack of local legitimacy, discrimination by local buyers or regulators, and the sheer cost of coordinating across distance. The standard remedy is expensive. The firm hires consultants, appoints local distributors, opens a representative office, or acquires a local company. The expatriate founder starts from a different place. Part of the foreignness has already been paid for, often at high personal cost, through the experience of migration itself. They have already learned to read an unfamiliar legal system, open a bank account without a local credit history, and negotiate with people whose assumptions differ from their own. That accumulated learning is a form of #human_capital that does not appear on any balance sheet. Vandor and Franke (2016) provided experimental evidence that international experience improves the ability to spot business opportunities, because exposure to a second context supplies a stock of ideas that can be recombined with the first. This is the cognitive engine of the phenomenon. Opportunity is often not invented from nothing. It is arbitrage across contexts: a product that exists in country A and is absent in country B, a price gap, a regulatory gap, a taste gap. The person who lives in both contexts sees the gap without having to search for it. 2.3 The two focal resources This article isolates two resources for close attention. The first is the transnational network. This is the set of relationships that the founder can activate across national borders: family, former classmates, ex-colleagues, community associations, religious congregations, professional bodies, alumni groups, and online communities. These relationships carry information, credit, referrals, reputation, and labour. They are the arteries of the business. The second is multilingual fluency. This is more than vocabulary. It includes the ability to conduct negotiation, humour, complaint, apology, and persuasion in more than one language, and to sense when a silence means agreement and when it means refusal. Each has been studied on its own. The argument of this paper is that they are complements, and that studying either in isolation understates the value of both. 3. Theoretical Framework Four theoretical traditions inform the framework developed in Section 7. None alone is sufficient. Together they specify where the advantage comes from, why it persists, and when it dissolves. 3.1 Social capital theory Nahapiet and Ghoshal (1998) divided #social_capital into three dimensions, and this division remains the most useful map of the terrain. The structural dimension asks who is connected to whom. It concerns network size, density, and the presence of #structural_holes, that is, gaps between clusters that are not otherwise connected. Burt (1992) argued that the person who bridges a structural hole enjoys an information and control advantage, because they see two worlds that cannot see each other. The relational dimension asks what flows through the tie. It concerns #trust, obligation, reciprocity, and identification. A structurally identical contact can be worth a great deal or nothing at all, depending on whether the parties trust each other. The cognitive dimension asks whether the parties share codes, narratives, and meanings that make communication possible. This is where language sits, and it is the dimension most often ignored in network research. The expatriate founder is, almost by definition, a bridge across a structural hole. Their family and school ties sit in one cluster. Their professional and market ties sit in another. Under Burt's logic they should hold an advantage. But the advantage is not automatic. It becomes real only when the relational and cognitive dimensions are also satisfied. A bridge that no one trusts, or that cannot be spoken across, carries no traffic. Granovetter's (1973) distinction between #strong_ties and #weak_ties adds a further nuance. Strong ties, typically family and close friends, are good for mobilising resources under uncertainty. They will lend money without a contract and work without a wage. Weak ties are good for novel information, because they reach beyond the founder's immediate circle. Expatriate founders often hold an unusual combination: very strong ties in the origin country and rapidly acquired weak ties in the host country. This combination is the raw material of the framework in Section 7. 3.2 The mixed embeddedness perspective Kloosterman and Rath's #mixed_embeddedness perspective, developed within economic geography, holds that migrant business cannot be explained by the resources of the migrant community alone. Ethnic solidarity is real, but it operates inside a structure of markets, laws, licensing rules, and visa regimes that the entrepreneur does not control. A community with strong internal trust but no legal right to work will not generate many firms. Solano (2020), studying Moroccan entrepreneurs in Amsterdam and Milan, demonstrates the point clearly. The same national-origin community produced different business patterns in the two cities because the opportunity structures and regulatory environments differed. Sandoz et al. (2022) extend this line, arguing that transnational entrepreneurship should be understood as spatially unequal: not all borders are equally crossable, and not all passports carry the same power. For this article, mixed embeddedness supplies an essential corrective. The framework below is not a claim that expatriate founders always outperform. It is a claim about a mechanism that operates only when the surrounding structure permits it. 3.3 The network view of internationalisation Johanson and Vahlne (2009) revised their own classic model of internationalisation and concluded that the central obstacle to entering a foreign market is not psychic distance in the abstract but outsidership: being outside the relevant business network. Insidership, by contrast, is what allows a firm to learn, build trust, and commit resources. This reframing is important because it converts internationalisation from a question of geography into a question of relationships. A firm can be physically close to a market and still be an outsider. A firm can be twelve time zones away and be an insider, if the right people vouch for it. The expatriate founder is frequently an insider in a market that competitors treat as distant. That is the whole argument, expressed in one sentence. The paper's task is to specify how. 3.4 The resource-based view and dynamic capabilities The resource-based view asks which resources produce sustained advantage. The standard answer is that a resource must be valuable, rare, difficult to imitate, and hard to substitute. Transnational networks and multilingual fluency pass this test, with qualifications. They are valuable, because they lower search and transaction costs. They are rare, because most founders in any given market do not have them. They are hard to imitate, because they are built over decades of lived experience and cannot be purchased at will. A competitor can hire a translator, but cannot hire the twenty years of a childhood spent in Lagos. Substitutability is the weak point. #machine_translation and digital marketplaces substitute, imperfectly, for some of what these resources do. Section 9 treats this threat directly. The dynamic capabilities extension matters too. Holding a network is not the same as using it. The capability lies in sensing which contact matters, seizing the moment to activate it, and reconfiguring the network as the business changes. Many founders hold enormous latent #network_capital and convert almost none of it. 3.5 Boundary spanning and language Barner-Rasmussen and colleagues (2014) showed that cultural and language skills function as resources for #boundary_spanning inside multinational firms, enabling individuals to exchange, link, facilitate, and intervene across organisational borders. The insight transfers directly to entrepreneurship. The expatriate founder is a one-person boundary spanner, performing inside a venture what a large firm must build an entire department to achieve. This is the theoretical bridge between the two focal resources. Language is not merely a communication tool sitting beside the network. Language is part of what constitutes the network, because the cognitive dimension of social capital is precisely the shared code that makes a tie usable. 4. Transnational Networks: What the Evidence Says 4.1 The functions a network performs It is easy to say that networks are useful and hard to say what they actually do. The literature suggests at least six distinct functions, and they are worth separating because they have different failure modes. Information. The network supplies market intelligence that no report contains: which importer pays late, which official can be reasoned with, which competitor is about to run out of cash, which product category is quietly growing. This information is timely, specific, and free. It is also unverified, which is a risk discussed in Section 9. Referral and legitimacy. A cold approach in an unfamiliar market rarely works. An introduction from a trusted third party converts a stranger into a provisional insider. This is the practical content of #legitimacy, and Bolzani, Marabello, and Honig (2020) show that legitimacy in transnational ventures is built simultaneously at several levels, in the home community, in the host market, and among institutional gatekeepers. Capital. Formal finance frequently fails migrant founders. Banks look for collateral, credit history, and residency stability, and many expatriate founders lack all three. Community-based finance fills the gap: family loans, rotating savings groups, supplier credit extended on the basis of shared origin. This capital is patient and cheap, but it is also bounded and can run out precisely when the firm needs to scale. Labour and skill. Networks supply staff who can be trusted quickly, who accept lower initial pay in exchange for the prospect of advancement, and who share the founder's tacit assumptions about how work is done. Distribution. For firms selling physical goods, the diaspora frequently is the first distribution channel. A food producer sells first to co-ethnic shops in the host country, then to mainstream retailers once volume and shelf credibility exist. Governance. In markets where courts are slow or unreliable, community reputation substitutes for contract enforcement. A supplier who cheats a co-ethnic buyer risks losing standing in a community that matters to their family. This is a real and powerful governance mechanism, and it is the reason so much cross-border trade in weakly institutionalised markets runs through kinship and origin ties. 4.2 Structure: bonding and bridging Two structural forms of network capital should be distinguished. #bonding_ties connect the founder to people like themselves: same origin, same language, same community. These ties are dense, redundant, and high in trust. They are excellent for mobilising resources fast and for enforcing agreements informally. #bridging_ties connect the founder to people unlike themselves: host-country professionals, mainstream customers, regulators, investors, and suppliers from other communities. These ties are sparse and lower in trust but far richer in novel information and in access to large markets. The evidence suggests that ventures relying only on bonding ties tend to stay small. They serve a community market, face intense competition from other community businesses, and hit a ceiling. Ventures that build bridging ties on top of a bonding base grow larger. The bonding network supplies the resources to survive the early years; the bridging network supplies the market that allows scale. This produces the first practical rule of the paper. The bonding network is a launchpad, not a destination. A founder who never leaves the #ethnic_enclave has used only half of the transnational advantage. 4.3 Direction: the home country is not only a market Much writing treats the origin country as a place to sell into. In practice, it is more often a place to source from and to build in. The most common transnational configurations found in the literature include the following. Source at home, sell in the host country. Manufacturing, agricultural produce, textiles, and handicrafts flow from the origin country into the host market, with the founder's family or former colleagues acting as trusted procurement agents. Build at home, sell in the host country. Software development, design, accounting, and customer support are performed by teams in the origin country at lower cost, while the customer relationship, the brand, and the pricing sit in the high-income host market. This model has expanded enormously with remote work infrastructure. Learn in the host country, deploy at home. The returnee founder brings a business model, a technology, or a management practice back and adapts it. Research on returnee entrepreneurs finds a persistent pattern of #reverse_knowledge_flow, though adaptation, not copying, is what determines success. Broker between third countries. The most sophisticated version. The founder sits in country A, sources from country B, and sells to country C, using origin ties in B and host-country credibility in A to reach a market in C that neither B nor C firms could serve efficiently. Trading hubs such as Dubai, Singapore, and Istanbul are dense with such firms. 4.4 Networks and speed Firms founded by internationally experienced entrepreneurs internationalise earlier and faster. This is one of the more robust findings in the literature and connects the transnational literature to the #born_global and #international_new_venture research streams originating with Oviatt and McDougall (1994). The mechanism is straightforward. Internationalisation is expensive mostly because of information and trust costs. If the founder already possesses both, the cost collapses, and there is no reason to wait. This is why the stage models of internationalisation, which assume a slow ladder from exporting to licensing to foreign direct investment, describe the expatriate founder poorly. The ladder was built for firms with no prior foreign relationships. The expatriate founder starts on the third rung. 4.5 Networks are perishable An important and under-appreciated finding is that transnational ties depreciate. A founder who left their origin country fifteen years ago and has not maintained contact holds a network that exists mostly in memory. Contacts retire, move, or die. Regulations change. The market the founder remembers may no longer exist. Maintaining a transnational network is therefore an ongoing cost, not a one-time endowment. It requires travel, visits, gifts, attendance at weddings and funerals, and the small reciprocal favours that keep obligation alive. Founders who treat the network as a bank account they can draw on at will typically discover that the account has been closed. 5. Multilingual Fluency: More Than Translation 5.1 Language as a business infrastructure Research on language in international business, reviewed by Piekkari, Gaibrois, and Johansson (2022) and earlier by Karhunen and colleagues (2018), has moved a long way from the assumption that language is a neutral conduit through which meaning passes unchanged. The current consensus treats language as constitutive: it shapes who can speak, who is heard, who is trusted, and who is excluded. Neeley (2017) documented what happens when a firm imposes a single corporate #lingua_franca. Employees fluent in that language gain informal power far beyond their formal role. Employees who are not fluent become quiet, appear less competent than they are, and withdraw from discussions. Status and language proficiency become entangled. For the expatriate entrepreneur, this dynamic is a source of advantage. In a room where three languages are in play, the person who moves between all three controls the flow of meaning. They know what was said, what was meant, and what was left out. 5.2 The specific business functions of language Language operates at several distinct points in the value chain, and the distinctions matter. Search. A great deal of commercially relevant information is never translated. Local regulations, tender notices, trade association bulletins, industry gossip, and customer complaints exist mostly in the local language. The multilingual founder searches a larger world. Negotiation. Negotiation is where the value of true fluency is highest. A negotiator working through an interpreter loses tempo, loses nuance, and loses the ability to build rapport. They cannot make a joke, cannot register a hesitation, cannot rephrase in the moment. Hurmerinta and colleagues have shown that language shapes which opportunities are even recognised, not merely how they are executed. Trust formation. Speaking a counterpart's language, even imperfectly, signals investment and respect. It says that the speaker has spent time and effort on the relationship before any money changed hands. This is a costly signal and is read as such. Team building. A multilingual founder can hire from a wider labour pool and can manage a team that shares no single common language, by acting as the hub through which meaning passes. Brand and marketing. Marketing copy that is translated literally usually fails. Marketing that is transcreated, that is, rebuilt from the intent rather than the words, succeeds. The founder who thinks in the target language can judge the difference. 5.3 Fluency is not binary Language competence should be treated as a spectrum with several dimensions rather than as a switch. Transactional competence is enough to buy, sell, and give instructions. It is the level at which most business language training stops. Relational competence is enough to build friendship, express warmth, apologise, and repair a damaged relationship. This level is where trust is built. Institutional competence is enough to read a contract, follow a regulatory hearing, or argue a tax position. This level protects the firm from being cheated. Many expatriate founders have relational competence in the origin language and institutional competence in the host language, or the reverse. Recognising which competence one lacks is more useful than a self-assessment of overall fluency. 5.4 Code-switching and its costs #code_switching, moving between languages according to context, is a skill and also a burden. It requires constant monitoring of who is in the room, what they understand, and what should be said in which language. Research on language in multinational settings documents anxiety, fatigue, and a sense of diminished self among people forced to operate in a second language for long periods. The founder who appears effortlessly cosmopolitan may be paying an unseen cognitive tax. There is also an identity cost. Constant movement between linguistic worlds can produce a #hybrid_identity that belongs fully to neither. This is often generative for business creativity, because it feeds the recombinant thinking described in Section 2. But it can be personally isolating, and the literature on migrant founders repeatedly finds loneliness as a theme. 5.5 The machine translation question Any serious paper written in the mid-2020s must confront the possibility that language advantage is being automated away. Neural machine translation and large language models now produce translation that is adequate for most transactional purposes and improving quickly. The honest assessment is that machine translation substitutes for transactional competence and does not yet substitute for relational or institutional competence. A machine can translate an invoice. It cannot sit at a dinner and read the room. It cannot know that a particular phrase, technically correct, will be heard as an insult by a person of a particular generation from a particular region. The strategic implication is uncomfortable but clear. The commercial value of merely knowing a language is declining. The commercial value of knowing a language and a culture and a network of people who speak it is not. The advantage is migrating from the linguistic layer to the relational layer. Founders who rely on language skill alone should expect that skill to be commoditised. Founders who use language as an instrument for building relationships are on firmer ground. 6. Method This is a conceptual and integrative review paper rather than an empirical study. Its purpose is theory building, not hypothesis testing, and its claims should be read accordingly. 6.1 Approach The paper follows the logic of an integrative literature review, which is appropriate when a field is fragmented across disciplines and the goal is synthesis rather than measurement of effect size. The procedure had four steps. Step one: scoping. Three literatures were identified as relevant: transnational and migrant entrepreneurship, language in international business, and network-based theories of internationalisation. The boundaries were set deliberately wide, because the central argument of the paper is that these literatures have been kept apart to their mutual cost. Step two: selection. Priority was given to peer-reviewed work published from 2019 onward, with earlier work included only where it supplies foundational theory that has not been superseded, such as Granovetter, Burt, Nahapiet and Ghoshal, Zaheer, and Johanson and Vahlne. Preference was given to systematic reviews and to studies published in established international business, entrepreneurship, and migration journals. Step three: extraction. From each source, three things were recorded: the mechanism proposed, the boundary conditions identified, and the empirical context studied. Mechanisms were then grouped by family. Step four: synthesis. The mechanisms were arranged into the framework presented in Section 7, and propositions were derived. Propositions are stated in a form that could be tested by subsequent empirical work. 6.2 Limitations of the method The review is not systematic in the strict PRISMA sense. No exhaustive database protocol was applied, and no inter-coder reliability was calculated. The synthesis therefore reflects the judgement of the author, and a different reviewer might weight the literature differently. The literature itself is unevenly distributed. Studies of Chinese, Indian, Turkish, and Moroccan diasporas are plentiful. Studies of African, Central Asian, and Pacific diasporas are comparatively thin. Research in English dominates, which is a notable irony in a paper about multilingualism, and it means that scholarship published in Chinese, Arabic, Spanish, Russian, and other languages is underrepresented here. Conclusions should be read with that skew in mind. Finally, the field suffers from survivorship bias. Researchers study firms that exist. Expatriate ventures that failed, and the networks that failed them, are largely invisible in the record. Any claim about the advantages of the position must be discounted accordingly. 7. An Integrated Framework and Propositions 7.1 The core claim The framework advanced here holds that #transnational_networks and #multilingual_fluency are complementary resources whose joint effect on cross-border expansion exceeds the sum of their separate effects. They are complementary because each removes a constraint that limits the usefulness of the other. A network without a shared language is a list of names. The founder may know that a contact exists but cannot build the relational depth needed to convert the contact into a partner, because the cognitive dimension of social capital is missing. A language without a network is a skill without an application. The founder can read the market, write the emails, and understand the meeting, but has no one to call and no one to vouch for them. They remain an outsider in Johanson and Vahlne's sense, however fluent. Held together, the two produce something neither produces alone: the capacity to enter a market as an insider at low cost and at speed. Proposition 1. The positive effect of transnational network breadth on the speed of foreign market entry is stronger when the entrepreneur possesses relational-level fluency in the language of the target market than when they possess only transactional-level fluency. 7.2 The three mechanisms The framework specifies three mechanisms that connect the resources to the outcome. Mechanism one: opportunity recognition through contextual arbitrage. The founder holds two mental models of how markets work, one from each context. Opportunities appear at the seams between them. This mechanism runs primarily through cognition and is amplified by language, because a language carries a way of categorising the world. The bilingual founder literally has more categories available. Proposition 2. Entrepreneurs with high multilingual fluency identify a greater number of distinct cross-border opportunities than monolingual entrepreneurs with equivalent international experience, and the difference is greatest in markets where formal market information is scarce. Mechanism two: liability reduction through borrowed legitimacy. The founder does not build legitimacy from zero in the target market. They borrow it from a trusted intermediary in the network, who vouches for them. Trust is transferred along the tie. This mechanism runs primarily through relationships, and language conditions it, because the intermediary must be able to explain the founder to the target in terms the target understands. Proposition 3. The effect of transnational ties on reducing perceived liability of foreignness is mediated by the intermediary's ability to communicate in the shared language of both parties, and is weakened where the intermediary and the target do not share a language. Mechanism three: resource assembly at reduced cost. Capital, labour, suppliers, and premises are obtained through the network at prices below the market rate, because the transaction is embedded in a relationship that carries its own enforcement. This mechanism runs primarily through governance. Proposition 4. Ventures that source key inputs through bonding ties exhibit lower initial transaction costs but a lower ceiling on growth than ventures that transition to bridging ties within the first five years of operation. 7.3 Boundary conditions The framework does not claim universal advantage. Four conditions determine whether the mechanisms operate. Institutional permission. Visa status, the right to work, licensing rules, and the ability to open a bank account or hold property determine whether the founder can act at all. Where these are denied, the richest network produces nothing. This is the mixed embeddedness constraint. Network quality, not size. A network of two thousand contacts who owe nothing is worth less than a network of twenty who owe a great deal. Reciprocity, not headcount, is the operative variable. Market formality. In highly formalised markets with strong contract enforcement, transparent information, and impersonal capital markets, the network advantage shrinks, because the functions the network performs are already supplied by institutions. In markets with #institutional_voids, the advantage is large. This predicts that expatriate advantage is greatest in emerging markets and in cross-border trade between them, and smallest in, for example, purely domestic e-commerce within a high-income economy. Product complexity. For commodity goods with clear specifications, the network advantage is modest, because buying can be done at arm's length. For complex, customised, or reputation-sensitive goods and services, the advantage is large, because the buyer must trust the seller before purchase. Proposition 5. The performance advantage of transnational entrepreneurs over comparable domestic entrepreneurs increases with the level of institutional voids in the target market and with the complexity of the product being traded. 7.4 A dynamic sequence The framework also implies a sequence rather than a static position. Four stages recur across the literature. Stage one, latent. The founder possesses the network and the languages but has not yet used them commercially. Most migrants remain here permanently. Stage two, activation. A specific opportunity is recognised, usually through an existing tie. The first venture is typically small, undercapitalised, and dependent on bonding ties. Stage three, bridging. The venture reaches the limits of the community market and must acquire host-country customers, formal finance, and non-co-ethnic staff. Many firms fail at this transition. It requires the founder to convert relational capital of one type into another. Stage four, institutionalisation. The venture's cross-border capability is embedded in systems rather than in the founder's personal relationships. Contracts replace handshakes. Employees replace cousins. The firm becomes sellable and survivable. Proposition 6. Ventures that fail to move from stage three to stage four remain dependent on the founder's personal ties and exhibit lower survival rates following founder exit than ventures that institutionalise their cross-border relationships. This last proposition contains the most important practical warning in the paper. A business built entirely on who the founder knows is a business that cannot be handed on, cannot be sold at a fair price, and cannot survive the founder's illness. The final task of the successful expatriate entrepreneur is to make themselves unnecessary. 8. Illustrative Patterns from the Literature This section does not present new fieldwork. It draws on patterns that recur across published studies to show the framework in operation. The cases are described at the level of pattern rather than named firm, in order to avoid attributing claims to specific companies that the literature does not support. 8.1 The technology corridor pattern The most studied case is the movement of engineers and entrepreneurs between high-income technology clusters and their countries of origin, particularly India, China, Taiwan, and Israel. The pattern documented across this literature is consistent. Founders acquire education and work experience in the host cluster, build professional networks there, and then establish ventures that place research, development, or delivery capacity in the origin country while retaining the customer relationship, the brand, and often the legal entity in the host country. The framework explains why this configuration is stable. The founder holds bridging ties in the host country, which supply customers and #venture_capital. They hold bonding ties in the origin country, which supply engineering talent that can be recruited and trusted quickly. Language operates on both sides: the founder can pitch a host-country investor in the investor's own professional register, and can also manage an origin-country team in the language in which those engineers think most precisely. Competitors who lack the dual position must choose. A purely host-country firm can raise money and win customers but pays high salaries and struggles to manage an offshore team it does not understand. A purely origin-country firm has cheap talent but cannot reach the customers or the investors. The expatriate founder does not have to choose. 8.2 The trade and sourcing pattern Studies of migrant traders in European and Gulf cities describe a second recurring configuration. The founder builds a firm that imports goods from the origin region and sells them into the host market, first to the co-ethnic community and later, if the transition to bridging succeeds, to the mainstream market. Solano's comparison of Moroccan entrepreneurs in Amsterdam and Milan is instructive here, because it shows that the same community produced different outcomes in different regulatory and market environments. This is the mixed embeddedness boundary condition made visible. The network was the same. The opportunity structure was not. The framework predicts what determines success within this pattern. Firms that remain in the community market compete on price against other community firms and are squeezed. Firms that acquire mainstream #market_entry, typically by meeting host-country certification standards, professionalising their packaging, and hiring host-country sales staff, escape the squeeze. That escape requires host-country institutional language competence, which many founders lack and must buy. 8.3 The services and remote delivery pattern The third pattern has expanded rapidly since remote work infrastructure became normal. Here the founder builds a professional services firm, in software, design, marketing, accounting, or consulting, that sells into a high-income market while delivering from a lower-cost origin market. This model is the purest expression of contextual arbitrage. The price of the same hour of skilled work differs by a large multiple across borders. The barrier to arbitraging that gap is not technology, which is freely available, but trust and communication. The customer must believe that work performed by people they will never meet, in a country they have never visited, will meet their standard. The expatriate founder is the guarantor of that belief. They stand in the middle, credible on both sides. The framework identifies the vulnerability. If the founder's only contribution is standing in the middle, the position is fragile. #digital_platforms are working steadily to disintermediate exactly this role, matching buyers and sellers directly and supplying escrow, ratings, and dispute resolution that substitute for personal trust. The durable version of this business is one where the founder adds capability, not only introduction: proprietary method, managed quality, domain specialisation, or accountability for outcomes rather than hours. 8.4 The returnee pattern The fourth pattern involves founders who go back. Research on returnee entrepreneurs in emerging economies finds that they tend to found more technology-intensive firms and to internationalise earlier than local counterparts, but also that they frequently misjudge the home market, having modelled it on a memory rather than a current reality. The framework accounts for both findings. The returnee carries superior #knowledge_transfer and bridging ties abroad, which explains the technology intensity and the export orientation. But their bonding ties at home have depreciated during their absence, and their mental model of the home market is stale. Where the returnee reinvests in local relationships and updates their model, the advantage is realised. Where they assume the home market is as they left it, the advantage evaporates. 8.5 What the patterns share Across all four, one structural feature recurs. The founder occupies a position that is difficult to occupy from either side alone, and the business is essentially the monetisation of that position. The strategic question for any expatriate founder is therefore simple to state and hard to answer: what exactly can I do from here that neither a purely domestic firm in the host country nor a purely domestic firm in the origin country can do, and how long will that remain true? 9. The Dark Side: Limits, Costs, and Failures A framework that only identifies advantages is a marketing document, not a theory. This section sets out the counter-forces. 9.1 Network closure and the trap of the enclave The same density that makes bonding ties useful can make them suffocating. A tightly closed community network circulates the same information among the same people. Novel information stops arriving. Everyone sees the same opportunities, so everyone enters the same businesses, so margins collapse. The community becomes a source of competition rather than support. Closure also imposes obligations that damage the firm. Family members must be employed regardless of competence. Prices must be discounted for relatives. Debts owed to community members must be honoured before commercial creditors. These are not irrational behaviours; they are the price of maintaining membership in the network that the founder depends on. But they are costs, and they are frequently invisible in the founder's own accounting. 9.2 Reputational contagion Trust that travels along ties is a double-edged instrument. If a founder's reputation is vouched for by the community, then a failure damages not only the founder but the vouching intermediary and, in some cases, the community's collective standing. This raises the stakes of failure and makes some founders excessively conservative, avoiding risks that a domestic founder with no such exposure would take. There is a converse risk. If another member of the community commits fraud, the founder may suffer by association, facing suspicion from banks, buyers, and regulators who apply the reputation of the group to the individual. 9.3 Discrimination and structural constraint The literature on migrant business is unambiguous that discrimination is a material constraint, not a background irritation. Access to bank credit is harder. Access to premises in prime locations is harder. Access to public procurement is harder. Access to the informal networks where large contracts are actually decided is harder. Nothing in the framework proposed here should be read as suggesting that transnational assets compensate for this. They do not. They provide an alternative route around some obstacles, which is not the same as removing the obstacles. A founder who succeeds despite discrimination has succeeded despite it, and the counterfactual, what they would have achieved without it, is not observable but is almost certainly larger. 9.4 Political and regulatory exposure The transnational founder is exposed to two political systems, and a deterioration in relations between them is a direct business risk. Trade restrictions, sanctions, visa suspensions, currency controls, and abrupt changes in taxation can sever the connection on which the business depends. Founders operating between countries with tense relations carry a #risk_management burden that domestic founders do not. Currency risk deserves specific mention. A business that earns in one currency and pays costs in another is running a foreign exchange position whether or not the founder recognises it. Sharp devaluations in the origin country can destroy the economics of a sourcing model overnight, or, less commonly, transform them favourably. 9.5 Founder dependence As Proposition 6 states, the deepest structural weakness of the model is that the transnational capability often lives entirely inside one person's head and one person's relationships. Such a firm has low transferable value. An acquirer buying it is buying a founder who may leave. A bank lending to it is lending against relationships it cannot inspect. Institutionalising the network is difficult precisely because the relationships are personal. The founder cannot simply hand a contact list to a subordinate and expect the relationship to transfer. What can be transferred is process: documented supplier terms, formal contracts, quality systems, multiple relationships per counterparty rather than one, and a management team that includes people with their own cross-border ties. 9.6 The commoditisation of the advantage Finally, and most seriously for the long term, the underlying advantage is eroding at the margins. #machine_translation reduces the value of ordinary language skill. Digital marketplaces reduce the value of ordinary intermediation. Global logistics platforms reduce the value of knowing a freight forwarder. Payment infrastructure reduces the value of knowing how to move money. What remains scarce is judgement: knowing which supplier will actually deliver in December, knowing which regulator's approval is real and which is theatre, knowing which partner will renegotiate after signature. That judgement is built from lived experience in a specific place among specific people, and it is not obviously automatable. The expatriate founder who wishes to remain valuable should be investing in judgement, not in the mechanical tasks that were once the visible surface of their advantage. 10. Implications for Practice 10.1 For the individual founder Several practical rules follow from the framework. Audit the network honestly. List the people who would take a call, and beside each name write what they can actually do, when they were last contacted, and what they are owed. Most founders discover that their real network is much smaller than they believed and much more depreciated. Invest in the weak side. Most expatriate founders are strong in one network and weak in the other. The bonding side usually requires maintenance. The bridging side usually requires construction. Identify which is weaker and allocate time accordingly, remembering that bridging ties take years to build and cannot be manufactured under deadline. Diagnose the level of language competence, not the fact of it. A founder who can chat comfortably in a language but cannot read a contract in it is exposed. A founder who can read a contract but cannot make small talk will struggle to build trust. Buy the missing competence deliberately, whether through study, hiring, or professional advice. Do not let #translation substitute for presence. Machine translation is adequate for documents. It is inadequate for the dinner where the deal is actually decided. Design for exit from day one. Write things down. Formalise supplier agreements. Build second relationships behind every first relationship. Ensure that at least one other person in the firm can pick up any given counterparty relationship. The goal is a firm that works when the founder is on a plane, ill, or gone. Manage the currency and political exposure explicitly. Know the size of the foreign exchange position. Know what happens to the business if the border closes for six months. Have an answer before the question is asked. 10.2 For students preparing for this path For students reading this in a business programme, three points are worth stating directly. First, the resources described here are being built now, whether or not the student notices. The classmates in a lecture hall in a foreign city are the weak ties of the future. Alumni networks are the single most reliably valuable bridging structure available to a young person, and they cost nothing to maintain other than attention. Second, language learning has a long payback period and should be started early. A person who begins a third language at twenty may have relational competence by twenty-eight. A person who begins at thirty-five, while running a company, generally will not. Third, the origin network is not a birthright. It is maintained or it decays. Students who study abroad and never return, never call, and never attend the family occasions will find, ten years later, that the network they assumed they had is gone. 10.3 For universities and business schools Institutions that educate international students hold a valuable and largely unexploited asset. Alumni from a single university, dispersed across forty countries, constitute exactly the kind of bridging structure that transnational ventures require. Most universities manage this asset as a fundraising list rather than as a commercial network. Curricula also lag. Internationalisation is still frequently taught as something a large domestic firm does. The expatriate founder inverts the sequence, and teaching should reflect that. Practical instruction in cross-border contracting, currency management, remote team leadership, and #cross_cultural_communication is more useful to this population than another module on the stage model of foreign market entry. 11. Policy Implications 11.1 The regulatory chokepoints Mixed embeddedness theory tells policymakers where to look. The binding constraints on expatriate entrepreneurship are rarely motivational and rarely about skills. They are usually administrative. The recurring chokepoints identified in the literature are the right to work independently rather than only for a sponsoring employer; the ability to open a business bank account without a long domestic credit history; the ability to obtain a licence in a regulated profession without full requalification; and the ability to remain in the country if the venture takes longer than the visa to succeed. Each of these is a policy variable. Countries that have loosened them, through startup visas, self-employment routes, and simplified licensing, have generally seen increases in venture formation among their migrant populations. Countries that keep residence tied to a single employer suppress entrepreneurship among some of their most internationally connected residents. 11.2 Origin countries and the diaspora Origin-country policy has often focused on #remittances, which are large but consumption-oriented. Investment and venture creation by the diaspora is a smaller flow with larger developmental effect, because it builds productive capacity rather than funding consumption. Effective origin-country policy in this area tends to share three features. It reduces the friction of investing from abroad, including property rights, repatriation of profits, and the ability to hold accounts. It provides credible information, so that diaspora investors are not relying on a fifteen-year-old mental model. And it creates convening structures, so that diaspora professionals can meet each other and domestic firms. The counter-warning from the literature is that diaspora engagement policy is frequently symbolic. Conferences are held, ministries are created, and little changes, because the underlying obstacles to doing business are untouched. A diaspora investor faces the same courts, the same customs officers, and the same power supply as everyone else. Fixing the general business environment does more for #diaspora_networks than any dedicated diaspora programme. 11.3 Host countries and the integration frame Host-country policy commonly treats migrant business through a social integration lens, funding small grants and advisory services aimed at self-employment as an alternative to unemployment. This is valuable but incomplete. It addresses the survival business and ignores the growth business. The framework here suggests a different emphasis. The highest-return intervention is not another business plan workshop. It is the deliberate construction of bridging ties: connecting migrant founders to mainstream customers, to procurement officers, to banks, and to non-co-ethnic professional networks. That is the transition at which firms fail, and it is the transition that policy is best placed to assist, because it is precisely a matter of introductions and credibility, which public institutions can confer. 12. Limitations and a Research Agenda 12.1 Limitations The paper is conceptual. Its propositions are argued, not tested. They are stated in falsifiable form for exactly that reason. The evidence base it draws on is skewed towards a small number of well-studied diasporas and towards research published in English. It is skewed towards firms that survived. It is skewed towards men, since the literature on women transnational entrepreneurs remains comparatively thin despite growing attention. The framework also treats the founder as a single decision-maker, which simplifies away the reality that many migrant ventures are family firms in which authority is distributed and contested, often across generations and across borders. 12.2 An agenda for future research Six directions appear most promising. Measure the network properly. Most studies measure transnational ties with crude proxies such as frequency of visits home. Proper network measurement, capturing structure, content, and reciprocity, would allow the propositions in Section 7 to be tested rather than asserted. Separate the levels of language competence. Studies routinely code language as a binary or as a self-reported score. Distinguishing transactional, relational, and institutional competence would sharpen findings considerably and would allow the substitution effect of machine translation to be estimated. Study the bridging transition directly. The failure point identified in Section 7 is stage three. Longitudinal work following firms across this transition, including the ones that fail, would be the single most valuable empirical contribution available. Estimate the machine translation effect. Is the language advantage actually eroding, and if so, at what rate and in which functions? This is an empirical question with a clear research design available, and it is currently unanswered. Extend beyond the studied diasporas. African, Central Asian, Latin American, and Pacific transnational entrepreneurship deserve far more attention than they receive. Take gender and generation seriously. The second generation, born in the host country, holds a different and under-theorised position: native host-country language and legitimacy, but weaker origin-country ties and often weaker origin-language competence. Whether the transnational advantage survives into the second generation, and in what form, is an open and important question. 13. Conclusion The expatriate entrepreneur is not an exotic case at the edge of business studies. In an era of large-scale mobility, they are an increasingly ordinary form of founder, and the theories built to explain the internationalisation of large domestic firms describe them poorly. This article has argued that their advantage rests on two resources that are usually studied apart and experienced together. Transnational networks supply information, legitimacy, capital, labour, distribution, and governance across borders. Multilingual fluency supplies the code without which those networks cannot be used, and the sensitivity without which trust cannot be built. Neither is sufficient alone. Together they allow a founder to enter as an insider a market that competitors must approach as outsiders, and that is the entire commercial logic of the position. The article has also argued against romanticism. The advantage is conditional. It depends on institutional permission, on the quality rather than the size of the network, on the degree of institutional voids in the target market, and on the complexity of the product. It carries real costs: obligation, closure, reputational exposure, political risk, and the quiet fatigue of living between languages. And it is eroding at its mechanical edges, as translation and intermediation are automated. What survives automation is judgement, relationship, and trust, and these are built slowly, in specific places, among specific people. The practical lesson for the founder, and for the student who will become one, is therefore not to accumulate contacts or to collect languages, but to build the kind of relationships in which a difficult favour can be asked and granted. That is what a transnational network really is. Everything else is a contact list. The final task of such a founder is paradoxical. Having built a firm that depends on their unique position between worlds, they must then build the firm out of that dependence, so that it can grow beyond them, be handed on, and survive them. The measure of success is not how indispensable the founder becomes. It is how dispensable they manage, eventually, to make themselves. References Azoulay, P., Jones, B. F., Kim, J. D., & Miranda, J. (2022). Immigration and entrepreneurship in the United States. American Economic Review: Insights, 4(1), 71 to 88. Barner-Rasmussen, W., Ehrnrooth, M., Koveshnikov, A., & Makela, K. (2014). Cultural and language skills as resources for boundary spanning within the MNC. Journal of International Business Studies, 45(7), 886 to 905. Bolzani, D., Marabello, S., & Honig, B. (2020). Exploring the multi-level processes of legitimacy in transnational social enterprises. Journal of Business Venturing, 35(3), 105941. Burt, R. S. (1992). Structural holes: The social structure of competition. Harvard University Press. Dabic, M., Vlacic, B., Paul, J., Dana, L. P., Sahasranamam, S., & Glinka, B. (2020). Immigrant entrepreneurship: A review and research agenda. Journal of Business Research, 113, 25 to 38. Duan, C., Kotey, B., & Sandhu, K. (2021). Transnational immigrant entrepreneurship: Effects of home-country entrepreneurial ecosystem factors. International Journal of Entrepreneurial Behaviour and Research, 27(3), 711 to 729. Elo, M., Taube, F. A., & Servais, P. (2022). Who is doing transnational diaspora entrepreneurship? Understanding formal identity and status. Journal of World Business, 57(1), 101240. Granovetter, M. S. (1973). The strength of weak ties. American Journal of Sociology, 78(6), 1360 to 1380. Hajro, A., Caprar, D. V., Zikic, J., & Stahl, G. K. (2021). Global migrants: Understanding the implications for international business and management. Journal of World Business, 56(2), 101192. Johanson, J., & Vahlne, J. E. (2009). The Uppsala internationalization process model revisited: From liability of foreignness to liability of outsidership. Journal of International Business Studies, 40(9), 1411 to 1431. Karhunen, P., Kankaanranta, A., Louhiala-Salminen, L., & Piekkari, R. (2018). Let us talk about language: A review of language-sensitive research in international management. Journal of Management Studies, 55(6), 980 to 1013. Kerr, S. P., & Kerr, W. R. (2020). Immigrant entrepreneurship in America: Evidence from the survey of business owners 2007 and 2012. Research Policy, 49(3), 103918. Kerr, W. R. (2019). The gift of global talent: How migration shapes business, economy and society. Stanford University Press. Nahapiet, J., & Ghoshal, S. (1998). Social capital, intellectual capital, and the organizational advantage. Academy of Management Review, 23(2), 242 to 266. Neeley, T. (2017). The language of global success: How a common tongue transforms multinational organizations. Princeton University Press. Oviatt, B. M., & McDougall, P. P. (1994). Toward a theory of international new ventures. Journal of International Business Studies, 25(1), 45 to 64. Piekkari, R., Gaibrois, C., & Johansson, M. (2022). A review of language-sensitive research in international business: A multi-paradigmatic reading. Journal of Comparative International Management, 25(1), 144 to 174. Piekkari, R., Welch, D., & Welch, L. S. (2014). Language in international business: The multilingual reality of global business expansion. Edward Elgar. Sandoz, L., Mittmasser, C., Riano, Y., & Piguet, E. (2022). A review of transnational migrant entrepreneurship: Perspectives on unequal spatialities. Zeitschrift fur Wirtschaftsgeographie, 66(3), 137 to 150. Sinkovics, N., & Reuber, A. R. (2021). Beyond disciplinary silos: A systematic analysis of the migrant entrepreneurship literature. Journal of World Business, 56(4), 101223. Solano, G. (2020). The mixed embeddedness of transnational migrant entrepreneurs: Moroccans in Amsterdam and Milan. Journal of Ethnic and Migration Studies, 46(10), 2067 to 2085. Tietze, S. (2021). Language, translation and management knowledge: A research overview. Routledge. Vandor, P., & Franke, N. (2016). See Paris and then found a business? The impact of cross-cultural experience on opportunity recognition capabilities. Journal of Business Venturing, 31(4), 388 to 407. Zaheer, S. (1995). Overcoming the liability of foreignness. Academy of Management Journal, 38(2), 341 to 363. #expatriate_entrepreneurship #transnational_entrepreneurship #diaspora_business #migrant_founders #global_business_expansion #multilingual_advantage #cross_border_networks #international_entrepreneurship #social_capital_theory #mixed_embeddedness_theory #language_in_international_business #immigrant_innovation #brain_circulation_economy #global_talent_mobility #business_research_for_students

A new kind of company has appeared. It is small in headcount, global in reach, and it lets software make choices that used to sit with people. Prices move without a manager approving them. Credit is granted, suppliers are chosen, job applicants are ranked, and contracts are signed, all by systems that learn from data and act at speed. This article asks a simple question with a hard answer. When machines decide, who owes the duties that company law has always placed on human beings? The article develops the idea of #fiduciary_governance for what it calls autonomous businesses: firms in which a material share of operational or strategic decisions is produced or executed by artificial intelligence with limited case by case human review. It argues that the classical fiduciary core of #duty_of_care, #duty_of_loyalty, good faith, and oversight does not disappear in such firms. It migrates. It moves upstream, away from the individual transaction and into the design of objective functions, the choice of training data, the setting of autonomy limits, and the quality of monitoring. The article compares how Delaware, the United Kingdom, the European Union, Singapore, the United Arab Emirates free zones, and selected offshore regimes for decentralised organisations treat this migration. It shows that the legal question is rarely whether an AI system can be a fiduciary, and almost always whether the humans behind it discharged duties that were already there. Using doctrinal analysis and a comparative reading of statutes, cases, and recent scholarship, the article proposes a five layer #governance_stack for internationally structured startups, together with a set of drafting tools: autonomy schedules, reversibility clauses, model change control, and a documented human accountability anchor. The article closes with limits, open questions, and a research agenda. Keywords: fiduciary duty; artificial intelligence; corporate governance; algorithmic accountability; startups; comparative law; agency law; regulatory arbitrage 1. Introduction 1.1 The setting For most of the history of company law, the decision maker was a person. A director read a board paper, asked questions, argued, and voted. If the decision went badly, courts asked whether that person had been careless, disloyal, or dishonest. The whole architecture of #corporate_governance grew around that assumption. Duties attach to persons because persons have minds, and minds can be interrogated for intent, care, and honesty. That assumption is now under quiet pressure. A growing class of young companies, especially in financial technology, logistics, marketplaces, insurance, recruitment, and digital health, run large parts of their operations through machine learning systems. The systems do not merely advise. They act. They set prices in milliseconds. They approve or refuse a loan before any employee sees the file. They freeze a merchant's account. In some firms, an internal agent writes and executes code that changes the product. This is what this article means by an #autonomous_business: not a company without humans, but a company where the ordinary path of a decision does not pass through a human mind, and where human involvement is exceptional rather than routine. A firm that uses a model to recommend and then has a manager approve each recommendation is doing something familiar. A firm that lets the model act and reserves human review for edge cases is doing something new. Startups are the sharpest version of this story. They are small, so automation is the only way to serve a large market with a small team. They are international almost from birth. And they are typically governed thinly, with small boards controlled by founders and a lead investor that meet to discuss growth rather than controls. High automation plus light governance is precisely where fiduciary risk concentrates. 1.2 The problem The problem is not that the law is silent. It is that the law speaks in a vocabulary that was designed for a different fact pattern. Take the #duty_of_care. In most systems, a director must act with the care, skill, and diligence that a reasonably diligent person would exercise. Courts test this by asking what the director knew, what they asked, and how they deliberated. But when a pricing engine causes a competition law problem, there is no deliberation to inspect. There is a training pipeline, a loss function, a deployment decision made months earlier, and a monitoring dashboard nobody read. The evidence of care, if it exists, is in documentation and telemetry, not in minutes. Take the #duty_of_loyalty. Loyalty forbids a fiduciary from preferring their own interest, or a third party interest, over the interest of the beneficiary. An #optimisation_objective can encode disloyalty without any human ever forming a disloyal intention. If a model is trained to maximise short term revenue and it learns to exploit vulnerable customers, or to steer users toward the products with the highest margin regardless of suitability, the harm looks like a loyalty breach, but the mental state is missing. The self dealing is structural, not personal. Take the international dimension. A typical #international_startup might be incorporated in Delaware, run its engineering team from Eastern Europe, hold intellectual property in Ireland, serve customers in the European Union, the Gulf, and Southeast Asia, and store data across three cloud regions. The obligations that bite on its AI systems are not the obligations of a single legal system. They are a layered set: the corporate law of the place of incorporation, the AI and data law of the place where users sit, the sectoral rules of the industries served, and increasingly the extraterritorial reach of European regulation. A director sitting in one country can breach a duty in another without ever leaving their desk. 1.3 The argument in short This article makes four claims. First, that fiduciary duties do not evaporate when decisions are automated. They relocate. The relevant human decision is no longer the transaction. It is the decision to delegate, the decision about how much autonomy to grant, the decision about what to optimise for, and the decision about how to watch. This is the core idea of what the article calls #fiduciary_migration. Second, that most jurisdictions can accommodate autonomous businesses without inventing new duties, but only if courts and regulators are willing to read existing duties at the level of system design rather than individual act. The tools exist. They are simply pointed at the wrong layer. Third, that the real gap for #startup_governance is not doctrinal but institutional. Startups lack the internal machinery, the committees, the model risk functions, the audit trails, that would allow a court to find that duties were discharged. Without that machinery, even a well intentioned founder cannot prove care. Fourth, that #cross_border structuring creates a specific and under examined risk: what looks like efficient jurisdiction shopping at incorporation can become an accountability vacuum when harm occurs, because the entity that made the design decision is not the entity that faces the claim. 1.4 Structure The article proceeds from literature and method, to a conceptual model of machine autonomy, to doctrinal analysis of each fiduciary duty, to comparative frameworks, to the specific problems of cross border structures, and finally to a proposed governance model and a set of recommendations. 2. Literature Review and Conceptual Background 2.1 Fiduciary law and its core Fiduciary law is one of the oldest bodies of private law and one of the least tidy. Its unifying idea is discretion over another person's interests. Where one party holds power that can be used to affect the legal or practical position of another, and where the other cannot easily watch or protect themselves, the law imposes duties that go beyond contract. Comparative and theoretical work has consistently identified loyalty as the irreducible core, with care, good faith, confidentiality, and disclosure orbiting around it (Criddle, Miller and Sitkoff, 2019; Gold, 2021). Two features matter here. First, fiduciary duties are triggered by a relationship, not by a job title: anyone holding discretionary power over another's interests may qualify. Second, they are prophylactic. They do not merely punish harm; they forbid situations in which harm becomes likely, which is why conflicts of interest are prohibited even when no loss occurs. Both features suggest that automation does not remove the trigger. Discretion is still being exercised, simply by a system that a human built, configured, and released. 2.2 Company law meets algorithms Scholarship on algorithms in the boardroom has moved through three phases. The first phase, roughly a decade ago, asked whether an algorithm could be a director. It produced provocative work and useful conceptual mapping, but little that a practitioner could use. The second phase asked how existing duties would apply if directors relied on machine output. Petrin (2019) analysed the consequences of algorithmic management for the role of the board and argued that directors would increasingly become supervisors of systems rather than decision makers. Enriques and Zetzsche (2020) warned against what they called the technology nirvana fallacy, the belief that better data and better models would automatically produce better governance, and argued that technology tends to reallocate rather than eliminate agency problems. Armour and Eidenmuller (2020) explored the self driving corporation and concluded that the more a firm's operations are automated, the more governance attention should shift to the specification and monitoring of automated processes. Hilb (2020) mapped a spectrum from assisted to autonomous governance, which remains one of the most useful conceptual contributions in this area. The third phase, current and unfinished, is about agentic systems. Kolt (2025) provides the most developed legal treatment, arguing that AI agents reproduce the classic markers of agency relationships, information asymmetry, discretionary authority, and divided loyalty, and that conventional solutions to principal agent problems, incentive design, monitoring, and enforcement, may not work against agents that operate at superhuman speed and scale. Chan and colleagues (2023) catalogue the harms that follow from increasingly agentic systems. Gabriel and colleagues (2024) explore the normative structure of advanced AI assistants and the loyalty questions they raise. What this literature has not done is join the corporate law analysis to the regulatory analysis and then apply the result to the specific situation of thinly governed, internationally structured startups. That is the gap this article addresses. 2.3 AI regulation as a governance input Parallel to the corporate law conversation, a large regulatory literature has grown around the European Union's #EU_AI_Act. Veale and Zuiderveen Borgesius (2021) gave an early critical reading of the draft, while Almada and Petit (2025) describe the finished instrument as a hybrid of product safety and fundamental rights logic. Laux, Wachter and Mittelstadt (2024) question whether risk regulation can carry the weight of trustworthiness, Novelli and colleagues (2024) propose a refined risk assessment model, and Schuett (2023) shows how the Act's risk management obligations map onto existing enterprise risk practice. This matters for fiduciary analysis for a reason that is often missed. Regulatory standards become evidence of what a reasonable director should have done. Once a jurisdiction sets out a risk management system, a logging obligation, and a human oversight requirement, a director in that jurisdiction who ignores them is not merely non compliant. They are careless in the fiduciary sense as well. #compliance_by_design and fiduciary care converge. 2.4 Accountability, auditing and the assurance layer A third body of work concerns how one would actually check an automated system. Raji and colleagues (2020) set out an internal audit framework; Metcalf and colleagues (2021) and Selbst (2021) analyse impact assessments and their institutional limits; Mokander and colleagues (2022) connect conformity assessment to auditing practice; Cobbe, Veale and Singh (2023) show that accountability in algorithmic supply chains is fragmented across many actors, which is exactly the situation of a startup that assembles its stack from third party vendors; and Schuett (2024) adapts the banking three lines of defence model to AI organisations. This literature supplies the operational content that fiduciary law needs. Fiduciary duties tell you that you must be careful. Assurance literature tells you what careful looks like. 2.5 Autonomous organisations and legal form Finally, there is work on organisations autonomous in a different sense: decentralised autonomous organisations, or #DAO structures, which operate through smart contracts and token based voting (Hassan and De Filippi, 2021; Kaal, 2021; Santana and Albareda, 2022; Bellavitis, Fisch and Momtaz, 2023; Sims, 2021). These organisations were the first to confront, in practice, the question of who owes duties when code governs. Their experience with legal wrappers, limited liability recognition, and liability leakage to token holders offers a preview of the problems facing AI heavy firms. 3. Methodology This article uses a doctrinal and comparative method, supplemented by conceptual modelling. The doctrinal component involves close reading of company law statutes and leading cases in the selected jurisdictions, together with the relevant regulatory instruments. Because the field is young and litigation is scarce, the analysis is necessarily forward looking. Where no case exists, the article reasons by analogy from established doctrine and identifies the most probable line of judicial development. The comparative component selects legal environments on the basis of relevance to internationally structured startups: Delaware, the United Kingdom, the European Union, Singapore, the United Arab Emirates common law free zones, and selected regimes offering legal recognition to autonomous organisations. Selection is purposive rather than representative. These are the jurisdictions that appear most often in the corporate structures of the firms under discussion. The conceptual component constructs two instruments: an autonomy spectrum, used to grade machine involvement in decision making, and a governance stack, used to allocate fiduciary responsibility across organisational layers. Three limitations should be stated at the outset. The article does not conduct empirical fieldwork. It does not resolve the philosophical question of machine agency, which it treats as unnecessary for the legal analysis. And it addresses private law duties and their regulatory interface rather than criminal liability, which raises distinct issues of intent. 4. A Conceptual Framework: The Autonomy Spectrum and the Autonomous Business 4.1 Why a spectrum Debate in this field is often confused because the word autonomous is used to describe very different arrangements. A recommendation engine and a self executing trading agent are both called AI. They pose completely different governance problems. A workable legal framework needs a way to grade them. The article proposes five levels of #decision_autonomy. Each level describes the relationship between the machine's output and the firm's action. Level 0: Instrumental. The system processes information; a human forms the judgment. Spreadsheets, dashboards, and search tools sit here. Fiduciary analysis is unchanged. Level 1: Advisory. The system recommends; a human evaluates and decides. The fiduciary question becomes one of #reliance: was reliance on the model reasonable, and did the human retain the capacity to disagree? Level 2: Delegated with review. The system decides, but a human reviews before the decision takes effect. The fiduciary question concerns the quality of review. If review is nominal, a rubber stamp performed at one second per case, the arrangement is functionally Level 3 and should be treated as such. Form should not defeat substance. Level 3: Delegated with exception handling. The system decides and acts. Humans intervene only on flagged exceptions, escalations, or complaints. Most automated pricing, fraud detection, and advertising allocation sits here. This is the heart of the autonomous business. Level 4: Agentic. The system plans, acts, and initiates further actions across systems and counterparties without a predetermined action set. It may negotiate, contract, deploy code, or spend money. This is where the sharpest doctrinal problems arise. Level 5: Constitutive. The system participates in choices ordinarily reserved to the board: capital allocation, market entry, changes to the business model. No mainstream legal system permits this without a human decision layer, though structures approximating it exist in some token governed organisations. 4.2 Defining the autonomous business An autonomous business, for the purposes of this article, is a firm in which a material proportion of decisions that affect third parties or the firm's risk profile occur at Level 3 or above, and in which human intervention is exceptional rather than routine. Three features characterise such firms. The first is #decision_velocity. Machine decisions occur at a rate that makes individual human review impossible in principle, not merely inconvenient. A payments firm making two million risk decisions a day cannot have a human look at each one. That is the point of the business, not a failure of governance. The second is #opacity. Even with full access to its models, a firm may not be able to reconstruct the reason for a particular output in a form that satisfies a court or a regulator. Post hoc explanation methods provide plausible narratives, not causal accounts. The third is #distributed_causation. The harmful outcome is rarely traceable to one act by one person. It emerges from a training dataset assembled by one team, an objective specified by another, a threshold set by a third, a vendor model licensed from outside, and a deployment decision made under commercial pressure. Traditional causation analysis struggles here. 4.3 The migration thesis The central conceptual move of this article is what it calls #fiduciary_migration. In a Level 0 or Level 1 firm, the fiduciary act is the decision. A director decides to approve a transaction, and the court asks whether that approval was careful, loyal, and honest. In a Level 3 or Level 4 firm, the fiduciary act is not the decision. It is the delegation and its conditions. The relevant human choices are: The choice to automate a category of decisions at all. The choice of what the system should optimise for, and what it should treat as a constraint. The choice of how much autonomy to grant, and where the boundaries lie. The choice of what to monitor, at what frequency, and with what escalation. The choice of whether and how a decision can be reversed. The choice of who is answerable when it goes wrong. Those six choices are made by people. They are documented, or they are not. They are deliberated, or they are not. They are revisited when circumstances change, or they are not. Fiduciary law can grip all of them without any doctrinal innovation at all. What it requires is that courts and regulators look upstream. The practical consequence is significant. It means that a startup cannot defend itself by saying that the model did it. It also means that a startup can defend itself, successfully, by showing that the six choices were made with care, documented, and reviewed. The defence lives in the paper trail, and the paper trail must be built before the harm, not after. 5. The Fiduciary Core Under Conditions of Automation 5.1 The duty of care and the business judgment rule The duty of care requires a director to act with reasonable diligence. In Delaware, the practical force of the duty is heavily softened by the #business_judgment_rule, a presumption that a decision made on an informed basis, in good faith, and in the honest belief that it serves the company, will not be second guessed by a court. The English formulation, in section 174 of the Companies Act 2006, requires reasonable care, skill and diligence measured by a dual objective and subjective standard. Continental systems apply comparable standards of prudent management. Three questions arise when decisions are automated. Is the delegation itself a business judgment? It should be, and probably is. A decision to automate underwriting is a classic business decision involving trade offs between speed, cost, accuracy, and risk. The rule should protect it, provided the decision was informed. What does informed mean here? This is the crux. To be informed about a delegation to an AI system, a board needs, at minimum, to understand what the system is for, what data it learns from, what it optimises, what its known failure modes are, what happens when it fails, and how the firm will know that it has failed. A board that approves an automation programme without asking these questions has not made an informed judgment. It has made an uninformed one, and the rule does not protect uninformed judgments. Can the board rely on experts? Corporate statutes generally permit directors to rely in good faith on reports from officers, employees, and experts selected with reasonable care. This is the most likely doctrinal home for #AI_reliance, but the protection is conditional: reliance must be in good faith, the expert must be competent, and the director must not have knowledge that makes reliance unwarranted. A director who is told by an engineer that a model is fair, and who has read three customer complaints alleging discriminatory outcomes, cannot hide behind the engineer. The likely evolution is a duty of #informed_delegation: a requirement that boards understand and approve the boundaries of machine autonomy in areas of material risk, even if they cannot follow the mathematics. Nobody expects a non executive director to price a derivative. Everybody expects them to know what exposures the company has taken and why. 5.2 The duty of loyalty and the objective function Loyalty is the harder problem, because it appears at first to be irrelevant. A model has no interests. It cannot self deal. But loyalty is not only about the fiduciary's personal enrichment. It is about the improper subordination of the beneficiary's interests to some other end. And an #objective_function is precisely a statement of what interests the system will serve. Consider three ways loyalty can be compromised without any disloyal human intention. Encoded conflict. A marketplace's ranking model is trained to maximise gross merchandise value. It learns to promote the merchants who pay the highest take rate. Buyers, who believe they are seeing the best match, are being shown the most profitable match. If the firm has told buyers that ranking is based on relevance, the divergence between the stated and the actual objective is a loyalty style wrong dressed as a technical artefact. Objective drift. A system is deployed with a balanced objective. Over time, one term dominates because the data distribution shifts, or because a downstream team retunes a weight to hit a quarterly target. Nobody decided to abandon the constraint. It simply eroded. Loyalty analysis has no natural home for erosion, but it should, because the harm is identical. Proxy misalignment. The firm wants customer wellbeing but can only measure engagement, so it optimises engagement. This is well documented in the technical literature and it is the mechanism by which many of the most publicised harms of the last decade occurred. The proxy is chosen by humans. That choice is a fiduciary act. The doctrinal implication is that the specification and periodic re examination of objective functions in systems affecting customers, employees, or the market should be treated as a matter of loyalty, not merely of care. That is a stricter standard, and it removes the shelter of the business judgment rule. The claim should be limited to cases where three conditions hold: the system affects third parties directly, the firm has made representations about how it behaves, and the divergence between the stated purpose and the operative objective is knowable to the firm. For #AI_agents that act on behalf of a user rather than the firm, the problem sharpens. An agent that books travel for a customer but is monetised by supplier commissions is a fiduciary in economic substance. The debate over data loyalty developed by Richards and Hartzog (2021, 2022) becomes a design question about who the agent works for. 5.3 The duty of oversight Oversight duties, which in Delaware jurisprudence require a good faith effort to implement a reasonable information and reporting system and not to consciously ignore red flags, are the most natural fit for autonomous businesses. Recent Delaware developments have made oversight claims somewhat more viable, particularly where the risk in question is central to the company's business. For an AI heavy firm, the mission critical risk is the model itself. It follows that a board that has no reporting system covering model performance, drift, complaint patterns, and incident escalation is exposed. What should such a system contain? At a minimum, the following should reach the board or a delegated committee on a defined cadence: Inventory of systems making Level 3 or higher decisions, with owner and purpose. Material incidents, near misses, and reversals, with root cause. Performance against defined thresholds, including disparate outcome metrics where legally relevant. Complaints and disputes arising from automated decisions, with resolution times. Changes to models, objectives, or thresholds since the last report. Third party dependencies and any material change in vendor models. This is not exotic. It is a #model_risk_management report of the kind that regulated banks have produced for years. The novelty is only that it is now needed in companies that are twenty people large and eighteen months old. The oversight duty also has a #red_flag dimension. Automated systems generate their own red flags: rising complaint volumes in one customer segment, a sudden change in approval rates, an unexplained shift in a key metric after a model update. A board that receives such signals and does not inquire is in the territory of conscious disregard, which is where oversight liability actually bites. 5.4 Good faith and candour Good faith prohibits conscious disregard of known duties. In the automation context it has a specific application: deliberate ignorance. A firm that chooses not to log decisions, not to measure outcomes across protected groups, and not to build a monitoring dashboard, because knowing would create legal exposure, is not merely careless. It has adopted a strategy of not knowing. Courts have historically treated wilful blindness harshly, and there is no reason for that to change. Candour is the duty to be honest with shareholders and, in some contexts, with the market. Its automation specific form concerns representations about what AI systems do. Overstating capability, understating autonomy, or describing a human in the loop that does not meaningfully exist creates exposure under securities law, consumer protection law, and fiduciary disclosure standards at once. #AI_washing, the practice of describing conventional software as artificial intelligence to attract investment, has already drawn enforcement attention in several markets, and it is a fiduciary issue as much as a securities issue. 5.5 The non delegation principle Company law contains a principle that is easy to forget: directors may delegate but may not abdicate. Delegation transfers execution; it does not transfer responsibility. The board remains accountable for the exercise of delegated power. This principle does most of the work in the autonomous business. It means: There must always be a human or a body that is answerable for each material automated decision system. This article calls that the #accountability_anchor. Delegation must be bounded. Unlimited authority is abdication, not delegation. Delegation must be revocable. A system that cannot be stopped, paused, or overridden has not been delegated to. It has been surrendered to. Delegation must be supervised in proportion to the risk it carries. These four requirements can be stated as a rule and written into a company's constitution. They are, in the author's view, the minimum content of #fiduciary_governance for firms operating at Level 3 and above. 5.6 The evidentiary problem A final point runs across all of the above. Fiduciary litigation is an exercise in reconstruction. The court asks what the fiduciary knew and did, and it does so through documents and testimony. In an autonomous business, the relevant record is technical: model cards, evaluation results, deployment approvals, configuration changes, logs, incident reports. If these do not exist, or exist but are not retained, or exist but cannot be linked to a particular decision, then the firm cannot prove that it acted with care even if it did. The practical consequence is that #audit_trail design is a fiduciary matter. A firm that cannot reconstruct why a decision was made has, by that fact alone, weakened its own defence. This inverts the usual intuition that logging is a technical or privacy issue. It is a governance issue, and boards should treat it as one. 6. Comparative Legal Frameworks 6.1 United States and Delaware Delaware remains the default incorporation choice for venture backed startups worldwide, including many with no meaningful United States operations. Its law therefore governs the internal affairs of a very large share of autonomous businesses. The relevant features are familiar. Directors owe care and loyalty. Care claims are usually neutralised by exculpation provisions permitted under the corporate statute, which means that in practice most fiduciary litigation is framed as loyalty or bad faith. Oversight claims sit within loyalty, since a sustained failure of oversight is analysed as a failure of good faith. For an autonomous business, this produces a specific litigation shape. A plaintiff will not sue for a careless model. They will sue for a board that never asked, never monitored, and ignored warnings, framing it as an oversight failure. That is the doctrinal channel through which #algorithmic_accountability will reach Delaware. Delaware also permits reliance on experts and the establishment of committees. A board that creates a technology and risk committee, charges it with model oversight, receives regular reports, and documents its inquiries has built a strong record. What Delaware does not do is regulate the systems themselves. That comes from federal sectoral regulators, state consumer protection and civil rights law, the growing body of state level AI statutes, and litigation under employment and credit discrimination law. A startup incorporated in Delaware but selling into California, New York, Colorado, and Illinois faces a patchwork harder to navigate than the corporate law itself. 6.2 United Kingdom The Companies Act 2006 codifies directors' duties in sections 171 to 177. Three are especially relevant. Section 172 requires a director to act in the way they consider, in good faith, would be most likely to promote the success of the company for the benefit of the members as a whole, having regard to factors including the long term consequences of decisions, the interests of employees, and relationships with suppliers and customers. This is the most explicitly future oriented duty in the major common law systems, and it maps unusually well onto automation. A director who deploys a system optimised for short term revenue, knowing that it degrades customer trust or creates long term regulatory exposure, has a section 172 problem. Section 174 sets the care standard, using an objective test and a subjective test that raises the bar for directors who possess greater expertise. The subjective limb matters for startups: a founder who is a machine learning researcher is held to a higher standard on AI matters than a lay director. Technical expertise increases duty. It does not reduce it. Section 177 governs conflicts, and can reach situations where a director has an interest in a vendor whose model the company licenses, a common and under examined conflict in the startup world. The United Kingdom has so far taken a lighter, sector led approach to AI regulation, relying on existing regulators applying cross cutting principles. Regulatory expectations therefore arrive through the financial conduct regulator, the data protection regulator, and the competition authority rather than through a single statute, which means #regulatory_mapping must be done sector by sector. 6.3 European Union The European Union has produced the most detailed regulatory framework, and its extraterritorial reach means that a startup anywhere in the world that places an AI system on the European market or whose system output is used in the Union may fall within scope. The AI Act adopts a risk based structure. Certain practices are prohibited outright. A defined set of high risk uses, which includes systems used in employment and worker management, access to essential private services including creditworthiness assessment, education, and several other categories, attracts substantial obligations. These include a risk management system, data governance requirements, technical documentation, automatic logging, transparency to deployers, human oversight measures, and requirements of accuracy, robustness and cybersecurity. Providers must operate a quality management system and conduct post market monitoring. Separate obligations attach to general purpose models. Three points deserve emphasis. First, the human oversight requirement is not satisfied by a person being nominally present. Oversight measures must enable the persons concerned to understand the system's capacities and limits, remain aware of automation bias, interpret output correctly, decide not to use the system, and intervene or stop it. This is, in effect, a legal specification of the #accountability_anchor. Second, the logging requirement makes the evidentiary problem tractable: a compliant firm will possess exactly the record that fiduciary litigation requires. Third, the risk management obligation is continuous rather than a one off certification, which maps directly onto the oversight duty in company law. Alongside the AI Act sit the General Data Protection Regulation, whose provisions on automated individual decision making and profiling remain a powerful constraint on Level 3 decisions affecting individuals, the reformed product liability regime, which extends to software and can capture defective AI systems, and, for larger firms, sustainability and due diligence obligations that require the identification of adverse impacts across operations. The Digital Services Act and Digital Markets Act add further layers for platforms. The practical effect is that European regulation converts many fiduciary questions into compliance questions with documented answers. That is good for the diligent firm and dangerous for the careless one, because non compliance is now easily provable. 6.4 Singapore Singapore has taken a facilitative approach, combining a strong corporate law tradition with voluntary but detailed AI governance guidance and regulator led principles for the financial sector concerning fairness, ethics, accountability and transparency in the use of AI and data analytics. Directors' duties require reasonable diligence and honest action in the company's interests, and Singapore courts have generally been reluctant to second guess commercial judgments but firm on process failures. The distinctive feature is that soft law here is dense, specific, and operational, providing model governance frameworks, self assessment tools, and testing toolkits. Where a regulator has published a detailed voluntary framework and a firm has ignored it, a court assessing care has an obvious benchmark. #soft_law hardens through the duty of care. 6.5 United Arab Emirates and the Gulf free zones For startups operating between Europe, Asia and Africa, the common law free zones of the United Arab Emirates, principally the Dubai International Financial Centre and the Abu Dhabi Global Market, have become significant. Both operate common law systems with their own courts, their own companies legislation modelled on English law, and their own data protection regimes influenced by European standards. Both have been active in publishing guidance on the use of AI and in creating regulatory sandboxes and tailored licensing categories for technology firms. Directors in these regimes owe duties closely modelled on the English formulation, so the analysis in section 6.2 largely transfers. The distinctive governance question is structural. A startup may establish a holding company in a free zone for treaty, capital raising, or licensing reasons while its engineering and model development sit elsewhere. The board of the free zone entity may have no practical visibility over the systems that generate the group's risk. That is a governance gap, and a common one. The wider Gulf environment is also marked by rapid national AI strategy adoption and a strong state role in deployment, which means public sector procurement often brings governance obligations by contract that exceed those imposed by general law. 6.6 China China regulates by application area, with distinct instruments addressing algorithmic recommendation, deep synthesis, and generative AI services, coupled with filing and security assessment requirements and a strong emphasis on content control and data localisation. Corporate law imposes duties of loyalty and diligence on directors and senior managers. For an international startup, the significance of China is usually not incorporation but market access and supply chain. A firm that uses a model developed in China, or that serves Chinese users, imports obligations that its Delaware board is unlikely to have considered. 6.7 Legal recognition of autonomous organisations A small number of jurisdictions have created legal forms intended to accommodate organisations governed substantially by code. Wyoming's limited liability company legislation permits algorithmically managed entities and has been used to give DAOs a recognised form; the Marshall Islands has enacted legislation recognising DAOs as limited liability entities; and foundation or association forms have been used elsewhere as legal wrappers. These regimes are instructive for two reasons. First, they show that legislatures can accommodate machine governance without granting #legal_personality to the machine. The entity is the person, the code is the management mechanism, and duties attach to whoever holds the power to change the code. Second, they show the liability leakage risk. Where a wrapper is absent or defective, participants may be treated as members of a general partnership, with unlimited personal liability. Autonomy without legal form does not remove liability. It distributes it, unpredictably, to whoever is standing closest. 7. Cross Border Problems Specific to International Startups 7.1 The mismatch between corporate law and operational reality The internal affairs of a company are governed, in most systems, by the law of the place of incorporation. The conduct of a company is governed by the law of the places where it acts and where its effects are felt. For a traditional business, those largely coincide. For an autonomous business, they routinely do not. Consider a common structure. A holding company is incorporated in Delaware because investors demand it. A subsidiary in Ireland or the Netherlands holds intellectual property and contracts with European customers. Engineers are employed through arrangements in one or more countries. Models are trained on cloud infrastructure elsewhere. Customers are in twenty countries. The board sits in San Francisco, London, and Dubai, and meets by video. Fiduciary duties here are owed at the level of each entity. But the decisions that create risk, what the model optimises for, how much autonomy it has, what is logged, are made once, centrally, and applied everywhere. The result is a #governance_asymmetry: a single design decision generates duties in several legal systems at once, and each entity's board is nominally responsible for consequences it did not choose and may not see. Directors of a European operating subsidiary may face exposure for a system whose objective function was set by a team reporting to the parent's chief technology officer. If that board never asked and never received reports, it has an oversight problem regardless of where the design happened. 7.2 Regulatory arbitrage and its limits Startups routinely engage in #regulatory_arbitrage, choosing jurisdictions on the basis of favourable rules. This is legitimate and, within limits, prudent. But arbitrage has three limits that founders systematically underestimate. The first is #extraterritoriality. European AI and data legislation applies by reference to market placement and to the location of affected persons, not to the location of the provider. Incorporating outside Europe does not remove the obligation if the system is used in Europe. The same logic applies to a growing number of regimes. The second is the fiduciary limit. A board that selects a jurisdiction primarily to avoid oversight obligations, and that thereby exposes the company to enforcement risk in its actual markets, may itself be in breach of the duty of care. Choosing a weak regime is a business decision, and business decisions must be informed. If the informed conclusion is that the company will be non compliant in the markets that generate its revenue, then the choice is not protected. The third is the investor limit. Investors increasingly conduct AI governance due diligence. A structure that cannot demonstrate compliance in its largest markets damages valuation and complicates later financing. Arbitrage that destroys enterprise value is not arbitrage. It is a mistake. 7.3 Conflict of laws and the location of an algorithmic act Where does an automated decision happen? The question sounds metaphysical and is intensely practical. If a model hosted in one country, trained on data from a second, operated by a company incorporated in a third, refuses credit to a person resident in a fourth, which law governs the harm? The place of the harmful effect is the usual starting point in modern conflict rules for non contractual obligations. Choice of law and forum clauses may redirect disputes, subject to consumer protection carve outs that are broad in Europe. Regulatory obligations attach independently of both. The result is that #conflict_of_laws analysis cannot be done once at incorporation. It has to be done per system, per market, per decision category. A useful discipline is a matrix mapping each Level 3 or higher system against the markets it touches, the persons it affects, and the applicable regime. That matrix belongs in the board pack. 7.4 The vendor chain Most startups do not build their own foundation models. They license them. They also license fraud detection, identity verification, credit data, and content moderation from specialist vendors. The firm's automated decisions are therefore the output of a chain of systems, most of which it did not build and cannot inspect. This creates an accountability problem that Cobbe, Veale and Singh (2023) describe well. Responsibility diffuses along the chain. The vendor says the customer configured it. The customer says the vendor built it. The foundation model provider disclaims warranties in a standard terms document that nobody negotiated. Fiduciary duty cuts through this. The board of the firm that faces the customer cannot outsource its responsibility to a supplier. It can, and must, manage it. That means #vendor_diligence with teeth: contractual rights to information, documented evaluation of the vendor's system against the firm's use case, notification obligations for material model changes, audit and testing rights where the risk warrants, and a fallback plan if the vendor's system fails or is withdrawn. A firm whose entire credit decision rests on a third party model, with no contractual right to be told when that model changes, has a governance defect, and the board should be told so. 7.5 Data, borders and the training pipeline Training data is where legal exposure is created quietly. Data localisation rules, cross border transfer restrictions, sectoral confidentiality obligations, intellectual property claims, and consent limitations all bear on whether a dataset may lawfully be used for a given purpose. The fiduciary point is narrow but important. Directors of a startup whose core asset is a model trained on data of uncertain provenance are sitting on a contingent liability that may exceed the company's value. The duty of care requires them to know whether that is the case. A #data_provenance record is not a nice to have. It is a condition of being able to say, truthfully, that the board understands the company's risk. 8. Liability, Personality and the Accountability Gap 8.1 The personality question, and why it is a distraction A recurring proposal is to grant AI systems some form of legal personality, so that they can bear duties and liabilities directly. The proposal has attracted sustained criticism, and the criticism is persuasive. Legal personality is a device for channelling liability to an asset pool and for enabling an entity to hold rights. Granting it to a model would achieve neither purpose usefully. The model has no assets. Any assets assigned to it would come from the firm, which is where liability would have landed anyway. And personality would create a shield, allowing firms to point at the machine and walk away. That is precisely the outcome fiduciary law exists to prevent. The better analysis, and the one this article adopts, is that AI systems are instrumentalities. The relevant questions are: who deployed it, who configured it, who profited from it, and who could have stopped it. Those questions have human answers. 8.2 Agency law as the working analogy Agency law offers the most productive analogy, as Kolt (2025) argues in detail. A principal who authorises an agent to act is bound by acts within the scope of that authority, and is liable for harms caused within the scope of the engagement. Authority can be actual or apparent. Apparent authority is created by the principal's own representations to third parties. Applied to autonomous systems, this yields workable results. A firm that deploys an agent to negotiate and conclude contracts has given it actual authority, and is bound. A firm that presents a chatbot to customers as able to make commitments has created apparent authority, and may be bound by what it promises, whether or not the firm intended the promise. Disputes over misleading statements made by customer facing bots have already been resolved on essentially this basis in some jurisdictions. The agency analogy also explains the limits. An agent that acts outside its authority does not bind the principal, but the principal may still be liable for negligently granting or supervising that authority. This maps precisely onto the fiduciary migration thesis. The wrong is in the authorisation and the supervision, not in the act. Where the analogy strains is in the remedies. Human agents are disciplined by reputation, incentives, and the fear of sanction. None of these operate on a model. Monitoring, the residual solution, is precisely what is hardest at machine speed and scale. This is why technical infrastructure, logging, evaluation, and circuit breakers, becomes a legal necessity rather than an engineering preference. 8.3 Product liability and the software turn The extension of product liability regimes to software and AI systems changes the picture for firms that supply systems to others. A defect can arise from behaviour after deployment, including from learning, and the absence of updates may itself amount to a defect. Burden of proof rules in some regimes have been softened where complexity makes proof excessively difficult for a claimant. For a startup that sells an AI product, this creates direct exposure to end users independent of contract; for one that uses an AI product, it creates a potential recovery route against the vendor, which is why the contractual rights described in section 7.4 matter. 8.4 The insurance layer Directors and officers policies may or may not respond to claims arising from automated decisions, depending on wording. Technology errors and omissions cover may exclude AI outputs. Cyber policies may not respond to a model failure that is not a security incident. A board that has not confirmed whether its cover responds to the firm's principal risk has not completed its risk analysis. #insurance_alignment is a modest but concrete obligation, and it is cheap to discharge. 8.5 The accountability gap restated Putting these strands together, the accountability gap in autonomous businesses is not a gap in the availability of legal duties. Duties are abundant. It is a gap in three other things. Traceability. The inability to link an outcome to a decision and a decision maker. Attribution. The diffusion of causal responsibility across teams, vendors, and time. Capacity. The absence, in small firms, of the internal functions that would allow duties to be discharged and proved. Any useful framework must address all three. Doctrine addresses none of them by itself. 9. Illustrative Scenarios The following scenarios are hypothetical constructs used for analysis. They are not accounts of real events and are not attributed to any real firm. They are designed to isolate the doctrinal issues discussed above. 9.1 The drifting price A logistics startup incorporated in Delaware, operating across six European markets, uses a dynamic pricing model. The model is instructed to maximise contribution margin subject to a set of constraints. Over eighteen months, three separate teams adjust weights to meet quarterly targets. One constraint, a cap on surge multipliers during emergencies, is disabled during a test and never restored. A storm causes prices in one city to rise sharply. Regulators open an inquiry. Media coverage follows. The company loses a licence in one market. The fiduciary analysis does not turn on the storm. It turns on change control. Who approved the removal of the constraint? Was the change recorded? Did anyone review the cumulative effect of eighteen months of adjustments against the objective the board originally approved? If the answer to all three is no, the board has an oversight problem, because the risk was central to the business, the failure was foreseeable, and no reporting system existed to catch it. #change_control is not a technical detail. It is the audit trail of a fiduciary decision. 9.2 The polite refusal A lending startup automates credit decisions using a vendor model combined with its own features. Approval rates for one applicant group fall sharply after a model update. Customer complaints rise. The complaints are handled by a support team using templates, and are never aggregated into a report that reaches management. Nine months later, a regulator finds a pattern of #disparate_outcomes. The company argues that it did not intend to discriminate and that the model is a vendor product. Neither argument helps. Intent is not required under most discrimination regimes. Vendor sourcing does not transfer responsibility away from the customer facing firm. The failure is that complaint data, which is the cheapest early warning signal any firm possesses, was collected and then discarded. The oversight duty is engaged not because the model was biased but because the firm built an organisation that was structurally incapable of noticing. 9.3 The autonomous negotiator A software startup deploys an agent that negotiates renewal terms within defined parameters, including latitude to offer discounts. Under pressure to hit a target, an engineer widens the discount ceiling without formal approval. The agent, optimising for renewal rate, discounts heavily to customers who would have renewed anyway, and in one case agrees terms that conflict with a most favoured nation clause in a large contract, triggering a repricing obligation. The issue is authority. The agent acted within the authority the firm actually gave it, so the firm is bound. The internal wrong is the unauthorised widening of the parameters. The board level wrong is having permitted an agent to bind the company without requiring a second approval for changes to its authority envelope. #authority_limits must be governed as seriously as bank payment limits, because they are the same kind of thing. 9.4 The unstoppable system A health technology startup deploys a triage model across clinics in three countries, embedded in workflow software supplied by a partner. When a safety concern emerges, the firm finds it cannot disable the model in one country without disabling the partner's entire platform, because the integration was built without a separate control. The absence of a #kill_switch turns a manageable incident into a crisis. The delegation was, in the sense described in section 5.5, an abdication: the firm surrendered a power it could not take back. Reversibility should be a condition of deployment, tested before launch, not discovered afterwards. 10. A Proposed Framework: The Fiduciary Governance Stack The framework proposed here has five layers. It is designed to be implementable by a company of twenty people, and to scale from there. 10.1 Layer one: the constitutional layer The first layer is documentary and sits in the company's foundational governance instruments: the charter or articles where permitted, and otherwise in a board approved governance policy that is expressly incorporated into the board's terms of reference. It should state four things. Autonomy is granted, not assumed. No system may make decisions at Level 3 or above in a defined list of sensitive categories without an express board or committee authorisation. Every authorisation has an anchor. Each authorised system has a named accountable executive who answers to the board for its behaviour. Accountability may not be assigned to a team, a function, or the system itself. Every authorisation is bounded and revocable. The authorisation specifies the decision categories, the value or volume limits, the constraints that may not be traded away, and the conditions under which the system must be paused. Objectives are approved, not inherited. The objective function of an authorised system, expressed in plain language, and the constraints that bound it, are approved at the same level as the authorisation, and material changes require re approval. These four statements are the constitutional core of #fiduciary_governance. They convert an abstract duty into a checkable rule. 10.2 Layer two: the board layer The second layer concerns how the board discharges oversight. For a startup, a full committee is often unrealistic. A workable minimum is a designated director with responsibility for AI oversight, supported by a standing item on the board agenda and a defined reporting pack. As the company grows, this becomes a technology and risk committee with a written charter. The reporting pack should contain the six items listed in section 5.3. It should be short. A two page report that is read is worth more than a forty page report that is filed. The board should also record its inquiries. The single most valuable defensive artefact in fiduciary litigation is a minute showing that a director asked a hard question and received an answer. #board_oversight is proved by curiosity, documented. 10.3 Layer three: the operational layer The third layer is the internal control environment. Adapting the three lines model to a small firm, as Schuett (2024) suggests for AI organisations, gives a workable structure. The first line is the product and engineering teams that build and run systems. They own the risk. They must maintain a system inventory; document each system's purpose, data, objective, constraints, known limitations, and evaluation results; operate change control so that no change to a model, objective, threshold, or authority envelope reaches production without recorded approval; and log decisions at a granularity that permits reconstruction. The second line is a risk and compliance function, which in a small company may be one person with a partial allocation or a fractional external appointment. It maintains the risk framework, challenges first line assessments, owns the incident process, and reports independently to the board. The third line is assurance, which for a startup will usually be external and periodic. The point is not bureaucracy. It is the separation of the person who builds from the person who checks. A startup where the engineer who deployed the model also assesses whether it is safe has no second line, and therefore no meaningful control. 10.4 Layer four: the assurance layer The fourth layer is verification. Its content is drawn from the auditing literature and from the conformity assessment logic of the European framework. The practical elements are: pre deployment evaluation against defined criteria, including performance across relevant subgroups where legally material; an #impact_assessment for systems affecting individuals, documenting foreseeable harms and mitigations; post deployment monitoring against thresholds, with automatic alerting; periodic re evaluation on a defined cadence and on trigger events such as a data shift or a vendor model update; and independent review, external where the risk is high. Assurance must be capable of producing a negative result. An evaluation that cannot fail is not an evaluation. Boards should ask what would have to be true for a system to be stopped, and whether anyone has the authority to stop it. 10.5 Layer five: the redress layer The fifth layer faces outward. It concerns the people affected by automated decisions. Its elements are: a clear route to contest an automated decision; human review that is genuine, meaning the reviewer has the information, competence, and authority to reach a different conclusion; aggregation of contests and complaints into a stream that reaches the second line and the board; and correction, meaning that a pattern of successful contests triggers a review of the system, not merely of the individual cases. This is where the loyalty duty becomes visible to the people who bear the consequences. It is also, as scenario 9.2 illustrates, the cheapest early warning system a firm can build. Firms that treat complaints purely as a cost centre are discarding their best risk data. 10.6 The five principles that hold the stack together Underneath the layers sit five principles, each statable in a sentence. Proportional autonomy. The degree of machine autonomy permitted should be proportionate to the reversibility of the decision and the severity of the harm if it is wrong. Traceability. Every material automated decision must be reconstructible: what system, what version, what inputs, what output, what happened next. Reversibility. No system should be deployed unless it can be stopped, and the stop mechanism should be tested. Human answerability. Every system has a human who answers for it. This is not a technical role. It is a governance role, and it should carry consequences. Objective transparency. What the system optimises for, and what it is forbidden to trade away, must be written down in language a director can read. These five principles are, in the author's submission, the operational translation of #fiduciary_duty for autonomous businesses. They are not a substitute for law. They are what discharging the law looks like. 11. Recommendations 11.1 For founders and boards Build the record before you need it. The single most consequential act a startup board can take is to insist that automated decisions are logged, that system documentation exists, and that changes are approved. None of this is expensive early. All of it is extremely expensive to retrofit under regulatory pressure. Approve autonomy explicitly, in writing: which systems may act without human review, in which categories, up to what limits, and what they may never do. Revisit annually. Appoint an anchor, one named person per material system, and record it in the minutes. Ask the stop question for every system: how would we know it is going wrong, who would notice, how quickly, and who can stop it. If there is no good answer, that is the finding. Diligence the vendor chain, insisting on notification of material model changes and on a fallback. And confirm that the insurance responds. That last step takes a morning. 11.2 For investors Investors are, in practice, the most effective regulators of early stage governance, because they control the money. Term sheets and shareholder agreements should treat AI governance as a reserved matter where the risk is material, requiring board approval for deployment of systems in sensitive categories. Diligence should include a system inventory, a documentation sample, and evidence of change control. Board seats should carry a specific oversight brief where the company operates at Level 3 or above. #investor_diligence at this level costs little and prevents a great deal. 11.3 For regulators and legislators First, calibrate by autonomy and by harm, not by sector or by named technology. A spectrum based approach of the kind sketched in section 4 provides a more durable basis for obligation than a list of technologies that will be obsolete in three years. Second, make proportionality real for small firms. Obligations rational for a large enterprise can be prohibitive for a company of fifteen people, and the result is not compliance but avoidance. Tiered documentation, standardised templates, and safe harbours for firms that adopt recognised frameworks would improve actual compliance more than higher penalties. Third, prioritise traceability. Logging and documentation obligations do the most work, because they make every other obligation enforceable and every fiduciary duty provable. If a legislature can do only one thing, it should do this. 11.4 For courts Courts will decide, in the next few years, whether existing duties reach system design. The argument of this article is that they can and should, without doctrinal innovation, by locating the fiduciary act upstream at the point of delegation, specification, and monitoring. Two specific propositions are offered for adoption. The first is that nominal human review does not convert a Level 3 system into a Level 2 system. Where review is systematically incapable of altering the outcome, the decision is automated in substance and should be analysed as such. The second is that a firm's inability to explain why an automated decision was made should not, by itself, defeat a claim. Where the firm chose not to build the capacity to explain, the evidential consequences of that choice should fall on the firm. This is not a reversal of the burden of proof. It is an application of the ordinary principle that a party cannot benefit from the non creation of evidence within its control. 12. Limitations and Future Research The article is doctrinal and conceptual. It does not test its claims against data on what startups actually do, and empirical work on the governance practices of AI heavy early stage firms is scarce and badly needed. It also assumes that the fiduciary framework is worth preserving. That assumption is contestable. Automation at scale may eventually require a shift from fiduciary regulation of firms to direct regulation of systems, with duties owed to affected persons rather than to the company. The article's position is that this shift is neither necessary nor desirable yet, but the question is open. The article does not resolve the position of #AI_agents that act for users rather than for firms. If such agents become the primary interface through which individuals transact, the loyalty questions they raise will require a body of law that does not yet exist. Finally, the article is written during rapid regulatory change. Several instruments discussed are in phased implementation, and the interpretation of key obligations, particularly human oversight and risk management, will be shaped by supervisory practice that has not yet accumulated. Four research directions follow: empirical mapping of governance practice in AI heavy startups; doctrinal analysis of the first wave of oversight litigation; comparative study of how the human oversight requirement is operationalised across supervisors; and work on the economics of proportionality, to identify the point at which governance obligations tip from protective to prohibitive for small firms. 13. Conclusion The temptation, when a machine causes harm, is to ask what the machine owed. It is the wrong question. Machines owe nothing. They have no interests to subordinate and no loyalty to breach. What they have is power, and that power was given to them by people who could have given less of it, bounded it more carefully, watched it more closely, or taken it back. Fiduciary law has always been about the responsible exercise of power over the interests of others. That is exactly what an autonomous business does at scale, thousands of times a day, in markets its directors have never visited, to people they will never meet. The duties are not obsolete. They are more relevant than they have ever been. What has changed is where they bite. They bite at the point of delegation, where a board decides how much a system may do without being asked. They bite at the point of specification, where a team decides what the system will maximise and what it may never sacrifice. They bite at the point of monitoring, where a firm decides whether it wants to know. And they bite at the point of reversal, where a firm discovers whether it kept the power it thought it had merely delegated. For the international startup, the difficulty is compounded by structure. The entity that bears the liability is often not the entity that made the choice. The market that imposes the rule is often not the market that granted the charter. The vendor that built the model is often not the party that faces the customer. In that environment, the only durable strategy is the one fiduciary law has always recommended: know what you are doing, write it down, watch what happens, and be able to stop. None of this requires a new legal concept. It requires that an old one be pointed in the right direction, and that the firms it governs build the machinery that lets it grip. #accountability does not scale automatically. It has to be designed. In an autonomous business, design is a fiduciary act. References Aggarwal, N. (2021) The norms of algorithmic credit scoring. Cambridge Law Journal, 80(1), pp. 42 to 73. Almada, M. and Petit, N. (2025) The EU AI Act: a medley of product safety and fundamental rights? Common Market Law Review, 62(1). Bellavitis, C., Fisch, C. and Momtaz, P.P. (2023) The rise of decentralized autonomous organizations (DAOs): a first empirical glimpse. Venture Capital, 25(2), pp. 187 to 203. Bruner, C.M. (2020) Distributed ledgers, artificial intelligence and the purpose of the corporation. Cambridge Law Journal, 79(3), pp. 431 to 458. Buiten, M., de Streel, A. and Peitz, M. (2023) The law and economics of AI liability. Computer Law and Security Review, 48, 105794. Chan, A., Salganik, R., Markelius, A. and others (2023) Harms from increasingly agentic algorithmic systems. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pp. 651 to 666. Chesterman, S. (2021) We, the Robots? Regulating Artificial Intelligence and the Limits of the Law. Cambridge: Cambridge University Press. Cihon, P., Schuett, J. and Baum, S.D. (2021) Corporate governance of artificial intelligence in the public interest. Information, 12(7), 275. Cobbe, J., Veale, M. and Singh, J. (2023) Understanding accountability in algorithmic supply chains. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pp. 1186 to 1197. Criddle, E.J., Miller, P.B. and Sitkoff, R.H. (eds) (2019) The Oxford Handbook of Fiduciary Law. New York: Oxford University Press. De Filippi, P., Mannan, M. and Reijers, W. (2022) The alegality of blockchain technology. Policy and Society, 41(3), pp. 358 to 372. Enriques, L. and Zetzsche, D.A. (2020) Corporate technologies and the tech nirvana fallacy. Hastings Law Journal, 72(1), pp. 55 to 98. Floridi, L. (2021) The European legislation on AI: a brief analysis of its philosophical approach. Philosophy and Technology, 34, pp. 215 to 222. Gabriel, I., Manzini, A., Keeling, G. and others (2024) The ethics of advanced AI assistants. Google DeepMind research report. Gramitto Ricci, S.A. (2020) Artificial agents in corporate boardrooms. Cornell Law Review, 105(3), pp. 869 to 918. Hacker, P. (2023) The European AI liability directives: critique of a half hearted approach and lessons for the future. Computer Law and Security Review, 51, 105871. Hacker, P., Engel, A. and Mauer, M. (2023) Regulating ChatGPT and other large generative AI models. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pp. 1112 to 1123. Hartzog, W. and Richards, N. (2022) Legislating data loyalty. Notre Dame Law Review Reflection, 97(1), pp. 356 to 368. Hassan, S. and De Filippi, P. (2021) Decentralized autonomous organization. Internet Policy Review, 10(2). Hilb, M. (2020) Toward artificial governance? The role of artificial intelligence in shaping the future of corporate governance. Journal of Management and Governance, 24, pp. 851 to 870. Kaal, W.A. (2021) Decentralized autonomous organizations: internal governance and external legal design. Annals of Corporate Governance, 5(4), pp. 237 to 306. Kolt, N. (2025) Governing AI agents. Notre Dame Law Review, 101 (forthcoming). Laux, J., Wachter, S. and Mittelstadt, B. (2024) Trustworthy artificial intelligence and the European Union AI Act: on the conflation of trustworthiness and acceptability of risk. Regulation and Governance, 18(1), pp. 3 to 32. Metcalf, J., Moss, E., Watkins, E.A., Singh, R. and Elish, M.C. (2021) Algorithmic impact assessments and accountability: the co construction of impacts. Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pp. 735 to 746. Mokander, J., Axente, M., Casolari, F. and Floridi, L. (2022) Conformity assessments and post market monitoring: a guide to the role of auditing in the proposed European AI regulation. Minds and Machines, 32, pp. 241 to 268. Mokander, J., Schuett, J., Kirk, H.R. and Floridi, L. (2024) Auditing large language models: a three layered approach. AI and Ethics, 4, pp. 1085 to 1115. Novelli, C., Casolari, F., Rotolo, A., Taddeo, M. and Floridi, L. (2024) Taking AI risks seriously: a new assessment model for the AI Act. AI and Society, 39, pp. 2493 to 2497. Petit, N. (2020) Big Tech and the Digital Economy: The Moligopoly Scenario. Oxford: Oxford University Press. Raji, I.D., Smart, A., White, R.N. and others (2020) Closing the AI accountability gap: defining an end to end framework for internal algorithmic auditing. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, pp. 33 to 44. Richards, N. and Hartzog, W. (2021) A duty of loyalty for privacy law. Washington University Law Review, 99(3), pp. 961 to 1021. Saetra, H.S. (2023) The AI ESG protocol: evaluating and disclosing the environment, social, and governance implications of artificial intelligence capabilities, assets, and activities. Sustainable Development, 31(2), pp. 1027 to 1037. Santana, C. and Albareda, L. (2022) Blockchain and the emergence of decentralized autonomous organizations (DAOs): an integrative model and research agenda. Technological Forecasting and Social Change, 182, 121806. Schuett, J. (2023) Risk management in the Artificial Intelligence Act. European Journal of Risk Regulation, 15(2), pp. 367 to 385. Schuett, J. (2024) Three lines of defense against risks from AI. AI and Society, 39, pp. 2967 to 2984. Selbst, A.D. (2021) An institutional view of algorithmic impact assessments. Harvard Journal of Law and Technology, 35(1), pp. 117 to 191. Sims, A. (2021) Decentralised autonomous organisations: governance, dispute resolution and regulation. Journal of Corporate Law Studies, 21(2). Smuha, N.A. (2021) Beyond the individual: governing AI's societal harm. Internet Policy Review, 10(3). Veale, M. and Zuiderveen Borgesius, F. (2021) Demystifying the draft EU Artificial Intelligence Act. Computer Law Review International, 22(4), pp. 97 to 112. Wachter, S., Mittelstadt, B. and Russell, C. (2021) Why fairness cannot be automated: bridging the gap between EU non discrimination law and AI. Computer Law and Security Review, 41, 105567. Zetzsche, D.A., Arner, D.W. and Buckley, R.P. (2020) Decentralized finance. Journal of Financial Regulation, 6(2), pp. 172 to 203. Hashtags #Fiduciary_Governance #Autonomous_Businesses #AI_Decision_Making #Corporate_Law #Startup_Law #AI_Regulation #Algorithmic_Accountability #Director_Duties #Comparative_Law #Tech_Governance #AI_Ethics #Legal_Research #Corporate_Compliance #Business_Law #Emerging_Technology

Since 2019, the Times Higher Education Impact Rankings have grown from a pilot exercise involving 450 universities into the most widely used global instrument for measuring how higher education institutions contribute to the United Nations Sustainable Development Goals. In 2026 the exercise was renamed the Sustainability Impact Ratings and moved to a membership model, a change that reveals how deeply the market for #global_recognition has become institutionalised. This article examines a group of actors who sit at the centre of that market but are rarely studied directly: the #educational_entrepreneurs who design, staff, fund and manage institutional participation in the ranking. Drawing on a structured review of peer reviewed research published between 2020 and 2025, together with documentary analysis of published ranking methodologies and results, the article maps the concrete practices these actors use. These include portfolio selection across the seventeen goals, the construction of evidence infrastructure, the production of documentary narratives, the strategic cultivation of partnerships, the embedding of sustainability in curricula, and the reorganisation of campus operations. The article then develops a five stage maturity framework that distinguishes cosmetic compliance from genuine institutional transformation. The central finding is that the ranking rewards two distinct capabilities at once: the capability to act on society, and the capability to document that action. Because the second capability is cheaper to build than the first, the instrument creates a persistent risk of decoupling between reported and real #societal_impact. The article closes with practical guidance for students who read these tables, for administrators who manage submissions, and for policymakers who use the results as evidence. Keywords: #Impact_Rankings, #Sustainable_Development_Goals, #higher_education, #educational_entrepreneurship, #university_rankings, #institutional_strategy, #societal_impact, #SDG_reporting 1. Introduction A university in Manila, a technical institute in Amman, a regional campus in Queensland and a research intensive institution in Manchester now share something they did not share ten years ago. Each of them can, if it chooses, submit a large body of evidence about its work on poverty, gender equality, clean water, decent work, climate action and eleven other goals to a single London based data operation, and receive back a numerical score and a place in a published table. That table is read by prospective students, by ministries of education, by donors, by employers and by the institutions themselves. It has become, in a very short time, one of the few ways a young or non elite university can appear on a #global_ranking at all. This article is about the people who make that submission happen and the choices they make along the way. I call them #educational_entrepreneurs, using the term in a broad sense. They are vice chancellors and rectors who decide that sustainability will be the institution's competitive position. They are pro vice chancellors for engagement, directors of sustainability, heads of institutional research, communications leads, and increasingly the specialist ranking officers whose entire job is to manage submissions. They are also faculty members who build community programmes and then have to translate that work into an auditable evidence file. What unites them is that they operate at the meeting point of two logics. One logic is #mission_driven and asks what the university is good for. The other is #market_driven and asks how the university will be seen. The #Impact_Rankings sit exactly on the seam between the two. The argument of this article is simple to state and harder to work through. The ranking measures two things that are related but not the same. The first is real activity in the world: research that gets used, graduates who change practice, campuses that cut emissions, partnerships that shift policy. The second is the institution's capacity to notice, collect, formalise and narrate that activity in the specific format an external assessor will accept. These are different organisational capabilities. The second one is significantly cheaper and faster to acquire than the first. Any institution that wants a better score therefore faces a fork in the road at every planning cycle, and the incentives do not point cleanly in one direction. This is not a claim that universities are dishonest. It is a claim about what happens when a complex social good is turned into a comparable number, a problem that scholars of quantification have described in many other fields. Existing research on the ranking has concentrated on its methodology and on whether its scores are valid measures of sustainability. That work is essential and this article draws on it heavily. What is missing is a systematic account written from the inside of the institution, one that treats optimisation not as a scandal to be exposed but as an ordinary #management_practice that can be done well or badly, honestly or dishonestly, with substance or without it. Students in particular deserve this account. They are told to use these tables to choose where to study. They are rarely told how the numbers are produced or which behaviours produce them. The article proceeds as follows. Section 2 reviews the relevant literature on rankings as calculative devices, on the entrepreneurial university, and on the specific critiques of the Impact Rankings. Section 3 explains the method. Section 4 sets out the anatomy of the instrument itself, because no strategy can be understood without understanding the scoring rules it responds to. Section 5 presents seven families of optimisation practice observed in the literature and in published institutional accounts. Section 6 offers a maturity framework. Section 7 discusses the risks, including #SDG_washing, equity effects and the consequences of the 2026 shift to a paid membership model. Section 8 draws out implications for students, institutions and policymakers. Section 9 states limitations and Section 10 concludes. 2. Literature Review and Conceptual Background 2.1 Rankings as devices that change what they measure The starting point for any serious discussion of #university_rankings is that they are not neutral mirrors. They are instruments that reorganise the thing they claim to describe. When a number is published and audiences begin to treat it as real, organisations reorganise themselves around the number, and the number then becomes a more accurate description of the organisation than it was when first invented. This is a well documented pattern in the sociology of quantification, and researchers have traced how numerical thinking becomes institutionalised inside organisations until it feels like plain common sense rather than a choice (Chun and Sauder, 2022). Applied to higher education, this means a ranking does not simply find the good universities. It teaches universities what good means and then rewards the ones that learn fastest. Hazelkorn has shown how the growth of rankings created an entirely new class of expertise inside institutions, with people whose professional standing depends on their ability to read, interpret and manipulate ranking data, and who therefore have an interest in the ranking continuing to matter (Hazelkorn, 2023). This is not a criticism of those individuals. It is a description of how #ranking_expertise becomes a durable feature of the sector. The methodological critique of global rankings is now mature. Gadd's assessment of the major global tables against agreed principles of responsible evaluation found that all of them fell short, and that the most visible ones fell shortest, largely because they blend unrelated indicators into a single composite score without a defensible theory of why those weights and not others (Gadd, 2021). Composite indicators of this kind are known to be sensitive to small changes in weighting, which means that a university can move several places without changing anything it actually does. The same work notes that rankings tend to favour institutions that are older, wealthier, English speaking and located in North America and Western Europe, which raises a #structural_equity problem that no amount of technical adjustment fully solves. More recent quantitative work using network methods has continued to find socio economic and territorial bias embedded in the way ranking scores distribute (Bellantuono et al., 2025). The relevant conclusion for this article is not that rankings are worthless. It is that #reactivity is the normal condition of ranked organisations, and that optimisation is therefore not a deviation from the system but the system working as designed. 2.2 The entrepreneurial university and the third mission The second literature this article draws on concerns the changing purpose of the university. Alongside teaching and research, institutions have been asked for several decades to take on a #third_mission that covers technology transfer, community engagement, policy advice and social innovation. A systematic review of this literature identified both the potential of the third mission and the constraints that limit it, including weak internal incentives, unclear metrics and the difficulty of persuading academic staff that engagement counts as real work (Compagnucci and Spigarelli, 2020). Research on #academic_engagement has similarly shown that the collaborative, advisory and community facing activities of academics are widespread but poorly captured by conventional performance systems (Perkmann et al., 2021). This creates a gap that the Impact Rankings entered. Universities were being told to matter to society but had no accepted way to show that they did. Traditional world rankings measure research volume, citation impact, reputation surveys, staff to student ratios and international outlook. None of these register whether a university reduced local water poverty or trained enough nurses for its region. The arrival of an instrument that promised to make the third mission visible was, for many institutions, genuinely welcome. That welcome is important. It explains why participation grew so quickly, and it explains why criticism of the ranking is often met with defensiveness from people who see it as the first tool that ever noticed their work. The #entrepreneurial_university concept is usually associated with commercialisation, spin outs and industry income. In this article I use a wider reading. An educational entrepreneur is someone who identifies an unmet need or an unclaimed position, assembles resources, accepts risk and builds something that did not exist before. Under that definition, the sustainability director who builds an institution wide SDG evidence system, secures a budget for it, persuades deans to cooperate and turns a middling institution into a globally visible one is doing entrepreneurial work, whether or not any money is made. 2.3 Higher education and the Sustainable Development Goals The 2030 Agenda gave universities a common vocabulary. Scholars have argued that higher education can contribute to sustainable development through research, teaching and community development, but that the sector still lacks agreed ways to document and evidence the enormous variety of activity involved (Chankseliani and McCowan, 2021). A large survey of academics across six regions found that engagement with the goals is real but uneven, frequently driven by individual enthusiasm rather than institutional strategy, and rarely supported by systematic measurement (Leal Filho, Salvia and Eustachio, 2023). Work on strategic integration has stressed that the goals only change behaviour when they are written into institutional strategy, resourced and monitored, rather than being announced and then left to volunteers (Leal Filho et al., 2023). This matters for the present argument. If SDG activity in most universities is fragmented and undocumented, then the first thing any ranking submission does is force the institution to look at itself. That act of looking has value independent of the score. Several institutions have reported that the exercise revealed programmes their own leadership did not know existed. #Self_knowledge is an underrated by product of #SDG_reporting. 2.4 The specific critique of the Impact Rankings The dedicated critical literature on this ranking is small but sharp. The most cited study conducted a content analysis of the methodology and concluded that the scoring structure allows a university to reach a high overall position while contributing little or nothing across most of the goals, because only four of the seventeen goals feed the overall score, and three of those four are the institution's own best performing goals (Bautista-Puig, Orduna-Malea and Perez-Esparrells, 2022). The same study raised concerns about the appropriateness and weighting of the research metrics used, and about the wide score bands that make positions difficult to interpret meaningfully. A later analysis using two years of data confirmed that the ranking does capture something real about institutional sustainability performance, while also finding large disparities in participation behaviour that are hard to explain by sustainability performance alone (Urbano et al., 2025). Work examining submission patterns directly has asked whether the observed behaviour represents meaningful measurement or #gaming_the_system, noting that universities steadily increase the number of goals they submit to over time in a way that is unlikely to reflect the sudden appearance of new sustainability activity, and more likely reflects growing skill at reporting (Dickson, 2025). Beyond higher education, the corporate literature has documented the phenomenon of #SDG_washing, where organisations select the goals that flatter them, report enthusiastically on those, and stay silent on the goals where their performance would look poor (Heras-Saizarbitoria, Urbieta and Boiral, 2022). The parallel with a ranking that explicitly counts only a university's best three goals is direct and uncomfortable. Research on sustainability reporting tools has likewise examined the merits and limits of sustainability rankings as reporting instruments, finding that their value depends heavily on the rigour of verification behind them (Patara and Dhalla, 2022). Finally, econometric work has modelled the relationship between what universities disclose and where they finish, finding that disclosure itself is strongly associated with score, and that the highest placed institutions differ from others in geography, in the information they choose to disclose, and in their orientation toward industry partnership and knowledge transfer (De la Poza et al., 2021). This is perhaps the single most important empirical finding for the purposes of this article. #Disclosure_behaviour is not a side effect of performance. It is part of performance. 3. Method This is a conceptual and documentary study rather than an empirical one. It combines three sources of material. First, a structured review of peer reviewed literature published between 2020 and 2025 on the Impact Rankings, on global university rankings more generally, on the sustainable development goals in higher education, and on the entrepreneurial university. Sources were identified through academic databases using combinations of terms covering impact rankings, sustainability rankings, SDG reporting in universities, third mission and quantification in higher education. Priority was given to studies that examined the ranking directly, and to reviews that synthesised broader bodies of work. Second, documentary analysis of the published methodology documents and results announcements for the ranking across its editions, including the scoring rules, the participation requirements, the indicator structure for individual goals, and the announcements accompanying the 2026 transition to a ratings and membership model. These documents are the primary rulebook that institutional actors respond to, and they are treated here as the object of strategy rather than as neutral background. Third, published institutional accounts, including statements by university leaders and sector commentary reported alongside the rankings, which describe in the institutions' own words what they did and why. The analysis proceeds by reading the methodology as an incentive structure, asking at each scoring rule what a rational and reasonably ethical administrator would do in response, and then checking that inferred behaviour against what the empirical literature and the institutional accounts actually report. Where the literature confirms the inferred behaviour, it is presented as an observed practice. Where it does not, it is flagged as a hypothesis. This approach has clear limits, which are stated in Section 9. It cannot establish frequency, and it cannot separate sincere action from strategic action inside any given institution. What it can do is map the option space that educational entrepreneurs operate in. 4. The Anatomy of the Instrument Strategy is a response to rules. To understand the strategy, the rules must be set out plainly. 4.1 Origins and growth The exercise began in 2019, when a set of performance metrics developed the previous year was published as a league table at a summit held at the Korea Advanced Institute of Science and Technology. The first edition included 450 universities from 76 countries, a self selected group willing to submit data and be scrutinised. In 2020 the scope was extended to cover all seventeen goals. By the 2022 edition the overall table included 1,406 universities from 106 countries and territories, led by Western Sydney University. The seventh edition, published in June 2025, included more than 2,500 institutions from 130 countries and territories. In the space of six years the ranking went from a pilot to the largest voluntary data collection exercise in global higher education. The eighth edition, published in 2026, arrived under a new name. The Impact Rankings became the Sustainability Impact Ratings, and participation became membership based through a newly created Sustainability Impact Network, with tiered annual fees set according to World Bank country income classifications. Institutions in high income and upper middle income nations pay 3,500 pounds a year, those in lower middle income nations pay 1,000 pounds, and those in low income nations pay nothing. The 2026 edition covered 1,646 universities from 116 countries and territories, a substantial fall in participation from the previous year. The University of Manchester finished first overall, ending Western Sydney University's four year run at the top, with Griffith University second and Western Sydney third. The Philippines contributed more institutions than any other single country, with 160. The publisher stated that the underlying #methodology was not changed by the move to ratings, and explained the renaming on the grounds that the participating group is self selecting rather than comprehensive, which makes the word ratings more accurate than the word rankings. 4.2 The four pillars and the seventeen tables Performance on each goal is assessed across four broad areas. #Research covers the volume, quality and use of publications relevant to that goal. #Stewardship covers how the institution manages its own resources, including its estate, its staff and its students. #Outreach covers work with local, regional, national and international communities. #Teaching covers the preparation of graduates who can deliver on the goals and carry sustainability into their careers. Each goal has its own set of indicators drawn from these areas, with fixed weights inside the goal. Results are published as seventeen individual tables plus one overall table. Data comes from two sources. Bibliometric indicators are supplied by a commercial publisher's citation database, using field weighted measures and subject classifications. Everything else is submitted by the institution itself, in the form of numbers and supporting evidence, and assessed by the ranking provider, whose evaluation of the evidence is final. 4.3 The scoring rule that drives everything This is the rule that shapes institutional behaviour more than any other. To be included in the overall table, a university must submit data for #SDG_17, which covers partnerships, and for at least three other goals. The overall score is then calculated from SDG 17, which contributes twenty two per cent, plus the institution's three highest scoring other goals, each contributing twenty six per cent. Read that again slowly, because almost everything in Section 5 follows from it. A university's overall position depends on partnerships plus its own best three goals. It does not depend on its worst goals. It does not depend on the thirteen goals it chose not to enter. An institution that performs superbly on gender equality, quality education and good health, and abysmally on climate action, responsible consumption and life below water, will not be penalised in the overall table for the second list, provided it simply does not submit to those goals or provided they do not rank among its best three. This is the structural feature that critics have identified as allowing high overall placement with limited contribution across most of the agenda (Bautista-Puig et al., 2022). It is also, from the institution's point of view, an entirely legal and openly published rule that would be irrational to ignore. A second consequence follows. Because the ranking assesses submitted evidence, an institution that does excellent work but submits nothing scores zero on that dimension. Where a data point is not provided, a value of zero is entered. #Silence_is_scored. This converts the reporting function from an administrative afterthought into a core determinant of the result. 5. Seven Families of Optimisation Practice The practices below are grouped into families. Each family is described, its evidentiary basis noted, and its position on the line between substance and cosmetics assessed. Most institutions use several families at once, and the same practice can be substantive in one institution and hollow in another. That ambiguity is the point. 5.1 Portfolio selection: choosing which goals to fight on The first and most consequential decision is which of the seventeen goals to enter, and which to invest in. A naive institution enters everything. A sophisticated one treats the goal set as an #investment_portfolio. It assesses where its existing research strengths lie, where its existing community work is concentrated, where its region has visible needs, and where the indicator set is achievable given its data. It then concentrates effort on a small number of goals where it can plausibly reach the top band, while entering others at lower intensity to build the evidence base for future years. This behaviour is visible in the data. Universities steadily increase the number of goals they submit to as they gain experience in the ranking, a pattern more consistent with growing reporting capability than with the sudden creation of new sustainability programmes (Dickson, 2025). Studies of submission behaviour have found significant and hard to explain disparities in how institutions choose to participate, which is what one would expect if #portfolio_choice were being managed strategically rather than reported passively (Urbano et al., 2025). Is this gaming? Not necessarily. A small university in a water stressed region that concentrates on clean water, climate action and partnerships is behaving exactly as good strategy and good ethics both recommend. Nobody benefits from a mid sized teaching institution pretending to have a marine biology programme. Concentration of effort is what strategy means. The problem arises only when concentration becomes concealment, when a genuinely poor performance on a goal is quietly kept out of the submission in order to protect the overall score. The instrument cannot distinguish between the two, and that is a design weakness rather than an institutional failing. 5.2 Building the evidence infrastructure The second family is the least visible and the most transformative. It concerns the machinery of #evidence. The submission requires an institution to state, for example, how many first generation students it admitted, how much water it recycled, what proportion of its energy came from renewable sources, how many of its policies address the rights of specific groups, whether it tracks food waste, and how it engages with local government on planning. Most universities do not hold this information in one place. It lives in the estates department, the human resources system, the student records system, the research office, the finance system and the private files of individual academics. The entrepreneurial move is to build a permanent internal architecture that collects this material continuously rather than in a panic each autumn. In practice this means appointing an owner for each goal, usually a senior academic or professional staff member; establishing a data dictionary so that terms mean the same thing across departments; embedding SDG tagging in the research information system so that publications and projects are classified as they are created rather than retrospectively; adding sustainability fields to procurement, estates and human resources reporting; and creating a central evidence repository with version control and clear provenance. This is expensive, slow and unglamorous. It is also the single practice most likely to produce real change, because an institution that measures something continuously will eventually manage it. The literature on strategic integration of the goals makes the same point from the other direction: integration only works when it is written into strategy and monitored, rather than delegated to enthusiasts (Leal Filho et al., 2023). #Measurement_infrastructure is the bridge between rhetoric and behaviour. It is also the practice most sharply divided along lines of institutional wealth. A university that can fund a ranking office, a data analyst and a sustainability team will out report a university that cannot, even if the second university does more good in the world. One senior scholar quoted in the sector commentary around the 2026 results made this precise point, observing that simply having the administrative capacity to tell a good story is beyond the human resources of many institutions. #Reporting_capacity is itself unequally distributed, which means the ranking partly measures administrative wealth. 5.3 Documentation and narrative construction Where the previous family concerns numbers, this one concerns text. Many indicators are assessed through evidence rather than counting. The assessor asks whether the university has a policy, whether it publishes it, whether it applies it and whether it can show that it does. Points are awarded according to answers, and the ranking provider's judgement of the evidence is final. This converts institutional writing into a scoring activity. Educational entrepreneurs respond by professionalising documentation. Policies that existed informally are written down and published. Web pages that described activities in vague language are rewritten to state clearly what was done, for whom, when and with what result. Annual sustainability reports are produced in a format that maps directly onto the indicator set. Case studies are prepared with named partners, dated activities and measurable outcomes. There is an obvious risk here and it should be named. #Narrative_optimisation can substitute for action. A policy can be written, published and never applied. A case study can inflate a modest workshop into a regional transformation. The corporate sustainability literature calls the general phenomenon SDG washing, and it describes exactly this: selective, enthusiastic reporting on flattering goals accompanied by silence elsewhere (Heras-Saizarbitoria et al., 2022). Because the ranking assesses documents, an institution skilled at documents has an advantage independent of its skill at outcomes. The honest version of this practice is also real, though. Many universities genuinely did not have a policy on, say, workers' rights in their supply chain, and wrote one because the ranking asked. Many genuinely did not publish their emissions figures, and now do. The requirement to say something in public creates accountability that did not previously exist. #Forced_disclosure has consequences, and some of them are good. 5.4 The partnership engine and the SDG 17 problem SDG 17 is compulsory and carries twenty two per cent of the overall score, more than double the weight of any other individual goal in relative terms once one accounts for the fact that only four goals count. Every serious competitor must therefore build a partnership function. The goal is assessed through the university's relationships with government, its cross sectoral collaboration, its international research collaboration, its publication of data on the goals, and its educational work on sustainable development beyond its own campus. In practice this drives several behaviours. Institutions formalise relationships with municipal and national government that were previously informal. They join international networks and consortia. They co author with overseas colleagues, which conveniently also improves performance in conventional rankings. They publish open data on their own SDG performance, which is itself a scored activity. The strategic elegance of SDG 17 is that it is the most gameable goal and simultaneously the most genuinely valuable. Signing a memorandum of understanding costs nothing and produces a document. Building a working relationship with a city government that changes municipal policy costs a great deal and produces a document that looks almost the same. #Partnership_inflation is a predictable outcome. Empirical modelling of what distinguishes top placed institutions found that they are more oriented toward transferring knowledge to industry and toward supporting strong institutions in their countries, which suggests the goal does capture something real, but it also suggests that the institutions best placed to exploit it are those already embedded in dense partnership ecosystems (De la Poza et al., 2021). 5.5 Curriculum embedding The teaching pillar rewards institutions that produce graduates equipped to deliver on the goals. Optimisation here takes several forms. At the shallow end, existing modules are relabelled. A module on business ethics becomes a module on responsible business and the goals, with no change to the reading list. Course catalogues are audited and every course that mentions anything relevant is tagged. This is cheap, fast and produces a large number very quickly. At the deeper end, institutions redesign programmes. They introduce required sustainability content across all degrees, create interdisciplinary problem based courses in which students work on local community problems, build in service learning with assessed community placements, and establish graduate attributes that all programmes must demonstrate. Some create new degrees entirely. This is slow, contested inside faculties, and genuinely changes what students learn. Both approaches raise the reported number. Only one changes the graduate. Research on the implementation of the goals in university programmes has stressed the difference between adding sustainability as a topic and embedding it as a way of thinking, and has offered frameworks for doing the latter (Leal Filho, Salvia and Frankenberger, 2021). The wider survey evidence suggests that in most institutions integration still stems from individual academics and small groups rather than from the whole institution, which is precisely the condition under which #curriculum_tagging outruns #curriculum_change (Leal Filho, Salvia and Eustachio, 2023). For students reading this article, the practical test is straightforward. Look past the ranking position and ask a specific question of the admissions office: which required modules, in the degree I would take, include assessed sustainability content, and what exactly is assessed? A university that has embedded the work will answer immediately. A university that has tagged the catalogue will send you a brochure. 5.6 Campus operations and stewardship The stewardship pillar covers the university as a physical and social organisation: its energy, water, waste, procurement, employment practices and treatment of students and staff. This is the family where optimisation is hardest to fake, because the underlying quantities are physical. A university cannot claim to have cut water consumption by forty per cent without meters that say so. Consequently this is where the ranking has arguably produced the most solid change. Institutions have installed metering, moved to renewable supply contracts, introduced waste tracking, changed procurement rules, and published estates data that was previously internal. Two forms of optimisation still exist here. The first is #boundary_setting: choosing which parts of the institution count as the institution. Does the joint venture campus count? Does the leased building count? Does the contracted catering company count? Different answers produce different numbers, and the choice is often defensible either way. The second is baseline selection, where an institution reports improvement against a year that happened to be unusually bad. Neither is dishonest in a legal sense. Both illustrate the general truth that quantified comparison invites definitional manoeuvre. 5.7 Recognition management: branding, recruitment and reputation The final family is the one that connects the score back to the market, and it is the reason the whole enterprise pays for itself. A strong Impact Ranking position is a marketing asset with unusually broad reach. For universities that will never appear in the top 200 of a conventional world ranking, it is often the only global table on which they can credibly claim excellence. It is used in international student recruitment, in bids for research funding, in negotiations with government, in donor communications and in staff recruitment. The publisher's own survey of participants in May 2025 reported that eighty four per cent of respondents regarded the exercise as an important tool for showcasing their contribution to the goals, which is a candid statement of what the ranking is for from the participants' point of view. Educational entrepreneurs therefore invest in the recognition layer: press releases timed to publication, campaign pages, social media, badges on prospectuses, sector conference appearances, and case studies written for the ranking's own publications. #Recognition_management is not incidental to impact strategy. In many institutions it is the business case that funds the rest of the work, which means the substantive programmes in Sections 5.2 to 5.6 exist partly because the marketing value in Section 5.7 justified their budget. This is an uncomfortable but important observation. #Reputational_incentive is often the mechanism by which real sustainability work gets funded at all. 6. A Maturity Framework for Impact Optimisation The seven families above can be arranged into a developmental sequence. The framework below is offered as an analytical tool rather than a validated instrument. Its purpose is to give students, administrators and researchers a shared vocabulary for distinguishing what looks similar from the outside. Stage 1. Absent. The institution does not participate. It may do substantial community work, but none of it is visible externally. Data lives in individual memories. Nothing is published. Stage 2. Cosmetic. The institution participates. A small team assembles a submission close to the deadline, drawing on whatever evidence can be found. Existing activities are relabelled with goal numbers. Policies are written to satisfy indicators and then filed. The score exists; the institution has not changed. This stage produces #SDG_washing in its purest form and is detectable by the gap between what the website claims and what students and staff experience. Stage 3. Systematic. The institution builds evidence infrastructure. Goals get owners. Data is collected continuously. Submissions improve year on year because the machinery improves, not because activity improves. The score rises substantially at this stage, which is why the stage is so attractive and so dangerous. It is entirely possible for an institution to reach a strong global position without having materially altered its contribution to society. This is the stage most consistent with the observed pattern of institutions steadily increasing their submitted goals over time (Dickson, 2025). Stage 4. Integrated. Measurement begins to change management. Because emissions, gender pay gaps, community reach and graduate outcomes are now visible to leadership every quarter, they enter budget discussions and performance reviews. Deans are held to account for them. Curriculum change moves from tagging to redesign. The institution starts to do differently what it previously only described differently. The score may rise more slowly at this stage than at Stage 3, which is a critical and counterintuitive point: the transition from reporting well to acting well can be #ranking_neutral in the short term. Stage 5. Transformative. The goals become the organising logic of the institution's strategy rather than a reporting framework applied to it. Research priorities, hiring, estate investment, partnerships and curriculum are set with reference to a defined societal mission, usually anchored in the institution's region. The ranking becomes a by product rather than a target. Institutions at this stage are typically willing to report goals on which they perform badly, because their legitimacy no longer depends on the table. The willingness to publish an unflattering number is, in my view, the single most reliable external indicator that an institution has reached this stage. The framework carries one clear warning. Stages 3 and 4 look identical from outside, and the score does not distinguish them. Any external user of the ranking who wants to know whether an institution is at Stage 3 or Stage 4 must look at evidence the ranking does not collect: staff turnover in sustainability roles, whether academic promotion criteria reward engagement, whether the institution publishes bad results, and whether students experience the curriculum described in the submission. 7. Discussion: Recognition, Impact and the Space Between 7.1 The decoupling problem The central tension identified in this article can now be stated formally. The ranking rewards a joint product of #real_impact and #documented_impact. Because the second component can be improved far more cheaply and quickly than the first, and because the score cannot distinguish the two, institutions face a persistent incentive to over invest in documentation relative to action. Over time this produces #decoupling: the reported institution and the actual institution drift apart. The literature supports the plausibility of this mechanism without settling the question of how common it is. Modelling shows that disclosure is strongly associated with score (De la Poza et al., 2021). Content analysis shows that the scoring structure permits high placement with limited breadth of contribution (Bautista-Puig et al., 2022). Submission pattern analysis shows behaviour more consistent with improving reporting than with improving performance (Dickson, 2025). At the same time, careful empirical work using two years of data concludes that the ranking does provide a robust overview of institutional sustainability performance, which suggests that decoupling has not swamped the signal (Urbano et al., 2025). The reasonable conclusion is that the ranking contains real information and real noise in proportions that vary by institution, and that it should be used as one input rather than as a verdict. 7.2 The equity question and the 2026 turn The move in 2026 to a membership model, with tiered fees and participation limited to network members, sharpens a question that was already present. #Participation_cost has always been non financial and high: staff time, systems, coordination. The reforms add a direct financial cost for all but low income countries, while explicitly designing the fee structure so that wealthier institutions subsidise the poorest and low income nations pay nothing. The immediate effect on coverage was visible. Participation fell from more than 2,500 institutions across 130 countries in 2025 to 1,646 institutions across 116 countries in 2026. Some of that fall may reflect institutions deciding the exercise was not worth the cost, which is a healthy outcome. Some may reflect institutions that valued the exercise but could not fund it, which is not. At the same time, the 2026 data showed movement in the opposite direction on outcomes. High income countries held 34 of the top 50 places, down from 39 the previous year. The number of lower middle income institutions in the top 50 doubled to four, and the number of upper middle income institutions in the top 200 rose from 34 to 38. The Philippines contributed more institutions than any other country. Commentators described this as a gradual diffusion of capacity rather than a change in leadership, which sounds right. #Global_South institutions are getting better at this exercise, and the exercise remains one of the very few global tables on which they can compete at all. Both things are true, and the tension between them is not resolvable by better methodology. There is also a governance point worth stating plainly. When the organisation that publishes a ranking also sells membership of the network required to enter it, and also sells the analytics, benchmarking and conference access that help institutions improve within it, the incentives of the evaluator and the evaluated become entangled. This is not an accusation of bad faith. It is an observation that the responsible rankings literature, which has consistently called for transparency and independence in evaluation design, would regard the arrangement as requiring unusually strong safeguards (Gadd, 2021). 7.3 What the ranking cannot see Several important things fall outside the instrument by construction. It cannot see #outcomes as distinct from activities. A university can report the number of community programmes it runs, but not whether the community is better off. Genuine impact evaluation of the kind used in development economics is expensive and slow, and no global ranking could sustain it. It cannot see academic freedom, or the quality of intellectual life, or whether an institution's research on climate is being quietly funded by interests that oppose climate action. It cannot see whether the gender equality policy is enforced when a senior professor breaks it. It cannot see #opportunity_cost. The staff time spent assembling a submission is time not spent teaching or researching. For a well resourced institution this is a rounding error. For a small institution in a low income country it may be a serious diversion of scarce administrative capacity, undertaken because the ranking is the only route to global visibility. Finally, it cannot see the institutions that do excellent work and choose not to participate. The tables describe a self selected group, a fact the publisher itself acknowledged when it renamed the exercise. Any student or policymaker who treats the table as a complete map of good universities is making an error of interpretation that no amount of methodological improvement will fix. 7.4 Is optimisation legitimate? I want to answer this directly rather than hedge. Optimisation is legitimate when the reported picture is true. If an institution builds systems that let it accurately describe the good it does, and then describes it in the format the assessor requires, it has done nothing wrong and has probably done itself and its region a favour along the way. The evidence infrastructure it builds will outlast the ranking. Optimisation is illegitimate when it creates a false picture, whether by strategic silence about known poor performance, by inflating trivial activity, by writing policies it does not enforce, or by concentrating institutional energy on the appearance of impact at the expense of impact itself. The line is not always easy to locate in practice, but it exists, and people inside institutions generally know which side of it they are on. The honest conclusion is that the ranking makes the wrong thing easy and the right thing hard, and then pays both the same. That is a design problem, not a moral failing of the people who work within it. 8. Implications 8.1 For students Students are the group most often told to use these tables and least often told how they work. Five practical points follow from this analysis. First, check which goals the university actually submitted to, not just its overall position. The individual goal tables are published. An institution with a strong overall score and entries in only four goals is telling you something specific about itself. Second, remember that #non_participation is not evidence of poor performance. Excellent institutions choose not to enter for reasons ranging from cost to principle. Third, test claims about #curriculum. Ask which required modules contain assessed sustainability content in the specific degree you would take. Fourth, look for the unflattering number. An institution that publishes its emissions, its gender pay gap and its student attrition, including when they are bad, is showing you a different kind of institution than one that publishes only awards. Fifth, use the ranking for what it is genuinely good at: finding institutions whose declared mission points toward the problems you care about. It is a decent search tool and a poor verdict. 8.2 For institutional leaders and administrators For those doing this work, the analysis suggests several priorities. Build the evidence infrastructure first, and build it as a permanent management system rather than a submission project. If the data only exists in the autumn, it is not a management system. Choose a portfolio that is honest about the institution's region and strengths, and defend that choice publicly. A clear and well argued concentration is more credible than a scattered claim to universal excellence. Watch the Stage 3 trap. If the score is rising and nothing inside the institution has changed, the score is measuring the reporting office. Set internal targets on outcomes that the ranking does not reward, and hold deans to them. Publish what looks bad. It costs little in the table, because unsubmitted or weak goals mostly do not count against the overall score, and it buys the one thing money cannot buy, which is #credibility. Recognise the labour. The people assembling submissions are typically overworked professional staff whose contribution is invisible in academic promotion systems. If the institution's global reputation rests on their work, its reward systems should say so. 8.3 For policymakers and funders Ministries increasingly cite these positions in strategy documents and in funding decisions. Three cautions apply. The table measures a self selected group and rewards a joint product of action and documentation. It should not be used as a proxy for national sustainability performance in higher education. Funding schemes that reward ranking position transmit the design flaws of the ranking directly into public policy, and will produce more reporting rather than more impact. Funding that rewards verified outcomes, even crude ones, does better. If a system wants its universities to serve society, the cheapest intervention is usually to change academic promotion criteria so that engagement counts. The literature has been saying this about the third mission for years (Compagnucci and Spigarelli, 2020; Perkmann et al., 2021), and no ranking will substitute for it. 9. Limitations This study is conceptual and documentary. It cannot establish how frequently each practice occurs, and it cannot measure the size of the decoupling effect it describes. The maturity framework in Section 6 is an analytical proposal, not a validated instrument, and it has not been tested against institutional data. The reliance on published methodology documents means the analysis reflects the rules as stated rather than as applied, and the ranking provider's internal assessment practices are not publicly observable. The institutional accounts drawn on are, by their nature, produced by parties with an interest in appearing successful. Finally, the analysis is written from outside the institutions it describes. A programme of interview based research with ranking officers, sustainability directors and senior leaders would test, and very likely complicate, much of what is argued here. That research is needed and does not yet exist in any substantial form. 10. Conclusion The Impact Rankings, now the Sustainability Impact Ratings, did something that no other global table had done. They told a large number of universities that the work they do for their communities counts, and they gave that work a number. For institutions long excluded from global visibility by rankings built around research volume and inherited prestige, this was not a small thing. It opened a route to #global_recognition that had been closed. It also created a market, and markets create specialists. The #educational_entrepreneurs described in this article are the specialists. They read the scoring rules, choose a portfolio, build the evidence machinery, write the narrative, cultivate the partnerships, redesign the curriculum, retrofit the campus and sell the result. Most of them believe in the work. Many of them use the ranking as a lever to get sustainability funded internally, which is a defensible and often admirable use of a flawed instrument. The difficulty is structural and will not be fixed by tinkering with weights. The ranking pays equally for doing good and for being able to describe doing good, and describing is cheaper. Until an instrument exists that can tell those apart, the number will contain both signal and noise, and the people who read it will need to do the work of separating them. For students, that means reading the tables with curiosity rather than deference, and asking the specific questions in Section 8.1. For institutions, it means resisting the temptation to stop at Stage 3, where the score is high and nothing has changed. For the sector, it means remembering what the ranking was invented to encourage. As one university leader put it when his institution reached the top of the 2026 ratings, the exercise asks institutions to consider what universities are good for, and not merely what they are good at. That is a question worth taking seriously, and it is not one that any table can answer on a university's behalf. Hashtags #Impact_Rankings #Sustainability_Impact_Ratings #Times_Higher_Education #Sustainable_Development_Goals #educational_entrepreneurship #university_rankings #higher_education_strategy #societal_impact #SDG_washing #global_recognition #university_leadership #sustainability_in_higher_education #ranking_optimization #academic_integrity #students_guide_to_rankings References Bautista-Puig, N., Orduna-Malea, E. and Perez-Esparrells, C. (2022) Enhancing sustainable development goals or promoting universities? An analysis of the Times Higher Education impact rankings. International Journal of Sustainability in Higher Education, 23(8), 211-231. https://doi.org/10.1108/IJSHE-07-2021-0309 Bellantuono, L., Lo Sasso, A., Amoroso, N., Monaco, A., Tangaro, S. and Bellotti, R. (2025) Network assortativity for a multidimensional evaluation of socio-economic territorial biases in university rankings. PLOS ONE, 20(6), e0323356. https://doi.org/10.1371/journal.pone.0323356 Caputo, F., Ligorio, L. and Pizzi, S. (2021) The contribution of higher education institutions to the SDGs: An evaluation of sustainability reporting practices. Administrative Sciences, 11(3), 97. https://doi.org/10.3390/admsci11030097 Chaleta, E., Saraiva, M., Leal, F., Fialho, I. and Borralho, A. (2021) Higher education and sustainable development goals (SDG): Potential contribution of the undergraduate courses of the School of Social Sciences of the University of Evora. Sustainability, 13(4), 1828. https://doi.org/10.3390/su13041828 Chankseliani, M. and McCowan, T. (2021) Higher education and the Sustainable Development Goals. Higher Education, 81(1), 1-8. https://doi.org/10.1007/s10734-020-00652-w Chun, H. and Sauder, M. (2022) The logic of quantification: Institutionalizing numerical thinking. Theory and Society, 51, 335-370. https://doi.org/10.1007/s11186-021-09453-1 Compagnucci, L. and Spigarelli, F. (2020) The third mission of the university: A systematic literature review on potentials and constraints. Technological Forecasting and Social Change, 161, 120284. https://doi.org/10.1016/j.techfore.2020.120284 De la Poza, E., Merello, P., Barbera, A. and Celani, A. (2021) Universities reporting on SDGs: Using THE Impact Rankings to model and measure their contribution to sustainability. Sustainability, 13(4), 2038. https://doi.org/10.3390/su13042038 Dickson (2025) Meaningful measurement of sustainable development or gaming the system? Patterns in submission to Times Higher Education's Impact Ranking. Sustainable Development. https://doi.org/10.1002/sd.70419 Gadd, E. (2020) University rankings need a rethink. Nature, 587, 523. Gadd, E. (2021) Mis-measuring our universities: Why global university rankings do not add up. Frontiers in Research Metrics and Analytics, 6, 680023. https://doi.org/10.3389/frma.2021.680023 Hazelkorn, E. (2023) The power in managing numbers: Changing interdependencies and the rise of ranking expertise. Higher Education, 86(4), 771-790. Heras-Saizarbitoria, I., Urbieta, L. and Boiral, O. (2022) Organizations engagement with sustainable development goals: From cherry-picking to SDG-washing? Corporate Social Responsibility and Environmental Management, 29(2), 316-328. https://doi.org/10.1002/csr.2202 Leal Filho, W., Salvia, A. L. and Eustachio, J. H. P. P. (2023) An overview of the engagement of higher education institutions in the implementation of the UN Sustainable Development Goals. Journal of Cleaner Production, 386, 135694. https://doi.org/10.1016/j.jclepro.2022.135694 Leal Filho, W., Salvia, A. L. and Frankenberger, F. (2021) A framework for the implementation of the Sustainable Development Goals in university programmes. Journal of Cleaner Production, 299, 126915. Leal Filho, W., Simaens, A., Paco, A., Hernandez-Diaz, P. M., Vasconcelos, C. R. P., Fritzen, B. and Mac-Lean, C. (2023) Integrating the Sustainable Development Goals into the strategy of higher education institutions. International Journal of Sustainable Development and World Ecology. https://doi.org/10.1080/13504509.2023.2167884 Leal Filho, W., Trevisan, L. V., Rampasso, I. S., Anholon, R., Dinis, M. A. P., Brandli, L. L., Sierra, J., Salvia, A. L., Pretorius, R., Nicolau, M., Eustachio, J. H. P. P. and Mazutti, J. (2023) When the alarm bells ring: Why the UN Sustainable Development Goals may not be achieved by 2030. Journal of Cleaner Production, 407, 137108. https://doi.org/10.1016/j.jclepro.2023.137108 Patara, S. and Dhalla, R. (2022) Sustainability reporting tools: Examining the merits of sustainability rankings. Journal of Cleaner Production, 366, 132960. Perkmann, M., Salandra, R., Tartari, V., McKelvey, M. and Hughes, A. (2021) Academic engagement: A review of the literature 2011-2019. Research Policy, 50(1), 104114. https://doi.org/10.1016/j.respol.2020.104114 Urbano, V. M., Arena, M., Azzone, G. and Mayeres, M. (2025) Sustainable development in higher education: An in-depth analysis of Times Higher Education Impact Rankings. Journal of Cleaner Production, 501, 145302. https://doi.org/10.1016/j.jclepro.2025.145302

Vocational education and training systems were designed for an industrial world in which job tasks changed slowly and a qualification earned at the age of twenty could reasonably support a career of forty years. That world is disappearing. The spread of machine learning, and more recently of generative systems that write text, produce code, and analyse images, is reshaping the task content of occupations far faster than most training institutions can revise a syllabus. This article asks a practical question: how can #vocational_education be redesigned so that it produces workers who remain valuable in an economy where routine cognitive work is increasingly automated? The paper uses a structured narrative review of scholarly and policy literature published mainly between 2020 and 2025, combined with a comparative reading of institutional practice, to build a conceptual framework. The framework treats reform not as an administrative exercise but as an act of #entrepreneurial_strategy inside education, borrowing from theories of disruptive innovation, effectuation, and lean venture design. Five reform levers are identified: modular and stackable credentials, employer-embedded delivery, teacher and trainer capability renewal, assessment redesign around judgement rather than recall, and financing models that reward outcomes instead of enrolment. The article argues that the central skill of the #AI_economy is not the ability to perform a task but the ability to define, delegate, evaluate, and correct work that machines partially perform. Training systems that continue to certify #task_execution alone will certify obsolescence. The paper closes with implications for policymakers, institutional leaders, employers, and students, and with an agenda for future research. Keywords: vocational education and training, artificial intelligence, skills, micro-credentials, entrepreneurship in education, labour market, lifelong learning, apprenticeship, curriculum reform, workforce development 1. Introduction For most of the twentieth century, the logic of skills training was stable. An employer needed people who could weld, machine a part, keep accounts, service an engine, or operate a switchboard. A college or a training centre taught those tasks. A certificate confirmed that the holder could do them. Employers trusted the certificate. Students trusted the employers. Governments funded the arrangement because it worked. The arrangement worked because of an assumption that is now under strain: that the tasks being taught would still exist, in recognisable form, for a long time after graduation. #Automation has always eroded some tasks, but it did so slowly enough that training institutions could adjust between cohorts. What is different now is the pace and the breadth. Systems built on large statistical models do not automate a single machine operation. They automate fragments of many occupations at once, including occupations that were previously considered safe because they involved language, analysis, or creativity. The evidence for this shift is no longer speculative. Eloundou, Manning, Mishkin, and Rock (2024) estimated that roughly eighty per cent of the United States workforce could have at least one tenth of their work tasks affected by large language models, and that around one fifth of workers could see half of their tasks affected. Felten, Raj, and Seamans (2021) mapped exposure to artificial intelligence across occupations and found that the pattern of exposure was quite different from earlier waves of automation, reaching into higher-paid analytical work rather than concentrating in manual routine work. The OECD (2023) reported that around a quarter of employment across member countries sits in occupations with high exposure to automation technologies, and that the composition of that exposure is shifting. The World Economic Forum (2025) reported that employers expect a large proportion of core job skills to change before the end of the decade and that a majority of the global workforce will require some form of training to keep pace. At the same time, controlled studies show that these systems increase productivity in real work rather than only in laboratory conditions. Noy and Zhang (2023) found large reductions in the time needed to complete mid-level writing tasks alongside gains in output quality. Brynjolfsson, Li, and Raymond (2025) studied customer support agents using a generative assistant and found meaningful average productivity gains, with the largest gains concentrated among the least experienced workers. Peng, Kalliamvakou, Cihon, and Demirer (2023) reported that developers using an AI coding assistant completed a standard programming task substantially faster than a control group. Dell'Acqua and colleagues (2023) found that consultants using a generative model performed better on tasks inside the model's competence but worse on tasks that fell just outside it, unless they had learned how to check the model's output. That last finding is the hinge of this article. The productivity benefit of an intelligent tool depends on whether the human user knows where the tool is reliable and where it is not. That knowledge is not a technical skill in the traditional sense. It is a form of professional judgement. It cannot be taught by demonstration and repetition alone, and it is not easily captured by a checklist-based competency standard. Yet checklist-based competency standards are the backbone of most #TVET systems in the world. This creates a mismatch. Training systems certify the reliable performance of defined tasks. The #labour_market increasingly rewards the reliable supervision of #partially_automated_work. The mismatch is not merely a curriculum problem. It runs through funding formulas, accreditation rules, teacher qualifications, assessment methods, employer engagement, and the basic unit of account in education, which is still the multi-year programme leading to a single terminal award. 1.1 Purpose and research questions This article treats the reform of vocational training as an entrepreneurial problem rather than an administrative one. Administrative reform asks how to improve an existing model. Entrepreneurial reform asks what the model would look like if it were designed today, for today's constraints, by someone with no institutional inheritance to protect. Both perspectives are necessary, but the second is under-represented in the literature on skills policy. The article addresses four questions. First, what specifically has changed in the demand for skills, and how confident can we be about the evidence? Second, why do traditional #vocational_education systems adapt slowly, and what are the structural causes of that slowness rather than the symptoms? Third, which #entrepreneurial_strategies, drawn from theories of innovation and venture design, can be transferred into skills-based education without damaging the public purposes that vocational systems serve? Fourth, what would an integrated reform framework look like in practice, and how could it be evaluated? 1.2 Contribution Three contributions are intended. The first is synthetic: the article brings together three literatures that rarely speak to one another, namely the economics of automation, the scholarship of vocational pedagogy, and the entrepreneurship literature on venture design and #disruptive_innovation. The second is conceptual: the article proposes a framework, described later as the Adaptive Vocational Enterprise model, that treats a training provider as a venture operating in an uncertain market rather than as a bureaucracy delivering a fixed product. The third is practical: the article translates the framework into decisions that a principal, a dean, a training manager, or a ministry official could act upon within an ordinary planning cycle. 1.3 Structure Section 2 reviews the literature. Section 3 sets out the method, which is a structured narrative review with a conceptual synthesis. Section 4 presents the analysis in five themes. Section 5 develops the framework. Section 6 discusses the framework against counter-arguments and risks. Section 7 sets out implications for practice and policy. Section 8 states limitations and a research agenda. Section 9 concludes. 2. Literature Review and Conceptual Background 2.1 The changing task content of work The modern economics of automation begins with the observation that technology does not replace jobs; it replaces tasks. A job is a bundle of tasks, and technology changes the composition of the bundle. Acemoglu and Restrepo (2020) provided influential evidence from industrial robots in United States labour markets, showing measurable negative effects on employment and wages in exposed local markets. Their broader theoretical work distinguishes between technologies that displace labour from tasks and technologies that create new tasks in which labour has an advantage. The balance between these two forces determines whether technological change raises or lowers the demand for workers. Autor (2022) argued that the future of work is characterised less by mass unemployment than by deep uncertainty about which tasks will remain in human hands and what those tasks will pay. Autor, Mindell, and Reynolds (2022) extended this into a policy argument, contending that the quality of jobs in an automated economy depends on institutional choices, including choices about training, rather than on technology alone. Bessen (2022) made a related point from the perspective of firms, noting that the diffusion of complex software tends to concentrate advantage among large firms that can afford to build the surrounding organisational capability, which has consequences for where good jobs appear. The arrival of general-purpose generative systems changed the shape of the argument. Earlier automation waves hollowed out middle-skill routine work while leaving both manual and analytical work relatively intact. Generative systems reach directly into analytical and communicative work. Eloundou and colleagues (2024) described these systems as general-purpose technologies whose exposure is broad rather than concentrated. Autor (2024) offered a more optimistic reading, arguing that if such systems can put expert judgement within reach of workers without elite credentials, they could rebuild middle-skill work rather than destroy it. Whether that optimistic path is realised depends heavily on whether #skills_based_education equips people to use the tools well. For vocational systems, the practical consequence is that the half-life of a taught task is shrinking. Deming and Noray (2020) showed that in fast-changing technical fields, the earnings advantage of a technical qualification decays as the specific content of the field turns over, and that workers who cannot refresh their skills lose ground. This is an argument for #lifelong_learning that rests on #evidence_based_practice rather than slogan. 2.2 What vocational education is for Vocational education has never had a single purpose. It has served economic purposes, by supplying #skilled_labour; social purposes, by offering routes into #decent_work for people who did not thrive in academic settings; and civic purposes, by forming people who can participate in an occupational community. Reform proposals that ignore any of these purposes tend to fail politically even when they succeed technically. McGrath and colleagues (2020) argued for a capabilities-based theory of vocational education, in which the aim is not only #employability but the expansion of what people are able to be and do. This matters for the present argument because a purely economic framing invites a narrow response to automation, namely training people for whatever tasks machines cannot yet perform, which is a race that has no finish line. A capabilities framing asks instead what durable human capacities are worth building. Brown, Lauder, and Cheung (2020) offered a sharp critique of the standard human capital story. They argued that the promise that more education automatically yields better jobs has weakened, because the value of a credential depends on the structure of the labour market, not only on the knowledge it certifies. Their argument is uncomfortable for #skills_based_education because it suggests that supplying more trained workers does not by itself create good work. It follows that training reform cannot be evaluated only by counting graduates. Guile and Unwin's collected scholarship on vocational learning, and the long tradition of workplace learning research associated with Billett, established that occupational expertise is built through participation in real practice, not only through instruction. This is one of the most robust findings in the field, and it has direct implications for the design of #work_based_learning in an automated economy, because judgement about machine output is precisely the kind of knowledge that develops through supervised participation rather than through lecture. 2.3 Credentials, micro-credentials, and the unit of account If task content changes quickly, the natural response is to make credentials smaller and faster to update. This is the logic behind #micro_credentials. Oliver (2022), writing for UNESCO, proposed a common definition and argued that micro-credentials must be assessed, quality assured, and meaningful to employers if they are to be more than marketing. Cedefop (2022) examined the role of micro-credentials in labour market education and training and found both promise and risk: promise in responsiveness, risk in fragmentation and in the difficulty of establishing genuine currency with employers. Wheelahan and Moodie (2021) provided the most serious critique. Drawing on the sociology of knowledge, they argued that micro-credentials tend to certify discrete competencies detached from the systematic knowledge that gives an occupation its coherence. A worker who holds twenty micro-credentials may still lack the organising knowledge that lets them make sense of a new situation. Their warning is not an argument against modularity as such, but against modularity without a spine. Any serious proposal for #stackable_credentials must therefore explain how the modules add up to something more than the sum of their parts. The European Union's approach to micro-credentials, together with national systems such as Singapore's lifelong learning framework, illustrates a middle path in which small units of learning are anchored to a national qualifications structure and to funding entitlements. The details vary, but the design principle is consistent: modularity works when it is attached to a #competency_frameworks architecture that defines how modules relate to one another and to occupational progression. 2.4 Digital and AI literacy as vocational content A second strand of literature asks what should now be taught. Long and Magerko (2020) proposed a definition of #AI_literacy centred on a set of competencies that allow a non-specialist to recognise, evaluate, and interact with artificial intelligence critically. Ng, Leung, Chu, and Qiao (2021) reviewed the emerging literature and organised AI literacy into knowing and understanding, using and applying, evaluating and creating, and ethical awareness. Van Laar and colleagues (2020) had already shown that #digital_skills for workers are not a single construct but a layered one, in which technical operation sits beneath information management, communication, collaboration, problem solving, and critical evaluation. The practical lesson for #vocational_education is that AI literacy is not a standalone course. A plumber, a nurse, a logistics coordinator, and a paralegal all need to evaluate machine output, but what counts as a good evaluation differs by occupation. Generic modules on the history of artificial intelligence will not build that judgement. AI literacy has to be embedded in occupational practice, which is a curriculum design problem rather than a content-acquisition problem. The critical literature is a necessary counterweight here. Selwyn (2022) cautioned that educational technology discourse tends to overstate benefits and understate the political and labour consequences of adoption. Holmes and Tuomi (2022) distinguished between what artificial intelligence in education can actually do and what it is claimed to do, and argued that many claims outrun the evidence. Bond and colleagues (2024), in a meta-review of systematic reviews, found persistent weaknesses in the rigour and ethical treatment of research in the field. These cautions should temper any reform proposal that assumes technology will solve the problems that technology helped create. 2.5 Entrepreneurship as a lens for institutional reform The entrepreneurship literature offers three concepts that transfer well to education reform. The first is #disruptive_innovation. Christensen, McDonald, Altman, and Palmer (2018) revisited the theory and clarified that disruption describes a process in which an entrant serves an overlooked segment with a simpler and cheaper offering and then improves until it satisfies mainstream demand. The theory has been misapplied often, but its core insight is relevant: incumbents are rarely defeated because they are incompetent. They are defeated because the criteria by which they judge good performance are the criteria of their existing customers. A public training college that measures success by enrolment in accredited multi-year programmes will systematically undervalue short, flexible, employer-embedded offerings, even when those offerings serve learners better. The second is effectuation. Sarasvathy (2022) described expert entrepreneurs as reasoning from available means towards emergent goals, rather than from fixed goals towards required means. Under genuine uncertainty, where the future distribution of outcomes is unknown, effectual reasoning outperforms predictive planning. Skills forecasting in an AI economy is a case of genuine uncertainty. Nobody knows with confidence which tasks will be automated in five years. An effectual training provider would therefore stop trying to forecast precisely and start building partnerships, running small experiments, and expanding what works. The third is the lean venture logic of building a minimum viable offering, measuring its effect, and iterating (Ries, 2011). The point is not the vocabulary but the cycle time. A training programme that takes three years to design, five years to accredit, and four years to run has a feedback loop of over a decade. In a domain where task content changes every two or three years, such a loop is not slow; it is broken. Chalmers, MacKenzie, and Carter (2021) examined how artificial intelligence changes venture creation itself, noting that the tools alter which entrepreneurial tasks require human input. Obschonka and Audretsch (2020) argued that data-rich environments change how entrepreneurial opportunities are recognised. Ratten and Jones (2021) reviewed entrepreneurship education and found a persistent gap between teaching about entrepreneurship and developing an #entrepreneurial_mindset through practice. Their finding mirrors the vocational literature exactly: knowing about something is not the same as being able to do it under uncertainty. 2.6 Synthesis of the gap The literature therefore establishes four things. Task content is changing quickly and broadly. Vocational systems are built around slow, task-centred certification. Modular credentials are a plausible response but risk incoherence. Entrepreneurial reasoning offers tools for acting under uncertainty that have not been systematically applied to vocational reform. What is missing is an integrated framework that connects the diagnosis to the institutional levers. Most policy documents describe what should change without specifying who has the authority to change it, what it costs, and how the change would be recognised as successful. This article attempts that connection. 3. Methodology 3.1 Research design This is a conceptual paper built on a structured narrative review. It does not claim to be a systematic review in the technical sense, and readers should not treat it as one. The aim is theory building rather than effect estimation. Three bodies of evidence were assembled. The first is empirical research on the labour market effects of artificial intelligence, drawn from economics journals, working paper series of established research institutes, and peer-reviewed experimental studies. Priority was given to studies with identified comparison groups or with transparent exposure measures. The second is scholarship on vocational education, workplace learning, credentials, and skills policy, drawn from specialist journals in the field and from the research outputs of international agencies with a long record in the area. The third is the entrepreneurship and innovation literature, drawn from management journals, used as a source of transferable concepts rather than as evidence about education. 3.2 Selection and inclusion Sources were selected to satisfy four criteria. They had to be published in the last five years, with a small number of exceptions where a theoretical anchor predates that window and has no adequate recent substitute. They had to be traceable to an identifiable author and outlet. They had to make a claim that could be checked against other sources. And they had to be relevant to at least one of the four research questions. Grey literature from international organisations was included because much of the best evidence on skills demand exists only in that form. It was treated with appropriate caution: statistics from employer surveys were reported as expectations rather than facts, and projections were not treated as findings. 3.3 Analytical approach Analysis proceeded in three stages. In the first, claims were extracted from each source and grouped by the aspect of the problem they addressed, producing an initial set of candidate themes. In the second, themes were consolidated where they overlapped and split where they concealed distinct mechanisms, yielding the five themes presented in Section 4. In the third, the entrepreneurship concepts were mapped onto the themes to identify which institutional lever each concept could plausibly move, producing the framework in Section 5. 3.4 Limitations of the method The method has three known weaknesses. Narrative reviews are vulnerable to selection bias, and although the inclusion criteria were stated in advance, no formal search protocol was registered. The evidence base on generative systems is very young, and several of the most cited studies are recent, which means their findings may not replicate. Finally, the framework is derived from literature rather than tested in the field, so it should be read as a hypothesis about what works rather than a demonstration that it does. These limits are revisited in Section 8. 4. Analysis: Five Structural Problems and Their Roots 4.1 Theme one: the curriculum clock runs slower than the technology clock Every training system has a curriculum clock, which is the time between recognising a change in occupational practice and delivering teaching that reflects it. In most public systems that clock runs for years. Industry bodies must agree a revised standard. Awarding organisations must revise qualifications. Regulators must approve them. Providers must retrain staff, buy equipment, and rewrite materials. Each step is defensible on its own. Together they produce a delay that guarantees obsolescence in a fast-moving field. The problem is not laziness. It is that the system was optimised for a different objective, namely consistency. A #national_qualification is valuable precisely because it means the same thing in every region. Consistency requires central approval, and central approval takes time. The trade-off between consistency and speed is real, and pretending otherwise is how reform proposals lose credibility with practitioners. The way out is not to abandon consistency but to change what is held constant. If the fixed element is the detailed task list, the system will always lag. If the fixed element is a set of durable capabilities, with task-level content delegated to providers and employers under a quality framework, then the system can update content quickly while keeping the qualification stable. This shifts the locus of #curriculum_agility from the centre to the edge, and it requires the centre to accept a supervisory rather than a prescriptive role. This is uncomfortable. It also has a precedent. Aviation, medicine, and financial services all maintain rapidly changing technical content inside stable professional standards, by requiring continuing professional development and periodic #revalidation rather than by rewriting the initial qualification every time a procedure changes. Vocational systems could adopt the same architecture. 4.2 Theme two: assessment measures the wrong thing #Competency_based_assessment asks whether a learner can perform a task to a specified standard. This is an excellent method when the task is stable and the standard is unambiguous. It becomes a poor method when the task is performed jointly by a human and a machine and the human contribution is supervisory. Consider a bookkeeping student. The traditional assessment asks the student to prepare a set of accounts. Software can now prepare a plausible draft in seconds. Assessing the student on the preparation of the draft is assessing something the labour market no longer pays for. What the market pays for is the ability to notice that the draft has misclassified a lease, to know why that matters, and to correct it. That is a different assessment task, and it is harder to write, harder to mark, and harder to standardise. The same pattern appears across occupations. In software, the value has moved from writing a function to specifying, reviewing, and testing it. In healthcare support roles, from recording observations to noticing when a machine-generated alert is implausible. In marketing, from producing copy to judging whether copy fits a brand and an audience. In each case the assessable skill is judgement about output produced with assistance, exercised under real constraints. The Dell'Acqua study is instructive here because it showed that the difference between benefiting from a machine and being misled by it lies exactly in this capacity. #Assessment_reform is therefore not a technical detail. It is the mechanism by which a system signals what it values. If assessment continues to reward unassisted task performance, teaching will continue to prepare students for unassisted task performance, whatever the curriculum documents say. Practical redesign points in three directions. Assessment should permit and often require the use of the tools that the workplace uses, since prohibiting them measures a performance that will never occur again after graduation. Assessment should include deliberately flawed #machine_outputs that the candidate must identify and repair. And assessment should be conducted, wherever possible, on real work with real consequences, which is the strongest argument for expanding #work_based_learning rather than simulating it. 4.3 Theme three: the teaching workforce is the binding constraint Reform documents routinely assume that if the curriculum changes, teaching will follow. It will not. The trainers in most vocational systems entered from industry, often a decade or more ago, and their industrial knowledge is frozen at the point of entry. Many are employed on terms that leave no time for continuing engagement with practice. Many teach in institutions that cannot afford the tools their students will use at work. This is the single most underestimated obstacle to modernisation. A curriculum that requires trainers to teach the critical evaluation of machine output cannot be delivered by trainers who have never used such a tool in anger. No amount of policy will change this. Only #instructor_capacity investment will, and that investment has a specific shape. It requires paid time. Trainers need protected hours for practice engagement, not evening webinars added to a full teaching load. It requires industry return, in the form of structured placements where trainers work in a firm for a period and bring the experience back. It requires dual-role employment, where practitioners teach part time while continuing to work, which is common in some systems and rare in others. And it requires that these things be treated as core cost rather than as discretionary professional development that is cut in the first budget round. There is a hard-headed argument here that appeals to finance directors. A trainer whose knowledge is five years out of date is producing graduates whose employment outcomes will decline, and in systems where funding is linked to outcomes, that decline is a financial risk. Investment in trainer currency is therefore not welfare spending. It is maintenance of the productive asset. 4.4 Theme four: the financing model rewards the wrong behaviour Most public training is funded by enrolment or by completion of accredited programmes. This creates two distortions. The first is a bias towards #long_programmes. A provider funded per enrolled student in a two-year programme has no financial reason to offer a six-week programme that gets the same student into work faster, even if the six-week programme is better for the student. The incentive is to lengthen, not to sharpen. The second is a bias against experimentation. A provider that wants to run a short pilot in a new field faces the full cost of design with no funding until the programme is accredited, which may take a year or more. Under those conditions no rational manager pilots anything. The system is not hostile to innovation; it simply does not pay for it. Reforming the funding model is therefore not a technical adjustment but the precondition for everything else. Three mechanisms deserve attention. #Outcome_linked_funding pays providers partly on employment and earnings outcomes rather than only on enrolment. It aligns provider incentives with student interests, but it must be designed carefully, because crude outcome measures encourage providers to select easy students and avoid difficult ones. Any serious design must adjust for intake characteristics and must protect #equity explicitly. #Individual_learning_entitlements give funding to the learner rather than the institution, allowing them to purchase training across providers over a working life. This creates a market signal that responsive providers can capture, and it supports #lifelong_learning structurally rather than rhetorically. The risk is that learners without good information make poor choices, so entitlements require independent guidance to work. #Innovation_funds provide small, fast grants for pilots with an evaluation requirement attached, so that a provider can test a new offering without waiting for accreditation. This is the direct application of #minimum_viable_program logic to a public system. The evaluation requirement is what distinguishes it from ordinary grant spending. 4.5 Theme five: employer engagement is treated as consultation rather than co-production Almost every vocational system has employer advisory committees. Very few have employers who are jointly responsible for delivery. The difference matters enormously. A consultative employer attends a meeting twice a year and comments on a draft standard. A co-producing employer provides equipment, releases staff to teach, hosts learners for substantial periods, sets real problems for assessment, and hires the graduates. Only the second arrangement produces the tight feedback loop that fast-changing occupations require. The obstacle is cost. Co-production is expensive for employers, and firms will only bear that cost if they capture enough of the benefit. Where labour is mobile and trained workers can be poached, individual firms under-invest. This is a classic collective action problem, and it explains why the systems with the strongest employer involvement, such as the German-speaking dual systems, rest on strong employer associations and social partnership rather than on individual firm goodwill. Countries without that institutional inheritance cannot simply import it. But they can build partial versions of it through sector skills bodies, #group_training_organisations that employ apprentices and place them across several firms, and tax or levy arrangements that spread the cost of training across the firms that benefit from it. The design principle is to make #public_private_partnership structurally rational rather than dependent on the civic virtue of individual employers. The AI economy sharpens this argument. Employers now hold something training providers cannot buy: current tools, current data, current workflows, and current problems. A training system that is not embedded in workplaces will teach yesterday's version of the technology on last year's software. #Ecosystem_partnerships are therefore not a nice addition to the model. They are the model. 5. The Adaptive Vocational Enterprise Framework 5.1 Overview The five problems above are interconnected, which is why single-lever reforms fail. Modular credentials without assessment reform simply chop an outdated curriculum into smaller pieces. Assessment reform without trainer investment produces standards nobody can teach. Trainer investment without funding reform is unaffordable. Funding reform without employer co-production rewards providers for outcomes they cannot influence. The framework proposed here therefore treats the training provider as an enterprise operating under uncertainty, with five interacting subsystems. It is called the #Adaptive_Vocational_Enterprise, and it has four layers. 5.2 Layer one: the durable capability spine At the centre of the model is a small set of capabilities that are expected to retain value even as tasks change. These form the spine to which all modular content is attached, which answers the coherence objection raised by Wheelahan and Moodie. Six capabilities are proposed, derived from the reviewed evidence rather than from aspiration. The first is #problem_framing, which is the ability to turn a messy situation into a defined task that can be delegated, whether to a person or a machine. This is the entry point for all assisted work, and it is the skill that determines whether a tool is useful. The second is #critical_evaluation of output, which is the ability to detect error, bias, omission, and implausibility in work that was produced with assistance. This capability sits at the heart of the Dell'Acqua finding and is the strongest single candidate for the core vocational skill of the decade. The third is #contextual_domain_knowledge, which is the systematic understanding of the field that makes evaluation possible. A machine output can only be judged as wrong by someone who knows what right looks like. This is the answer to the fantasy that domain knowledge is now obsolete because the machine holds it. The machine holds a representation of it. The worker must hold enough to audit the representation. The fourth is #tool_fluency, which is the practical ability to use current instruments, including the ability to learn the next one quickly. Note that fluency is deliberately placed fourth. It is the fastest-decaying capability and therefore the least suitable as an organising principle for a qualification, even though it is what most reform rhetoric focuses on. The fifth is #ethical_reasoning and regulatory reasoning, which is the ability to recognise when an assisted decision raises questions of privacy, safety, fairness, accountability, or legal compliance, and to know what to do about it. As #data_governance obligations tighten across sectors, this moves from a soft skill to a licence-to-operate skill. The sixth is #adaptive_learning capacity, meaning the disposition and the practical method to acquire new skills without formal instruction. Deming and Noray's evidence on skill decay implies that this capability determines lifetime earnings in technical fields more than any specific technique does. These six form the assessable spine. They are not generic soft skills. Each is expressed differently in each occupation, and each is assessed in occupational context. 5.3 Layer two: the modular content shell Around the spine sits a shell of short, updatable modules covering current tools, current regulations, current procedures, and current sector practice. These are the parts of the curriculum expected to change every one to three years. Modules in the shell should be designed to be replaced rather than revised. This is a design decision borrowed from software: components with short lifespans should be loosely coupled so that replacing one does not require redesigning the whole. In practice this means each module has explicit inputs, explicit outputs, and an explicit link to one or more spine capabilities, so that when the tool changes the module can be swapped without touching the qualification. The shell is where #micro_credentials belong. Attached to the spine, they are coherent. Detached from it, they are the fragmentation that critics rightly fear. The rule is simple: no module is offered unless it can state which spine capability it develops and how that development is assessed. 5.4 Layer three: the delivery enterprise The delivery layer is where entrepreneurial method applies most directly. #Effectual_reasoning replaces predictive planning. Instead of asking which skills will be needed in 2030, which nobody knows, the provider asks what it can build now with the partners, staff, equipment, and learners it actually has. It then commits small amounts, observes results, and expands what works. This is not a licence for aimlessness. It is a disciplined response to genuine uncertainty, and it is what expert entrepreneurs do (Sarasvathy, 2022). Short cycles replace long ones. A new offering is designed as a #minimum_viable_program: the smallest version that can genuinely test whether learners gain and employers hire. It runs with a small cohort, is evaluated against pre-declared measures, and is either scaled, revised, or stopped. The discipline lies in the willingness to stop, which is the part of lean method that public institutions find hardest, because in a bureaucracy a closed programme is a failure rather than a resolved experiment. #Portfolio_thinking replaces single bet planning. A provider should run a portfolio of offerings across a risk spectrum, with a large core of proven programmes funding a smaller set of experiments. This is standard practice in venture-backed organisations and almost unknown in public training, where every programme is expected to succeed and therefore none is genuinely new. Distribution is treated as a first-class problem. Entrepreneurs know that a good product without distribution fails. In training, distribution means the pathway by which a learner discovers the programme, believes it will help, can afford it, can attend it, and can convert it into a job. Most vocational marketing addresses only the first of these. #Learner_agency is not served by a brochure; it is served by clear information about outcomes, transparent costs, flexible scheduling, and a credible route to employment. 5.5 Layer four: the governance and trust layer The entrepreneurial layers create risk, and the governance layer contains it. This is the part most often omitted by enthusiastic reformers, and its omission is why many market-based education reforms have produced poor outcomes for vulnerable learners. Four guarantees are proposed. Quality assurance must apply to the spine, not to the shell. Regulators should verify that a provider can evidence the six capabilities and can assess them credibly. They should not attempt to approve every module, because doing so reinstates the slow clock the reform is designed to remove. #Quality_assurance shifts from input approval to outcome verification, with sampling and audit rather than universal pre-approval. #Outcome_transparency must be mandatory and comparable. Providers should publish employment and earnings outcomes, disaggregated by learner background, in a common format. Without this, an outcome-funded market rewards good marketing rather than good teaching. Equity protection must be explicit and funded. Outcome-linked funding without intake adjustment will push providers to avoid learners who need the most help. The adjustment must be generous enough to make such learners financially attractive to serve, not merely tolerable. #Portability must be guaranteed. Credentials earned with one provider must be recognised by others, which requires a national or sectoral registry of learning records. Without portability, modularity traps learners in a single provider's ecosystem, which is the opposite of the intended effect. 5.6 The framework in one sentence Hold the capabilities stable, let the content churn, give providers freedom to experiment on the content, hold them accountable for the capabilities and the outcomes, and pay for the trainer time and employer partnership that make it possible. 6. Discussion 6.1 Does this simply repackage existing ideas? Several elements of the framework are familiar. Modularity, employer engagement, and outcome funding all appear in existing policy. The claim to novelty rests on three points. First, the framework specifies the interaction between levers rather than listing them. It argues that each lever fails in isolation and states why. Policy documents typically present a menu; this article presents a system with dependencies. Second, the framework identifies the capability spine as the thing to hold constant, and it derives that spine from evidence about how assisted work actually goes wrong, rather than from a generic list of twenty-first century skills. The elevation of critical evaluation of machine output to a central assessable capability is, as far as the reviewed literature shows, not yet standard in vocational qualification design. Third, the framework imports entrepreneurial method as a governance principle rather than as a subject to be taught. Most discussion of entrepreneurship in education concerns teaching students to start businesses. This article concerns teaching institutions to behave like ventures under uncertainty, which is a different and less explored proposition (Ratten and Jones, 2021). 6.2 The strongest objection: this will produce shallow training The most serious criticism of any modular, fast-cycling model is that it trades depth for speed. Wheelahan and Moodie's argument deserves to be answered directly rather than acknowledged and passed over. The answer is that the model deliberately protects depth in the spine and permits shallowness only in the shell. Contextual domain knowledge is one of the six capabilities precisely because the alternative is a worker who can operate a tool but cannot tell when it is wrong. If a provider builds a qualification entirely from shell modules with no spine, the model has been implemented incorrectly, and quality assurance should catch it. There is, however, a real risk that in practice the spine will be neglected, because spine capabilities are expensive to teach and difficult to assess, while shell modules are cheap and easy to sell. Market pressure will push providers towards the shell. This is why the governance layer must be strong, and why regulators should audit spine assessment specifically rather than counting modules. A reform that removes central curriculum control without strengthening central outcome verification will produce exactly the shallow training the critics predict. 6.3 The second objection: the evidence base is thin and young This is correct and should not be minimised. The strongest empirical studies of generative systems at work are recent, mostly conducted in a handful of firms and occupations, and often in settings favourable to the technology. Bond and colleagues (2024) documented weaknesses in rigour across educational artificial intelligence research generally. Selwyn (2022) warned that the field is prone to hype cycles. It is entirely possible that productivity effects observed in customer support or software development will not generalise, or that they will fade as tasks are reorganised. The framework is designed to be robust to that uncertainty, which is the main practical justification for effectual rather than predictive reasoning. If the technology advances more slowly than expected, a system organised around problem framing, critical evaluation, domain knowledge, and adaptive learning capacity is not damaged. Those capabilities were valuable before, and they remain valuable. The bet on the spine is a low-regret bet. The bet on any specific tool is not, which is why tool fluency is placed in the replaceable shell. 6.4 The third objection: institutions cannot behave like startups Public training institutions face constraints that no startup faces. They cannot select their students, they must serve regions that are not commercially attractive, they carry statutory obligations, and they cannot fail, because their failure strands learners. To ask them to be entrepreneurial can look like asking a hospital to be a technology company. The response is that the framework does not ask institutions to accept startup risk. It asks them to adopt startup method inside a bounded portion of their activity, funded specifically for that purpose, with governance that prevents learner harm. The experimental portfolio should be a minority of provision, ring-fenced, evaluated, and closed when it fails. This is closer to how well-run research and development functions inside a large firm than to how a startup operates. The distinction matters, because reform proposals that ignore public obligations are correctly resisted by the people who carry those obligations. 6.5 Equity, and who bears the cost of adjustment There is a distributional question running underneath this entire argument. The evidence from Brynjolfsson, Li, and Raymond (2025) suggests that assistive systems can compress performance differences by raising the productivity of less experienced workers. That is a hopeful finding for #equity, because it implies these tools could be levellers. But access to tools is not equally distributed, and neither is the training that turns access into capability. If the workers who most need the productivity gain are the ones least likely to receive good training in how to supervise machine output, the technology will widen gaps rather than close them. Brown, Lauder, and Cheung (2020) warned that the returns to education depend on the structure of opportunity, not only on the education itself. A training system that responds to the AI economy only by serving learners who already have advantages will have failed by its own stated purposes. The practical implication is that #upskilling provision for incumbent workers in lower-paid occupations should be treated as a first-order priority rather than as a residual category funded after initial training. This is the opposite of how most systems allocate money. 6.6 The role of the student Finally, the framework has implications for how students are addressed. A system that expects learners to return repeatedly across a working life must treat them as long-term participants rather than as a cohort to be processed once. That means keeping records they own, giving guidance that is independent of the provider selling the course, and being honest about outcomes, including when a programme has poor employment results. It also means being honest with students about the nature of the bet they are making. A student entering training today should be told plainly that the specific tools they learn will change, that the capability to relearn is the thing of lasting value, and that no qualification will be sufficient for a career. That is a harder message than the traditional promise, but it is a true one, and the traditional promise is becoming a false one. 7. Implications 7.1 For policymakers Change the unit of funding before changing the curriculum. Curriculum reform without funding reform produces documents, not practice. Introduce individual entitlements with independent guidance, add outcome-linked elements with strong intake adjustment, and create a dedicated innovation fund with mandatory evaluation. Move regulation from pre-approval of content to verification of outcomes. Publish comparable outcome data. Fund a national learning record so credentials are portable. Fund trainer currency as a core cost. Mandate and pay for industry return periods. Permit and encourage dual-role employment. Treat #policy_reform for incumbent workers as equal in priority to initial training, and target it at occupations where exposure is high and existing training provision is weak. 7.2 For institutional leaders Identify the spine capabilities for each of your occupational areas and check whether your assessment actually measures them. In most institutions it will not. That gap is the first project. Build a portfolio. Ring-fence a small proportion of provision for experiments, define in advance what success looks like, and be willing to close what does not work. The willingness to close is the whole discipline. Make employer co-production a condition of programme approval in fast-moving fields. If no employer will provide equipment, staff time, or real problems, that is information about whether the programme should run. Invest in your #trainers before you invest in technology. A well-equipped workshop taught by someone whose knowledge is a decade old is a waste of capital. 7.3 For employers Recognise that consultation is not enough. If you need workers who can supervise assisted work, you must let learners do assisted work in your workplace, on your tools, with your data governance rules, under supervision. Consider sectoral arrangements that spread training costs, because the poaching problem is real and will otherwise keep every individual firm under-investing. Hire on #demonstrated_capability rather than on programme length. If your recruitment continues to filter on qualification titles, providers will continue to produce qualification titles. 7.4 For students Choose programmes that let you use #current_tools and that assess your judgement about output rather than only your ability to produce output unaided. Ask providers directly what their graduates earn and whether they are employed. Treat any single credential as a starting point. Build the habit of learning without instruction, because that habit, more than any specific technique, determines what happens to your earnings across a working life. Learn the domain deeply enough to know when the machine is wrong. That is where the remaining human advantage sits, and it is the part that cannot be shortcut. 8. Limitations and Future Research The limitations of this article are substantial and should be stated plainly. The framework is derived from literature and reasoning, not from field trial. It has not been implemented anywhere in the integrated form described. Its components have precedents in different systems, but the interaction effects claimed in Section 5 are hypotheses. The evidence on generative systems is early. Several of the most influential studies were conducted in specific firms and occupations, and their generalisability is unknown. If the productivity findings do not replicate, the urgency of the argument weakens, though the case for the capability spine does not. The article draws heavily on evidence from high-income countries. Vocational systems in lower-income settings face different constraints, including infrastructure, connectivity, and the size of the #informal_sector, and the framework has not been tested against those conditions. Applying it without adaptation would be a mistake. The article does not address in detail the question of who owns the data generated by learners in assisted training environments, which is a serious #data_governance question that deserves separate treatment. Four research priorities follow. First, experimental or quasi-experimental evaluation of assessment redesign, comparing employment outcomes for graduates assessed on unassisted performance against those assessed on supervision of assisted work. This is the single most valuable study the field could run. Second, longitudinal study of #skill_decay and renewal in occupations with high AI exposure, building on the approach of Deming and Noray, to establish how quickly vocational content actually loses value. Third, evaluation of #micro_credentials attached to an explicit capability spine against micro-credentials offered standalone, to test the coherence argument empirically rather than theoretically. Fourth, research on trainer capability, which is badly under-studied relative to its importance. We know little about what interventions actually restore industrial currency in vocational teaching staff, and almost nothing about how long the effect lasts. 9. Conclusion Vocational education has always been a bet on the future. The bet was that the tasks taught today would still be worth money tomorrow. For a long time that bet paid off, because the tasks changed slowly enough for the system to keep up. The bet is now much harder to win at the task level, because tasks are being redefined faster than qualifications can be rewritten. The response cannot be to guess the future task list more accurately, because nobody can. The response has to be structural: hold constant the capabilities that survive task change, let the task-level content churn freely, and build institutions capable of changing that content quickly without waiting for permission. That is an entrepreneurial proposition, and it is uncomfortable for systems built on consistency and central control. But the discomfort of reform should be weighed against the alternative, which is a training system that continues to issue certificates for work that machines already do. Certificates of that kind do not protect anyone. They simply move the moment of disappointment from the classroom to the labour market, where the student bears it alone. The most valuable thing a #training_system can now give a young person is not a task. It is the ability to frame a problem, direct a powerful tool at it, know enough about the field to tell when the tool has failed, fix it, and then do the whole thing again next year with a tool that did not exist when they graduated. Building institutions that can teach that, assess it, and be funded for it is the work of the next decade. It will not be done by adjusting the syllabus. It will be done by rebuilding the machine that produces the syllabus, and that is a task for people willing to think like founders inside systems that were never designed for them. References Acemoglu, D. and Restrepo, P. (2020). Robots and jobs: Evidence from US labor markets. Journal of Political Economy, 128(6), 2188 to 2244. https://doi.org/10.1086/705716 Agrawal, A., Gans, J. and Goldfarb, A. (2022). Power and Prediction: The Disruptive Economics of Artificial Intelligence. Boston: Harvard Business Review Press. Autor, D. (2022). The labor market impacts of technological change: From unbridled enthusiasm to qualified optimism to vast uncertainty. NBER Working Paper No. 30074. Cambridge, MA: National Bureau of Economic Research. https://doi.org/10.3386/w30074 Autor, D. (2024). Applying AI to rebuild middle class jobs. NBER Working Paper No. 32140. Cambridge, MA: National Bureau of Economic Research. https://doi.org/10.3386/w32140 Autor, D., Mindell, D. A. and Reynolds, E. B. (2022). The Work of the Future: Building Better Jobs in an Age of Intelligent Machines. Cambridge, MA: MIT Press. Bacigalupo, M., Weikert Garcia, L., Mansoori, Y. and O'Keeffe, W. (2020). EntreComp Playbook: Entrepreneurial Learning Beyond the Classroom. Luxembourg: Publications Office of the European Union. Bessen, J. (2022). The New Goliaths: How Corporations Use Software to Dominate Industries, Kill Innovation, and Undermine Regulation. New Haven: Yale University Press. Bond, M., Khosravi, H., De Laat, M., Bergdahl, N., Negrea, V., Oxley, E., Pham, P., Chong, S. W. and Siemens, G. (2024). A meta systematic review of artificial intelligence in higher education: A call for increased ethics, collaboration, and rigour. International Journal of Educational Technology in Higher Education, 21, article 4. https://doi.org/10.1186/s41239-023-00436-z Brown, P., Lauder, H. and Cheung, S. Y. (2020). The Death of Human Capital? Its Failed Promise and How to Renew It in an Age of Disruption. New York: Oxford University Press. Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2), 889 to 942. https://doi.org/10.1093/qje/qjae044 Cedefop (2022). Microcredentials for Labour Market Education and Training: First Look at Mapping Microcredentials in European Labour Market Related Education, Training and Learning. Luxembourg: Publications Office of the European Union. Cedefop (2023). The Future of Vocational Education and Training in Europe: Synthesis Report. Luxembourg: Publications Office of the European Union. Chalmers, D., MacKenzie, N. G. and Carter, S. (2021). Artificial intelligence and entrepreneurship: Implications for venture creation in the fourth industrial revolution. Entrepreneurship Theory and Practice, 45(5), 1028 to 1053. https://doi.org/10.1177/1042258720934581 Christensen, C. M., McDonald, R., Altman, E. J. and Palmer, J. E. (2018). Disruptive innovation: An intellectual history and directions for future research. Journal of Management Studies, 55(7), 1043 to 1078. https://doi.org/10.1111/joms.12349 Dell'Acqua, F., McFowland, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F. and Lakhani, K. R. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Working Paper 24-013. Boston: Harvard Business School. Deming, D. J. and Noray, K. (2020). Earnings dynamics, changing job skills, and STEM careers. Quarterly Journal of Economics, 135(4), 1965 to 2005. https://doi.org/10.1093/qje/qjaa021 Ehlers, U. D. (2020). Future Skills: The Future of Learning and Higher Education. Karlsruhe: Springer. Eloundou, T., Manning, S., Mishkin, P. and Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306 to 1308. https://doi.org/10.1126/science.adj0998 Felten, E., Raj, M. and Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strategic Management Journal, 42(12), 2195 to 2217. https://doi.org/10.1002/smj.3286 Holmes, W. and Tuomi, I. (2022). State of the art and practice in AI in education. European Journal of Education, 57(4), 542 to 570. https://doi.org/10.1111/ejed.12533 International Labour Organization (2021). Shaping Skills and Lifelong Learning for the Future of Work. Geneva: International Labour Office. Long, D. and Magerko, B. (2020). What is AI literacy? Competencies and design considerations. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, 1 to 16. https://doi.org/10.1145/3313831.3376727 McGrath, S., Powell, L., Alla-Mensah, J., Hilal, R. and Suart, R. (2020). New VET theories for new times: The critical capabilities approach to vocational education and training and its potential for theorising a transformed and transformative VET. Journal of Vocational Education and Training, 72(4), 465 to 487. https://doi.org/10.1080/13636820.2020.1786440 Mollick, E. (2024). Co-Intelligence: Living and Working with AI. New York: Portfolio. Ng, D. T. K., Leung, J. K. L., Chu, S. K. W. and Qiao, M. S. (2021). Conceptualizing AI literacy: An exploratory review. Computers and Education: Artificial Intelligence, 2, 100041. https://doi.org/10.1016/j.caeai.2021.100041 Noy, S. and Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187 to 192. https://doi.org/10.1126/science.adh2586 Obschonka, M. and Audretsch, D. B. (2020). Artificial intelligence and big data in entrepreneurship: A new era has begun. Small Business Economics, 55, 529 to 539. https://doi.org/10.1007/s11187-019-00202-4 OECD (2021). OECD Skills Outlook 2021: Learning for Life. Paris: OECD Publishing. https://doi.org/10.1787/0ae365b4-en OECD (2023). OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market. Paris: OECD Publishing. https://doi.org/10.1787/08785bba-en Oliver, B. (2022). Towards a Common Definition of Micro-Credentials. Paris: UNESCO. Peng, S., Kalliamvakou, E., Cihon, P. and Demirer, M. (2023). The impact of AI on developer productivity: Evidence from GitHub Copilot. arXiv preprint arXiv:2302.06590. Ratten, V. and Jones, P. (2021). Entrepreneurship and management education: Exploring trends and gaps. The International Journal of Management Education, 19(1), 100431. https://doi.org/10.1016/j.ijme.2020.100431 Ries, E. (2011). The Lean Startup: How Constant Innovation Creates Radically Successful Businesses. London: Portfolio Penguin. Rios, J. A., Ling, G., Pugh, R., Becker, D. and Bacall, A. (2020). Identifying critical 21st-century skills for workplace success: A content analysis of job advertisements. Educational Researcher, 49(2), 80 to 89. https://doi.org/10.3102/0013189X19890600 Sarasvathy, S. D. (2022). Effectuation: Elements of Entrepreneurial Expertise, 2nd edition. Cheltenham: Edward Elgar Publishing. Selwyn, N. (2022). The future of AI and education: Some cautionary notes. European Journal of Education, 57(4), 620 to 631. https://doi.org/10.1111/ejed.12532 Susskind, D. (2020). A World Without Work: Technology, Automation and How We Should Respond. London: Allen Lane. UNESCO (2023). Guidance for Generative AI in Education and Research. Paris: UNESCO. van Laar, E., van Deursen, A. J. A. M., van Dijk, J. A. G. M. and de Haan, J. (2020). Determinants of 21st-century skills and 21st-century digital skills for workers: A systematic literature review. SAGE Open, 10(1). https://doi.org/10.1177/2158244019900176 Vuorikari, R., Kluzer, S. and Punie, Y. (2022). DigComp 2.2: The Digital Competence Framework for Citizens. Luxembourg: Publications Office of the European Union. https://doi.org/10.2760/115376 Wheelahan, L. and Moodie, G. (2021). Analysing micro-credentials in higher education: A Bernsteinian analysis. Journal of Curriculum Studies, 53(2), 212 to 228. https://doi.org/10.1080/00220272.2021.1887358 World Economic Forum (2025). The Future of Jobs Report 2025. Geneva: World Economic Forum. Hashtags #vocational_education_reform #skills_for_the_AI_economy #modernizing_TVET #entrepreneurial_education_strategy #future_ready_skills #micro_credentials_that_work #AI_and_the_labour_market #work_based_learning_matters #lifelong_learning_now #disrupting_skills_training #teach_judgement_not_tasks #employer_co_production #outcome_based_funding #adaptive_vocational_enterprise #students_and_the_AI_economy

Since 2019, the Times Higher Education Impact Rankings have grown from a pilot exercise involving 450 universities into the most widely used global instrument for measuring how higher education institutions contribute to the United Nations Sustainable Development Goals. In 2026 the exercise was renamed the Sustainability Impact Ratings and moved to a membership model, a change that reveals how deeply the market for #global_recognition has become institutionalised. This article examines a group of actors who sit at the centre of that market but are rarely studied directly: the #educational_entrepreneurs who design, staff, fund and manage institutional participation in the ranking. Drawing on a structured review of peer reviewed research published between 2020 and 2025, together with documentary analysis of published ranking methodologies and results, the article maps the concrete practices these actors use. These include portfolio selection across the seventeen goals, the construction of evidence infrastructure, the production of documentary narratives, the strategic cultivation of partnerships, the embedding of sustainability in curricula, and the reorganisation of campus operations. The article then develops a five stage maturity framework that distinguishes cosmetic compliance from genuine institutional transformation. The central finding is that the ranking rewards two distinct capabilities at once: the capability to act on society, and the capability to document that action. Because the second capability is cheaper to build than the first, the instrument creates a persistent risk of decoupling between reported and real #societal_impact. The article closes with practical guidance for students who read these tables, for administrators who manage submissions, and for policymakers who use the results as evidence. Keywords: #Impact_Rankings, #Sustainable_Development_Goals, #higher_education, #educational_entrepreneurship, #university_rankings, #institutional_strategy, #societal_impact, #SDG_reporting 1. Introduction A university in Manila, a technical institute in Amman, a regional campus in Queensland and a research intensive institution in Manchester now share something they did not share ten years ago. Each of them can, if it chooses, submit a large body of evidence about its work on poverty, gender equality, clean water, decent work, climate action and eleven other goals to a single London based data operation, and receive back a numerical score and a place in a published table. That table is read by prospective students, by ministries of education, by donors, by employers and by the institutions themselves. It has become, in a very short time, one of the few ways a young or non elite university can appear on a #global_ranking at all. This article is about the people who make that submission happen and the choices they make along the way. I call them #educational_entrepreneurs, using the term in a broad sense. They are vice chancellors and rectors who decide that sustainability will be the institution's competitive position. They are pro vice chancellors for engagement, directors of sustainability, heads of institutional research, communications leads, and increasingly the specialist ranking officers whose entire job is to manage submissions. They are also faculty members who build community programmes and then have to translate that work into an auditable evidence file. What unites them is that they operate at the meeting point of two logics. One logic is #mission_driven and asks what the university is good for. The other is #market_driven and asks how the university will be seen. The #Impact_Rankings sit exactly on the seam between the two. The argument of this article is simple to state and harder to work through. The ranking measures two things that are related but not the same. The first is real activity in the world: research that gets used, graduates who change practice, campuses that cut emissions, partnerships that shift policy. The second is the institution's capacity to notice, collect, formalise and narrate that activity in the specific format an external assessor will accept. These are different organisational capabilities. The second one is significantly cheaper and faster to acquire than the first. Any institution that wants a better score therefore faces a fork in the road at every planning cycle, and the incentives do not point cleanly in one direction. This is not a claim that universities are dishonest. It is a claim about what happens when a complex social good is turned into a comparable number, a problem that scholars of quantification have described in many other fields. Existing research on the ranking has concentrated on its methodology and on whether its scores are valid measures of sustainability. That work is essential and this article draws on it heavily. What is missing is a systematic account written from the inside of the institution, one that treats optimisation not as a scandal to be exposed but as an ordinary #management_practice that can be done well or badly, honestly or dishonestly, with substance or without it. Students in particular deserve this account. They are told to use these tables to choose where to study. They are rarely told how the numbers are produced or which behaviours produce them. The article proceeds as follows. Section 2 reviews the relevant literature on rankings as calculative devices, on the entrepreneurial university, and on the specific critiques of the Impact Rankings. Section 3 explains the method. Section 4 sets out the anatomy of the instrument itself, because no strategy can be understood without understanding the scoring rules it responds to. Section 5 presents seven families of optimisation practice observed in the literature and in published institutional accounts. Section 6 offers a maturity framework. Section 7 discusses the risks, including #SDG_washing, equity effects and the consequences of the 2026 shift to a paid membership model. Section 8 draws out implications for students, institutions and policymakers. Section 9 states limitations and Section 10 concludes. 2. Literature Review and Conceptual Background 2.1 Rankings as devices that change what they measure The starting point for any serious discussion of #university_rankings is that they are not neutral mirrors. They are instruments that reorganise the thing they claim to describe. When a number is published and audiences begin to treat it as real, organisations reorganise themselves around the number, and the number then becomes a more accurate description of the organisation than it was when first invented. This is a well documented pattern in the sociology of quantification, and researchers have traced how numerical thinking becomes institutionalised inside organisations until it feels like plain common sense rather than a choice (Chun and Sauder, 2022). Applied to higher education, this means a ranking does not simply find the good universities. It teaches universities what good means and then rewards the ones that learn fastest. Hazelkorn has shown how the growth of rankings created an entirely new class of expertise inside institutions, with people whose professional standing depends on their ability to read, interpret and manipulate ranking data, and who therefore have an interest in the ranking continuing to matter (Hazelkorn, 2023). This is not a criticism of those individuals. It is a description of how #ranking_expertise becomes a durable feature of the sector. The methodological critique of global rankings is now mature. Gadd's assessment of the major global tables against agreed principles of responsible evaluation found that all of them fell short, and that the most visible ones fell shortest, largely because they blend unrelated indicators into a single composite score without a defensible theory of why those weights and not others (Gadd, 2021). Composite indicators of this kind are known to be sensitive to small changes in weighting, which means that a university can move several places without changing anything it actually does. The same work notes that rankings tend to favour institutions that are older, wealthier, English speaking and located in North America and Western Europe, which raises a #structural_equity problem that no amount of technical adjustment fully solves. More recent quantitative work using network methods has continued to find socio economic and territorial bias embedded in the way ranking scores distribute (Bellantuono et al., 2025). The relevant conclusion for this article is not that rankings are worthless. It is that #reactivity is the normal condition of ranked organisations, and that optimisation is therefore not a deviation from the system but the system working as designed. 2.2 The entrepreneurial university and the third mission The second literature this article draws on concerns the changing purpose of the university. Alongside teaching and research, institutions have been asked for several decades to take on a #third_mission that covers technology transfer, community engagement, policy advice and social innovation. A systematic review of this literature identified both the potential of the third mission and the constraints that limit it, including weak internal incentives, unclear metrics and the difficulty of persuading academic staff that engagement counts as real work (Compagnucci and Spigarelli, 2020). Research on #academic_engagement has similarly shown that the collaborative, advisory and community facing activities of academics are widespread but poorly captured by conventional performance systems (Perkmann et al., 2021). This creates a gap that the Impact Rankings entered. Universities were being told to matter to society but had no accepted way to show that they did. Traditional world rankings measure research volume, citation impact, reputation surveys, staff to student ratios and international outlook. None of these register whether a university reduced local water poverty or trained enough nurses for its region. The arrival of an instrument that promised to make the third mission visible was, for many institutions, genuinely welcome. That welcome is important. It explains why participation grew so quickly, and it explains why criticism of the ranking is often met with defensiveness from people who see it as the first tool that ever noticed their work. The #entrepreneurial_university concept is usually associated with commercialisation, spin outs and industry income. In this article I use a wider reading. An educational entrepreneur is someone who identifies an unmet need or an unclaimed position, assembles resources, accepts risk and builds something that did not exist before. Under that definition, the sustainability director who builds an institution wide SDG evidence system, secures a budget for it, persuades deans to cooperate and turns a middling institution into a globally visible one is doing entrepreneurial work, whether or not any money is made. 2.3 Higher education and the Sustainable Development Goals The 2030 Agenda gave universities a common vocabulary. Scholars have argued that higher education can contribute to sustainable development through research, teaching and community development, but that the sector still lacks agreed ways to document and evidence the enormous variety of activity involved (Chankseliani and McCowan, 2021). A large survey of academics across six regions found that engagement with the goals is real but uneven, frequently driven by individual enthusiasm rather than institutional strategy, and rarely supported by systematic measurement (Leal Filho, Salvia and Eustachio, 2023). Work on strategic integration has stressed that the goals only change behaviour when they are written into institutional strategy, resourced and monitored, rather than being announced and then left to volunteers (Leal Filho et al., 2023). This matters for the present argument. If SDG activity in most universities is fragmented and undocumented, then the first thing any ranking submission does is force the institution to look at itself. That act of looking has value independent of the score. Several institutions have reported that the exercise revealed programmes their own leadership did not know existed. #Self_knowledge is an underrated by product of #SDG_reporting. 2.4 The specific critique of the Impact Rankings The dedicated critical literature on this ranking is small but sharp. The most cited study conducted a content analysis of the methodology and concluded that the scoring structure allows a university to reach a high overall position while contributing little or nothing across most of the goals, because only four of the seventeen goals feed the overall score, and three of those four are the institution's own best performing goals (Bautista-Puig, Orduna-Malea and Perez-Esparrells, 2022). The same study raised concerns about the appropriateness and weighting of the research metrics used, and about the wide score bands that make positions difficult to interpret meaningfully. A later analysis using two years of data confirmed that the ranking does capture something real about institutional sustainability performance, while also finding large disparities in participation behaviour that are hard to explain by sustainability performance alone (Urbano et al., 2025). Work examining submission patterns directly has asked whether the observed behaviour represents meaningful measurement or #gaming_the_system, noting that universities steadily increase the number of goals they submit to over time in a way that is unlikely to reflect the sudden appearance of new sustainability activity, and more likely reflects growing skill at reporting (Dickson, 2025). Beyond higher education, the corporate literature has documented the phenomenon of #SDG_washing, where organisations select the goals that flatter them, report enthusiastically on those, and stay silent on the goals where their performance would look poor (Heras-Saizarbitoria, Urbieta and Boiral, 2022). The parallel with a ranking that explicitly counts only a university's best three goals is direct and uncomfortable. Research on sustainability reporting tools has likewise examined the merits and limits of sustainability rankings as reporting instruments, finding that their value depends heavily on the rigour of verification behind them (Patara and Dhalla, 2022). Finally, econometric work has modelled the relationship between what universities disclose and where they finish, finding that disclosure itself is strongly associated with score, and that the highest placed institutions differ from others in geography, in the information they choose to disclose, and in their orientation toward industry partnership and knowledge transfer (De la Poza et al., 2021). This is perhaps the single most important empirical finding for the purposes of this article. #Disclosure_behaviour is not a side effect of performance. It is part of performance. 3. Method This is a conceptual and documentary study rather than an empirical one. It combines three sources of material. First, a structured review of peer reviewed literature published between 2020 and 2025 on the Impact Rankings, on global university rankings more generally, on the sustainable development goals in higher education, and on the entrepreneurial university. Sources were identified through academic databases using combinations of terms covering impact rankings, sustainability rankings, SDG reporting in universities, third mission and quantification in higher education. Priority was given to studies that examined the ranking directly, and to reviews that synthesised broader bodies of work. Second, documentary analysis of the published methodology documents and results announcements for the ranking across its editions, including the scoring rules, the participation requirements, the indicator structure for individual goals, and the announcements accompanying the 2026 transition to a ratings and membership model. These documents are the primary rulebook that institutional actors respond to, and they are treated here as the object of strategy rather than as neutral background. Third, published institutional accounts, including statements by university leaders and sector commentary reported alongside the rankings, which describe in the institutions' own words what they did and why. The analysis proceeds by reading the methodology as an incentive structure, asking at each scoring rule what a rational and reasonably ethical administrator would do in response, and then checking that inferred behaviour against what the empirical literature and the institutional accounts actually report. Where the literature confirms the inferred behaviour, it is presented as an observed practice. Where it does not, it is flagged as a hypothesis. This approach has clear limits, which are stated in Section 9. It cannot establish frequency, and it cannot separate sincere action from strategic action inside any given institution. What it can do is map the option space that educational entrepreneurs operate in. 4. The Anatomy of the Instrument Strategy is a response to rules. To understand the strategy, the rules must be set out plainly. 4.1 Origins and growth The exercise began in 2019, when a set of performance metrics developed the previous year was published as a league table at a summit held at the Korea Advanced Institute of Science and Technology. The first edition included 450 universities from 76 countries, a self selected group willing to submit data and be scrutinised. In 2020 the scope was extended to cover all seventeen goals. By the 2022 edition the overall table included 1,406 universities from 106 countries and territories, led by Western Sydney University. The seventh edition, published in June 2025, included more than 2,500 institutions from 130 countries and territories. In the space of six years the ranking went from a pilot to the largest voluntary data collection exercise in global higher education. The eighth edition, published in 2026, arrived under a new name. The Impact Rankings became the Sustainability Impact Ratings, and participation became membership based through a newly created Sustainability Impact Network, with tiered annual fees set according to World Bank country income classifications. Institutions in high income and upper middle income nations pay 3,500 pounds a year, those in lower middle income nations pay 1,000 pounds, and those in low income nations pay nothing. The 2026 edition covered 1,646 universities from 116 countries and territories, a substantial fall in participation from the previous year. The University of Manchester finished first overall, ending Western Sydney University's four year run at the top, with Griffith University second and Western Sydney third. The Philippines contributed more institutions than any other single country, with 160. The publisher stated that the underlying #methodology was not changed by the move to ratings, and explained the renaming on the grounds that the participating group is self selecting rather than comprehensive, which makes the word ratings more accurate than the word rankings. 4.2 The four pillars and the seventeen tables Performance on each goal is assessed across four broad areas. #Research covers the volume, quality and use of publications relevant to that goal. #Stewardship covers how the institution manages its own resources, including its estate, its staff and its students. #Outreach covers work with local, regional, national and international communities. #Teaching covers the preparation of graduates who can deliver on the goals and carry sustainability into their careers. Each goal has its own set of indicators drawn from these areas, with fixed weights inside the goal. Results are published as seventeen individual tables plus one overall table. Data comes from two sources. Bibliometric indicators are supplied by a commercial publisher's citation database, using field weighted measures and subject classifications. Everything else is submitted by the institution itself, in the form of numbers and supporting evidence, and assessed by the ranking provider, whose evaluation of the evidence is final. 4.3 The scoring rule that drives everything This is the rule that shapes institutional behaviour more than any other. To be included in the overall table, a university must submit data for #SDG_17, which covers partnerships, and for at least three other goals. The overall score is then calculated from SDG 17, which contributes twenty two per cent, plus the institution's three highest scoring other goals, each contributing twenty six per cent. Read that again slowly, because almost everything in Section 5 follows from it. A university's overall position depends on partnerships plus its own best three goals. It does not depend on its worst goals. It does not depend on the thirteen goals it chose not to enter. An institution that performs superbly on gender equality, quality education and good health, and abysmally on climate action, responsible consumption and life below water, will not be penalised in the overall table for the second list, provided it simply does not submit to those goals or provided they do not rank among its best three. This is the structural feature that critics have identified as allowing high overall placement with limited contribution across most of the agenda (Bautista-Puig et al., 2022). It is also, from the institution's point of view, an entirely legal and openly published rule that would be irrational to ignore. A second consequence follows. Because the ranking assesses submitted evidence, an institution that does excellent work but submits nothing scores zero on that dimension. Where a data point is not provided, a value of zero is entered. #Silence_is_scored. This converts the reporting function from an administrative afterthought into a core determinant of the result. 5. Seven Families of Optimisation Practice The practices below are grouped into families. Each family is described, its evidentiary basis noted, and its position on the line between substance and cosmetics assessed. Most institutions use several families at once, and the same practice can be substantive in one institution and hollow in another. That ambiguity is the point. 5.1 Portfolio selection: choosing which goals to fight on The first and most consequential decision is which of the seventeen goals to enter, and which to invest in. A naive institution enters everything. A sophisticated one treats the goal set as an #investment_portfolio. It assesses where its existing research strengths lie, where its existing community work is concentrated, where its region has visible needs, and where the indicator set is achievable given its data. It then concentrates effort on a small number of goals where it can plausibly reach the top band, while entering others at lower intensity to build the evidence base for future years. This behaviour is visible in the data. Universities steadily increase the number of goals they submit to as they gain experience in the ranking, a pattern more consistent with growing reporting capability than with the sudden creation of new sustainability programmes (Dickson, 2025). Studies of submission behaviour have found significant and hard to explain disparities in how institutions choose to participate, which is what one would expect if #portfolio_choice were being managed strategically rather than reported passively (Urbano et al., 2025). Is this gaming? Not necessarily. A small university in a water stressed region that concentrates on clean water, climate action and partnerships is behaving exactly as good strategy and good ethics both recommend. Nobody benefits from a mid sized teaching institution pretending to have a marine biology programme. Concentration of effort is what strategy means. The problem arises only when concentration becomes concealment, when a genuinely poor performance on a goal is quietly kept out of the submission in order to protect the overall score. The instrument cannot distinguish between the two, and that is a design weakness rather than an institutional failing. 5.2 Building the evidence infrastructure The second family is the least visible and the most transformative. It concerns the machinery of #evidence. The submission requires an institution to state, for example, how many first generation students it admitted, how much water it recycled, what proportion of its energy came from renewable sources, how many of its policies address the rights of specific groups, whether it tracks food waste, and how it engages with local government on planning. Most universities do not hold this information in one place. It lives in the estates department, the human resources system, the student records system, the research office, the finance system and the private files of individual academics. The entrepreneurial move is to build a permanent internal architecture that collects this material continuously rather than in a panic each autumn. In practice this means appointing an owner for each goal, usually a senior academic or professional staff member; establishing a data dictionary so that terms mean the same thing across departments; embedding SDG tagging in the research information system so that publications and projects are classified as they are created rather than retrospectively; adding sustainability fields to procurement, estates and human resources reporting; and creating a central evidence repository with version control and clear provenance. This is expensive, slow and unglamorous. It is also the single practice most likely to produce real change, because an institution that measures something continuously will eventually manage it. The literature on strategic integration of the goals makes the same point from the other direction: integration only works when it is written into strategy and monitored, rather than delegated to enthusiasts (Leal Filho et al., 2023). #Measurement_infrastructure is the bridge between rhetoric and behaviour. It is also the practice most sharply divided along lines of institutional wealth. A university that can fund a ranking office, a data analyst and a sustainability team will out report a university that cannot, even if the second university does more good in the world. One senior scholar quoted in the sector commentary around the 2026 results made this precise point, observing that simply having the administrative capacity to tell a good story is beyond the human resources of many institutions. #Reporting_capacity is itself unequally distributed, which means the ranking partly measures administrative wealth. 5.3 Documentation and narrative construction Where the previous family concerns numbers, this one concerns text. Many indicators are assessed through evidence rather than counting. The assessor asks whether the university has a policy, whether it publishes it, whether it applies it and whether it can show that it does. Points are awarded according to answers, and the ranking provider's judgement of the evidence is final. This converts institutional writing into a scoring activity. Educational entrepreneurs respond by professionalising documentation. Policies that existed informally are written down and published. Web pages that described activities in vague language are rewritten to state clearly what was done, for whom, when and with what result. Annual sustainability reports are produced in a format that maps directly onto the indicator set. Case studies are prepared with named partners, dated activities and measurable outcomes. There is an obvious risk here and it should be named. #Narrative_optimisation can substitute for action. A policy can be written, published and never applied. A case study can inflate a modest workshop into a regional transformation. The corporate sustainability literature calls the general phenomenon SDG washing, and it describes exactly this: selective, enthusiastic reporting on flattering goals accompanied by silence elsewhere (Heras-Saizarbitoria et al., 2022). Because the ranking assesses documents, an institution skilled at documents has an advantage independent of its skill at outcomes. The honest version of this practice is also real, though. Many universities genuinely did not have a policy on, say, workers' rights in their supply chain, and wrote one because the ranking asked. Many genuinely did not publish their emissions figures, and now do. The requirement to say something in public creates accountability that did not previously exist. #Forced_disclosure has consequences, and some of them are good. 5.4 The partnership engine and the SDG 17 problem SDG 17 is compulsory and carries twenty two per cent of the overall score, more than double the weight of any other individual goal in relative terms once one accounts for the fact that only four goals count. Every serious competitor must therefore build a partnership function. The goal is assessed through the university's relationships with government, its cross sectoral collaboration, its international research collaboration, its publication of data on the goals, and its educational work on sustainable development beyond its own campus. In practice this drives several behaviours. Institutions formalise relationships with municipal and national government that were previously informal. They join international networks and consortia. They co author with overseas colleagues, which conveniently also improves performance in conventional rankings. They publish open data on their own SDG performance, which is itself a scored activity. The strategic elegance of SDG 17 is that it is the most gameable goal and simultaneously the most genuinely valuable. Signing a memorandum of understanding costs nothing and produces a document. Building a working relationship with a city government that changes municipal policy costs a great deal and produces a document that looks almost the same. #Partnership_inflation is a predictable outcome. Empirical modelling of what distinguishes top placed institutions found that they are more oriented toward transferring knowledge to industry and toward supporting strong institutions in their countries, which suggests the goal does capture something real, but it also suggests that the institutions best placed to exploit it are those already embedded in dense partnership ecosystems (De la Poza et al., 2021). 5.5 Curriculum embedding The teaching pillar rewards institutions that produce graduates equipped to deliver on the goals. Optimisation here takes several forms. At the shallow end, existing modules are relabelled. A module on business ethics becomes a module on responsible business and the goals, with no change to the reading list. Course catalogues are audited and every course that mentions anything relevant is tagged. This is cheap, fast and produces a large number very quickly. At the deeper end, institutions redesign programmes. They introduce required sustainability content across all degrees, create interdisciplinary problem based courses in which students work on local community problems, build in service learning with assessed community placements, and establish graduate attributes that all programmes must demonstrate. Some create new degrees entirely. This is slow, contested inside faculties, and genuinely changes what students learn. Both approaches raise the reported number. Only one changes the graduate. Research on the implementation of the goals in university programmes has stressed the difference between adding sustainability as a topic and embedding it as a way of thinking, and has offered frameworks for doing the latter (Leal Filho, Salvia and Frankenberger, 2021). The wider survey evidence suggests that in most institutions integration still stems from individual academics and small groups rather than from the whole institution, which is precisely the condition under which #curriculum_tagging outruns #curriculum_change (Leal Filho, Salvia and Eustachio, 2023). For students reading this article, the practical test is straightforward. Look past the ranking position and ask a specific question of the admissions office: which required modules, in the degree I would take, include assessed sustainability content, and what exactly is assessed? A university that has embedded the work will answer immediately. A university that has tagged the catalogue will send you a brochure. 5.6 Campus operations and stewardship The stewardship pillar covers the university as a physical and social organisation: its energy, water, waste, procurement, employment practices and treatment of students and staff. This is the family where optimisation is hardest to fake, because the underlying quantities are physical. A university cannot claim to have cut water consumption by forty per cent without meters that say so. Consequently this is where the ranking has arguably produced the most solid change. Institutions have installed metering, moved to renewable supply contracts, introduced waste tracking, changed procurement rules, and published estates data that was previously internal. Two forms of optimisation still exist here. The first is #boundary_setting: choosing which parts of the institution count as the institution. Does the joint venture campus count? Does the leased building count? Does the contracted catering company count? Different answers produce different numbers, and the choice is often defensible either way. The second is baseline selection, where an institution reports improvement against a year that happened to be unusually bad. Neither is dishonest in a legal sense. Both illustrate the general truth that quantified comparison invites definitional manoeuvre. 5.7 Recognition management: branding, recruitment and reputation The final family is the one that connects the score back to the market, and it is the reason the whole enterprise pays for itself. A strong Impact Ranking position is a marketing asset with unusually broad reach. For universities that will never appear in the top 200 of a conventional world ranking, it is often the only global table on which they can credibly claim excellence. It is used in international student recruitment, in bids for research funding, in negotiations with government, in donor communications and in staff recruitment. The publisher's own survey of participants in May 2025 reported that eighty four per cent of respondents regarded the exercise as an important tool for showcasing their contribution to the goals, which is a candid statement of what the ranking is for from the participants' point of view. Educational entrepreneurs therefore invest in the recognition layer: press releases timed to publication, campaign pages, social media, badges on prospectuses, sector conference appearances, and case studies written for the ranking's own publications. #Recognition_management is not incidental to impact strategy. In many institutions it is the business case that funds the rest of the work, which means the substantive programmes in Sections 5.2 to 5.6 exist partly because the marketing value in Section 5.7 justified their budget. This is an uncomfortable but important observation. #Reputational_incentive is often the mechanism by which real sustainability work gets funded at all. 6. A Maturity Framework for Impact Optimisation The seven families above can be arranged into a developmental sequence. The framework below is offered as an analytical tool rather than a validated instrument. Its purpose is to give students, administrators and researchers a shared vocabulary for distinguishing what looks similar from the outside. Stage 1. Absent. The institution does not participate. It may do substantial community work, but none of it is visible externally. Data lives in individual memories. Nothing is published. Stage 2. Cosmetic. The institution participates. A small team assembles a submission close to the deadline, drawing on whatever evidence can be found. Existing activities are relabelled with goal numbers. Policies are written to satisfy indicators and then filed. The score exists; the institution has not changed. This stage produces #SDG_washing in its purest form and is detectable by the gap between what the website claims and what students and staff experience. Stage 3. Systematic. The institution builds evidence infrastructure. Goals get owners. Data is collected continuously. Submissions improve year on year because the machinery improves, not because activity improves. The score rises substantially at this stage, which is why the stage is so attractive and so dangerous. It is entirely possible for an institution to reach a strong global position without having materially altered its contribution to society. This is the stage most consistent with the observed pattern of institutions steadily increasing their submitted goals over time (Dickson, 2025). Stage 4. Integrated. Measurement begins to change management. Because emissions, gender pay gaps, community reach and graduate outcomes are now visible to leadership every quarter, they enter budget discussions and performance reviews. Deans are held to account for them. Curriculum change moves from tagging to redesign. The institution starts to do differently what it previously only described differently. The score may rise more slowly at this stage than at Stage 3, which is a critical and counterintuitive point: the transition from reporting well to acting well can be #ranking_neutral in the short term. Stage 5. Transformative. The goals become the organising logic of the institution's strategy rather than a reporting framework applied to it. Research priorities, hiring, estate investment, partnerships and curriculum are set with reference to a defined societal mission, usually anchored in the institution's region. The ranking becomes a by product rather than a target. Institutions at this stage are typically willing to report goals on which they perform badly, because their legitimacy no longer depends on the table. The willingness to publish an unflattering number is, in my view, the single most reliable external indicator that an institution has reached this stage. The framework carries one clear warning. Stages 3 and 4 look identical from outside, and the score does not distinguish them. Any external user of the ranking who wants to know whether an institution is at Stage 3 or Stage 4 must look at evidence the ranking does not collect: staff turnover in sustainability roles, whether academic promotion criteria reward engagement, whether the institution publishes bad results, and whether students experience the curriculum described in the submission. 7. Discussion: Recognition, Impact and the Space Between 7.1 The decoupling problem The central tension identified in this article can now be stated formally. The ranking rewards a joint product of #real_impact and #documented_impact. Because the second component can be improved far more cheaply and quickly than the first, and because the score cannot distinguish the two, institutions face a persistent incentive to over invest in documentation relative to action. Over time this produces #decoupling: the reported institution and the actual institution drift apart. The literature supports the plausibility of this mechanism without settling the question of how common it is. Modelling shows that disclosure is strongly associated with score (De la Poza et al., 2021). Content analysis shows that the scoring structure permits high placement with limited breadth of contribution (Bautista-Puig et al., 2022). Submission pattern analysis shows behaviour more consistent with improving reporting than with improving performance (Dickson, 2025). At the same time, careful empirical work using two years of data concludes that the ranking does provide a robust overview of institutional sustainability performance, which suggests that decoupling has not swamped the signal (Urbano et al., 2025). The reasonable conclusion is that the ranking contains real information and real noise in proportions that vary by institution, and that it should be used as one input rather than as a verdict. 7.2 The equity question and the 2026 turn The move in 2026 to a membership model, with tiered fees and participation limited to network members, sharpens a question that was already present. #Participation_cost has always been non financial and high: staff time, systems, coordination. The reforms add a direct financial cost for all but low income countries, while explicitly designing the fee structure so that wealthier institutions subsidise the poorest and low income nations pay nothing. The immediate effect on coverage was visible. Participation fell from more than 2,500 institutions across 130 countries in 2025 to 1,646 institutions across 116 countries in 2026. Some of that fall may reflect institutions deciding the exercise was not worth the cost, which is a healthy outcome. Some may reflect institutions that valued the exercise but could not fund it, which is not. At the same time, the 2026 data showed movement in the opposite direction on outcomes. High income countries held 34 of the top 50 places, down from 39 the previous year. The number of lower middle income institutions in the top 50 doubled to four, and the number of upper middle income institutions in the top 200 rose from 34 to 38. The Philippines contributed more institutions than any other country. Commentators described this as a gradual diffusion of capacity rather than a change in leadership, which sounds right. #Global_South institutions are getting better at this exercise, and the exercise remains one of the very few global tables on which they can compete at all. Both things are true, and the tension between them is not resolvable by better methodology. There is also a governance point worth stating plainly. When the organisation that publishes a ranking also sells membership of the network required to enter it, and also sells the analytics, benchmarking and conference access that help institutions improve within it, the incentives of the evaluator and the evaluated become entangled. This is not an accusation of bad faith. It is an observation that the responsible rankings literature, which has consistently called for transparency and independence in evaluation design, would regard the arrangement as requiring unusually strong safeguards (Gadd, 2021). 7.3 What the ranking cannot see Several important things fall outside the instrument by construction. It cannot see #outcomes as distinct from activities. A university can report the number of community programmes it runs, but not whether the community is better off. Genuine impact evaluation of the kind used in development economics is expensive and slow, and no global ranking could sustain it. It cannot see academic freedom, or the quality of intellectual life, or whether an institution's research on climate is being quietly funded by interests that oppose climate action. It cannot see whether the gender equality policy is enforced when a senior professor breaks it. It cannot see #opportunity_cost. The staff time spent assembling a submission is time not spent teaching or researching. For a well resourced institution this is a rounding error. For a small institution in a low income country it may be a serious diversion of scarce administrative capacity, undertaken because the ranking is the only route to global visibility. Finally, it cannot see the institutions that do excellent work and choose not to participate. The tables describe a self selected group, a fact the publisher itself acknowledged when it renamed the exercise. Any student or policymaker who treats the table as a complete map of good universities is making an error of interpretation that no amount of methodological improvement will fix. 7.4 Is optimisation legitimate? I want to answer this directly rather than hedge. Optimisation is legitimate when the reported picture is true. If an institution builds systems that let it accurately describe the good it does, and then describes it in the format the assessor requires, it has done nothing wrong and has probably done itself and its region a favour along the way. The evidence infrastructure it builds will outlast the ranking. Optimisation is illegitimate when it creates a false picture, whether by strategic silence about known poor performance, by inflating trivial activity, by writing policies it does not enforce, or by concentrating institutional energy on the appearance of impact at the expense of impact itself. The line is not always easy to locate in practice, but it exists, and people inside institutions generally know which side of it they are on. The honest conclusion is that the ranking makes the wrong thing easy and the right thing hard, and then pays both the same. That is a design problem, not a moral failing of the people who work within it. 8. Implications 8.1 For students Students are the group most often told to use these tables and least often told how they work. Five practical points follow from this analysis. First, check which goals the university actually submitted to, not just its overall position. The individual goal tables are published. An institution with a strong overall score and entries in only four goals is telling you something specific about itself. Second, remember that #non_participation is not evidence of poor performance. Excellent institutions choose not to enter for reasons ranging from cost to principle. Third, test claims about #curriculum. Ask which required modules contain assessed sustainability content in the specific degree you would take. Fourth, look for the unflattering number. An institution that publishes its emissions, its gender pay gap and its student attrition, including when they are bad, is showing you a different kind of institution than one that publishes only awards. Fifth, use the ranking for what it is genuinely good at: finding institutions whose declared mission points toward the problems you care about. It is a decent search tool and a poor verdict. 8.2 For institutional leaders and administrators For those doing this work, the analysis suggests several priorities. Build the evidence infrastructure first, and build it as a permanent management system rather than a submission project. If the data only exists in the autumn, it is not a management system. Choose a portfolio that is honest about the institution's region and strengths, and defend that choice publicly. A clear and well argued concentration is more credible than a scattered claim to universal excellence. Watch the Stage 3 trap. If the score is rising and nothing inside the institution has changed, the score is measuring the reporting office. Set internal targets on outcomes that the ranking does not reward, and hold deans to them. Publish what looks bad. It costs little in the table, because unsubmitted or weak goals mostly do not count against the overall score, and it buys the one thing money cannot buy, which is #credibility. Recognise the labour. The people assembling submissions are typically overworked professional staff whose contribution is invisible in academic promotion systems. If the institution's global reputation rests on their work, its reward systems should say so. 8.3 For policymakers and funders Ministries increasingly cite these positions in strategy documents and in funding decisions. Three cautions apply. The table measures a self selected group and rewards a joint product of action and documentation. It should not be used as a proxy for national sustainability performance in higher education. Funding schemes that reward ranking position transmit the design flaws of the ranking directly into public policy, and will produce more reporting rather than more impact. Funding that rewards verified outcomes, even crude ones, does better. If a system wants its universities to serve society, the cheapest intervention is usually to change academic promotion criteria so that engagement counts. The literature has been saying this about the third mission for years (Compagnucci and Spigarelli, 2020; Perkmann et al., 2021), and no ranking will substitute for it. 9. Limitations This study is conceptual and documentary. It cannot establish how frequently each practice occurs, and it cannot measure the size of the decoupling effect it describes. The maturity framework in Section 6 is an analytical proposal, not a validated instrument, and it has not been tested against institutional data. The reliance on published methodology documents means the analysis reflects the rules as stated rather than as applied, and the ranking provider's internal assessment practices are not publicly observable. The institutional accounts drawn on are, by their nature, produced by parties with an interest in appearing successful. Finally, the analysis is written from outside the institutions it describes. A programme of interview based research with ranking officers, sustainability directors and senior leaders would test, and very likely complicate, much of what is argued here. That research is needed and does not yet exist in any substantial form. 10. Conclusion The Impact Rankings, now the Sustainability Impact Ratings, did something that no other global table had done. They told a large number of universities that the work they do for their communities counts, and they gave that work a number. For institutions long excluded from global visibility by rankings built around research volume and inherited prestige, this was not a small thing. It opened a route to #global_recognition that had been closed. It also created a market, and markets create specialists. The #educational_entrepreneurs described in this article are the specialists. They read the scoring rules, choose a portfolio, build the evidence machinery, write the narrative, cultivate the partnerships, redesign the curriculum, retrofit the campus and sell the result. Most of them believe in the work. Many of them use the ranking as a lever to get sustainability funded internally, which is a defensible and often admirable use of a flawed instrument. The difficulty is structural and will not be fixed by tinkering with weights. The ranking pays equally for doing good and for being able to describe doing good, and describing is cheaper. Until an instrument exists that can tell those apart, the number will contain both signal and noise, and the people who read it will need to do the work of separating them. For students, that means reading the tables with curiosity rather than deference, and asking the specific questions in Section 8.1. For institutions, it means resisting the temptation to stop at Stage 3, where the score is high and nothing has changed. For the sector, it means remembering what the ranking was invented to encourage. As one university leader put it when his institution reached the top of the 2026 ratings, the exercise asks institutions to consider what universities are good for, and not merely what they are good at. That is a question worth taking seriously, and it is not one that any table can answer on a university's behalf. Hashtags #Impact_Rankings #Sustainability_Impact_Ratings #Times_Higher_Education #Sustainable_Development_Goals #educational_entrepreneurship #university_rankings #higher_education_strategy #societal_impact #SDG_washing #global_recognition #university_leadership #sustainability_in_higher_education #ranking_optimization #academic_integrity #students_guide_to_rankings References Bautista-Puig, N., Orduna-Malea, E. and Perez-Esparrells, C. (2022) Enhancing sustainable development goals or promoting universities? An analysis of the Times Higher Education impact rankings. International Journal of Sustainability in Higher Education, 23(8), 211-231. https://doi.org/10.1108/IJSHE-07-2021-0309 Bellantuono, L., Lo Sasso, A., Amoroso, N., Monaco, A., Tangaro, S. and Bellotti, R. (2025) Network assortativity for a multidimensional evaluation of socio-economic territorial biases in university rankings. PLOS ONE, 20(6), e0323356. https://doi.org/10.1371/journal.pone.0323356 Caputo, F., Ligorio, L. and Pizzi, S. (2021) The contribution of higher education institutions to the SDGs: An evaluation of sustainability reporting practices. Administrative Sciences, 11(3), 97. https://doi.org/10.3390/admsci11030097 Chaleta, E., Saraiva, M., Leal, F., Fialho, I. and Borralho, A. (2021) Higher education and sustainable development goals (SDG): Potential contribution of the undergraduate courses of the School of Social Sciences of the University of Evora. Sustainability, 13(4), 1828. https://doi.org/10.3390/su13041828 Chankseliani, M. and McCowan, T. (2021) Higher education and the Sustainable Development Goals. Higher Education, 81(1), 1-8. https://doi.org/10.1007/s10734-020-00652-w Chun, H. and Sauder, M. (2022) The logic of quantification: Institutionalizing numerical thinking. Theory and Society, 51, 335-370. https://doi.org/10.1007/s11186-021-09453-1 Compagnucci, L. and Spigarelli, F. (2020) The third mission of the university: A systematic literature review on potentials and constraints. Technological Forecasting and Social Change, 161, 120284. https://doi.org/10.1016/j.techfore.2020.120284 De la Poza, E., Merello, P., Barbera, A. and Celani, A. (2021) Universities reporting on SDGs: Using THE Impact Rankings to model and measure their contribution to sustainability. Sustainability, 13(4), 2038. https://doi.org/10.3390/su13042038 Dickson (2025) Meaningful measurement of sustainable development or gaming the system? Patterns in submission to Times Higher Education's Impact Ranking. Sustainable Development. https://doi.org/10.1002/sd.70419 Gadd, E. (2020) University rankings need a rethink. Nature, 587, 523. Gadd, E. (2021) Mis-measuring our universities: Why global university rankings do not add up. Frontiers in Research Metrics and Analytics, 6, 680023. https://doi.org/10.3389/frma.2021.680023 Hazelkorn, E. (2023) The power in managing numbers: Changing interdependencies and the rise of ranking expertise. Higher Education, 86(4), 771-790. Heras-Saizarbitoria, I., Urbieta, L. and Boiral, O. (2022) Organizations engagement with sustainable development goals: From cherry-picking to SDG-washing? Corporate Social Responsibility and Environmental Management, 29(2), 316-328. https://doi.org/10.1002/csr.2202 Leal Filho, W., Salvia, A. L. and Eustachio, J. H. P. P. (2023) An overview of the engagement of higher education institutions in the implementation of the UN Sustainable Development Goals. Journal of Cleaner Production, 386, 135694. https://doi.org/10.1016/j.jclepro.2022.135694 Leal Filho, W., Salvia, A. L. and Frankenberger, F. (2021) A framework for the implementation of the Sustainable Development Goals in university programmes. Journal of Cleaner Production, 299, 126915. Leal Filho, W., Simaens, A., Paco, A., Hernandez-Diaz, P. M., Vasconcelos, C. R. P., Fritzen, B. and Mac-Lean, C. (2023) Integrating the Sustainable Development Goals into the strategy of higher education institutions. International Journal of Sustainable Development and World Ecology. https://doi.org/10.1080/13504509.2023.2167884 Leal Filho, W., Trevisan, L. V., Rampasso, I. S., Anholon, R., Dinis, M. A. P., Brandli, L. L., Sierra, J., Salvia, A. L., Pretorius, R., Nicolau, M., Eustachio, J. H. P. P. and Mazutti, J. (2023) When the alarm bells ring: Why the UN Sustainable Development Goals may not be achieved by 2030. Journal of Cleaner Production, 407, 137108. https://doi.org/10.1016/j.jclepro.2023.137108 Patara, S. and Dhalla, R. (2022) Sustainability reporting tools: Examining the merits of sustainability rankings. Journal of Cleaner Production, 366, 132960. Perkmann, M., Salandra, R., Tartari, V., McKelvey, M. and Hughes, A. (2021) Academic engagement: A review of the literature 2011-2019. Research Policy, 50(1), 104114. https://doi.org/10.1016/j.respol.2020.104114 Urbano, V. M., Arena, M., Azzone, G. and Mayeres, M. (2025) Sustainable development in higher education: An in-depth analysis of Times Higher Education Impact Rankings. Journal of Cleaner Production, 501, 145302. https://doi.org/10.1016/j.jclepro.2025.145302

Cities such as Dubai, Doha, Riyadh, Singapore, and Toronto have become magnets for workers from many countries. In these places, a single team may contain people from ten or more nationalities, and a large share of staff may leave within two or three years, producing #turnover rates far above the global average. This article calls such places high-mobility hubs and asks a simple question: what does leadership have to do differently when both the culture mix and the turnover rate are unusually high at the same time? Drawing on an integrative review of recent scholarship in international human resource management, cross-cultural leadership, and global talent management, the article develops a conceptual framework built on four leadership capabilities: cultural agility, structural fairness, communication design, and knowledge continuity. The review shows that most existing models were built for settled multinational subsidiaries, where leaders had time to build relationships and where staff turnover was moderate. High-mobility hubs break these assumptions. Speed of trust formation, transparency of reward systems, deliberate design of language norms, and the capture of knowledge before people leave become central rather than secondary concerns. The article also examines how workforce nationalisation policies in the Gulf Cooperation Council create a dual talent pool that leaders must manage without creating resentment on either side. Twelve propositions are offered for future testing, together with practical guidance for managers and a research agenda. The article is written for students and early-career professionals who will spend part of their working lives in such environments. Keywords: cross-cultural leadership; talent management; workforce mobility; cultural intelligence; inclusive leadership; emerging markets; international human resource management 1. Introduction A team meeting in a business district in the Gulf may include an Indian software engineer, a Filipino accountant, an Egyptian marketing manager, a British operations director, an Emirati graduate on a national development programme, and a Chinese supplier liaison joining by video call. Each person brings a different set of expectations about how decisions get made, how disagreement is voiced, how much closeness is normal between a manager and a subordinate, and how long they expect to stay in the country. Two years later, half of them may have moved on. This is the ordinary reality of #cross_cultural_leadership in cities that this article calls high-mobility hubs. The term describes urban labour markets that combine three features. First, a very high proportion of foreign-born workers, sometimes a clear majority of the workforce. Second, high rates of entry and exit driven by #visa_policy and short contracts, so that the composition of the workforce is always changing. Third, rapid growth in sectors that did not exist locally a decade earlier, which forces organisations to hire quickly and at scale. The Gulf states are the clearest example, but the pattern also appears in Singapore, in parts of Southeast Asia, in fast-growing African commercial centres, and in immigration-heavy cities in Canada and Australia. #high_mobility_hubs are not simply places with diverse workforces. They are places where diversity is combined with churn. The distinction matters because the leadership literature has largely studied diversity without churn. Research on multicultural teams has produced a rich understanding of how cultural differences create both process gains and process losses, and has shown that the effect of diversity on performance is indirect, working through team processes and shaped by context (Stahl & Maznevski, 2021). Research on cross-cultural leadership has mapped how societal values shape what followers expect from leaders and which behaviours count as effective (Den Hartog & De Hoogh, 2024). Research on #global_talent_management has traced how multinational enterprises identify, deploy, and retain people in strategic roles across borders (Caligiuri, Collings, De Cieri, & Lazarova, 2024). Each stream is valuable. None of them was designed for a setting in which the average tenure of a professional employee is short, where the legal right to remain in the country is tied to the job, and where the national government is actively reshaping the composition of the workforce through policy. This creates a practical problem. A manager in a stable subsidiary in Germany or Japan can invest in relationships over years, allow trust to develop slowly, and rely on shared organisational memory. A manager in a high-mobility hub cannot. The people who understand why a process works the way it does may be gone before the next audit. The team that finally learned to work together may lose three members in a single quarter. Norms that a leader carefully built must be rebuilt for every new arrival. The leadership task is therefore not only to bridge cultural differences but to bridge them repeatedly and quickly, under conditions where the reward for building a strong relationship may be lost within a year. There is also an institutional layer. Several hub economies have introduced #workforce_nationalization programmes designed to raise the share of citizens in private-sector employment. Emiratisation in the United Arab Emirates, Saudisation in Saudi Arabia, Qatarisation in Qatar, and similar policies elsewhere use quotas, incentives, subsidies, and penalties to change hiring behaviour (Elbanna, 2022). These policies are legitimate responses to a real structural imbalance, and they are central to national #economic_diversification strategies. But they place line managers in a difficult position. Managers must meet quantitative targets for hiring citizens while also retaining the experienced foreign staff on whom current operations depend, and they must do so without allowing either group to feel that the other is being favoured. The literature on nationalisation has focused mainly on policy design and macro outcomes. It has said much less about what the policy feels like inside a team, and what leadership behaviours make the policy work rather than backfire (Elsharnouby, Elbanna, Obeidat, & Mauji, 2024). This article addresses that gap. Its purpose is threefold. First, to define the high-mobility hub as a distinct organisational context and to specify what makes it different from the settings in which established leadership theory was developed. Second, to synthesise recent research into an integrated conceptual framework of the leadership capabilities that this context demands. Third, to translate that framework into propositions that researchers can test and into practices that managers and students can use. The article proceeds as follows. Section 2 reviews the relevant literature across five streams. Section 3 develops the conceptual framework and states twelve propositions. Section 4 describes the review method. Section 5 presents seven themes that emerge from the synthesis. Section 6 discusses theoretical contributions. Section 7 sets out practical implications for managers and human resource professionals. Section 8 addresses students and early-career professionals directly. Section 9 states limitations and a research agenda. Section 10 concludes. 2. Conceptual Background and Literature Review 2.1 Defining the high-mobility hub The idea of a global city is familiar in international business research, where such cities are understood as places with dense international connections, strong professional service ecosystems, and a cosmopolitan population. The high-mobility hub shares these features but adds two more that are directly relevant to leadership. The first is the temporary residence assumption. In many hub economies, the right of a foreign professional to live in the country depends on holding a job there. Residence is renewable rather than permanent. Even where long-term visa routes exist, most workers plan their lives around a horizon of a few years rather than a career. This shapes behaviour in ways that leaders often misread. An employee who declines a long development programme may not lack ambition. She may simply be uncertain whether she will still be in the country when it finishes. An employee who negotiates hard on salary rather than on career progression may be acting rationally, because salary is portable and a promotion inside one country may not be. The second is compositional volatility. In a stable workforce, the leader manages a known group. In a hub, the group is a moving population. Research on global work has emphasised that the where, who, and how of international employment have all become less predictable, with traditional long-term assignments giving way to a wider mix of short assignments, commuter arrangements, remote work, and locally hired international staff (Lazarova, Caligiuri, Collings, & De Cieri, 2023). In hub economies this mix is not an exception. It is the norm. Together these two features mean that the unit of leadership is not a stable team but a stream of people passing through a role. The organisation persists; the individuals rotate. #labour_mobility becomes the background condition of everything a leader does. 2.2 Cross-cultural leadership: from dimensions to behaviour Early cross-cultural management research was dominated by value dimensions, which described national cultures along axes such as individualism, #power_distance, and uncertainty avoidance. These frameworks were enormously influential and remain the vocabulary that most practitioners use. They have also been criticised, and the criticism has grown sharper. Comparative testing of the instruments used to measure these value models has shown weak convergence between them, which raises questions about whether they are measuring what they claim to measure (Taras, Steel, & Stackhouse, 2023). Treating a nation as a single culture ignores internal variation, and using a national average to predict an individual's behaviour is a well-known error. Recent reviews have moved the field toward a more behavioural and contextual position. Culture is now more often treated as something that shapes the expectations followers hold about leaders, rather than as something that determines what leaders do (Den Hartog & De Hoogh, 2024). The same behaviour can carry different meanings in different settings. Direct feedback in front of others may be read as honest and efficient in one context and as a threat to #face_saving in another, which is why a shared #feedback_culture cannot be assumed. Silence in a meeting may signal agreement, disagreement, deference, or simply the difficulty of thinking aloud in a second language. What matters for leadership is not knowing an abstract score for a country but reading the meaning that a specific group of people attaches to a specific act. This shift has a direct consequence for #high_mobility settings. When a leader faces a team of twelve nationalities, a dimension-based approach becomes unusable. There is no single cultural profile to adapt to. #ethnocentrism, the assumption that one's own way is the natural way, is dangerous, but so is the opposite error of trying to memorise a separate rulebook for each nationality. What is needed is a form of #contextual_intelligence: a general capability to read situations, hold multiple interpretations open, and adjust behaviour without abandoning consistency. 2.3 Cultural intelligence, cultural agility, and global mindset Three closely related constructs address that capability. #cultural_intelligence refers to an individual's capability to function effectively in culturally diverse situations, usually described as having motivational, cognitive, metacognitive, and behavioural components. #global_mindset refers to an orientation that combines openness to difference with the cognitive ability to integrate diverse perspectives. Systematic review work has shown that these constructs overlap substantially and are often measured in overlapping ways, which has slowed theoretical progress even as it has confirmed their practical relevance (Yari, Lankut, Alon, & Richter, 2020). #cultural_agility offers a more action-focused alternative. It is defined as the ability to vary one's response to context, and it explicitly recognises that adaptation is not the only correct answer. A culturally agile professional chooses between three responses: adapting to the local way of doing things, adopting a way that is different from their own but not local, or integrating and combining approaches, and knowing when to hold firm rather than adapt (Caligiuri, 2021). This last point matters. A leader in a hub is required to apply company-wide standards on safety, ethics, and equal treatment even where local practice differs. Cultural agility does not mean cultural surrender. Research on how such competencies develop suggests that international experience alone does not produce them. What produces them is experience that is demanding, that requires real interaction rather than observation, and that is accompanied by reflection (Caligiuri & Caprar, 2022; Caligiuri, 2023). For high-mobility hubs, the practical implication is that the leader's own #adaptive_leadership capability is the primary instrument. There is no time to learn each culture. There is only time to learn how to learn quickly. 2.4 Multicultural teams: gains, losses, and faultlines The evidence on culturally diverse teams is more nuanced than either optimists or pessimists suggest. Diversity does not directly cause better or worse performance. It works through #team_processes. Diversity increases the range of information and perspectives available, which can improve #creativity, support #innovation, and raise the quality of #decision_making. It also increases the risk of misunderstanding, raises the demand for #conflict_management, reduces social integration, and can slow convergence on a decision. Whether gains or losses dominate depends on context, task, and, critically, on how the team is led (Stahl & Maznevski, 2021). One mechanism deserves particular attention in hubs. #faultlines are hypothetical dividing lines that split a team into #subgroups when several attributes align. A team with three South Asian junior engineers and three Western senior managers has a strong faultline, because nationality, seniority, and possibly pay all line up in the same direction. A team with the same nationalities spread across seniority levels has a weak one. Faultlines are more damaging than diversity itself, because they turn difference into an us-and-them structure. Hub organisations are unusually prone to strong faultlines because recruitment channels, visa categories, and pay bands often correlate with nationality. When the whole warehouse team comes from one country and the whole finance team from another, the organisation has built a faultline into its structure without ever deciding to. Work on multinational teams has also shown that gaps between members can be actively bridged. Individuals who span cultural boundaries, translating not only language but norms and expectations, improve team functioning, and their contribution is often invisible to formal performance systems (Backmann, Kanitz, Tian, Hoffmann, & Hoegl, 2020). Identifying and supporting such bridge-builders is a leadership act, not an accident. 2.5 Inclusive leadership, psychological safety, and voice If cultural agility is the leader's internal capability, #inclusive_leadership is its external expression. A systematic review of the field defines inclusive leadership around two needs that must be satisfied at once: the need for #belonging, meaning acceptance as a full member of the group, and the need for #uniqueness, meaning recognition of what makes a person distinct rather than pressure to assimilate (Korkmaz, van Engen, Knappert, & Schalk, 2022). Satisfying only one produces failure. Belonging without uniqueness is assimilation, in which people are accepted only if they behave like the dominant group. Uniqueness without belonging is tokenism, in which people are visible but not influential. The mechanism through which inclusive leadership works is now reasonably well established. Inclusive behaviours raise #psychological_safety, the shared belief that speaking up will not lead to punishment or humiliation, and psychological safety in turn increases #employee_voice, the willingness to raise problems and ideas (Jiang, Ding, Wang, & Li, 2022; Mohase, Donald, & Israel, 2025). Voice is the input on which learning, error correction, and innovation depend. In a high-mobility hub, the barriers to voice are stacked. A person may be speaking a second language, may come from a context where challenging a senior person is inappropriate, may hold a visa tied to the employer, and may have been in the organisation for only weeks. Any one of these would suppress voice. Together they make silence the rational default. A leader who interprets that silence as agreement is making a serious error. Related work stresses that inclusion is not achieved by good intentions but by leader character and by deliberate design of the systems in which people work (Seijts & Milani, 2022), and that leaders need an active mindset about diversity, seeing it as a resource that must be worked with rather than a fact to be tolerated (van Knippenberg & van Ginkel, 2022). 2.6 Global talent management in unstable conditions Talent management scholarship has moved from a focus on identifying high-potential individuals toward a broader concern with how talent systems fit their context. Reviews stress that context matters, that practices which work in one national and institutional setting may fail in another, and that the transfer of talent practices from headquarters to subsidiaries is a persistent source of friction (Gallardo-Gallardo, Thunnissen, & Scullion, 2020). Even the meaning of the word talent varies across organisations and languages, which has consequences for how systems are designed and how fair they appear to employees (Vardi & Collings, 2023). Recent critical reviews argue that global talent management has been slow to adapt to the new organisational reality: more flexible mobility, more remote work, greater attention to employee wellbeing, and a labour market in which valuable people have more options than before (Caligiuri et al., 2024). Evidence that multinationals are managing talent effectively on a global scale is thin. In emerging markets specifically, the fit between imported talent practices and local societal expectations is often poor, and firm-level practices must be evaluated against societal and individual outcomes rather than only against corporate templates (Pereira, Collings, Wood, & Mellahi, 2022). Two categories of people matter especially in hubs and are under-studied. #self_initiated_expatriates are professionals who move abroad on their own initiative rather than being sent by an employer. They form a large share of the professional workforce in Gulf cities, yet talent systems are often designed around the assumption of a company-assigned expatriate with a defined assignment length and a repatriation plan. #host_country_nationals, the citizens of the country in which the operation sits, have similarly received less research attention than their importance warrants, even though they are the group most affected by nationalisation policy and the group whose long-term commitment the organisation most needs (Michailova, Fee, & DeNisi, 2023). 2.7 Institutional context: nationalisation and the dual labour market Gulf labour markets are structurally unusual. Foreign nationals form the overwhelming majority of the private-sector workforce, while citizens have historically been concentrated in public employment. Governments have responded with nationalisation programmes that combine quotas, wage support, training subsidies, and penalties for non-compliance. The #institutional_context is therefore an active force rather than a backdrop. A thematic review of this literature identifies four levels at which such policies operate, legal, organisational, human development, and socio-cultural, and argues that programmes fail when they are pursued at only one level, for example by imposing quotas without addressing the skills pipeline or the perceptions that employers and citizens hold of each other (Elbanna, 2022). The perception problem is well documented. Qualitative work with local job seekers finds that the private sector is often seen as offering longer hours, weaker job security, and less prestige than public employment, while employers hold stereotypes about the commitment of citizen employees. These beliefs are self-reinforcing, and they shape the #diversity_climate of every workplace in which the two groups meet (Elsharnouby et al., 2024). Meanwhile, nationalisation reshapes the position of migrant workers, affecting job security, remittances, and the sense of belonging of communities that have lived in these countries for decades (Sahu, 2025). For leadership, the result is a #dual_labour_market inside the organisation itself. Two groups, one protected and promoted by policy, the other holding most of the current operational knowledge, work side by side under different implicit contracts. Every promotion decision, every training allocation, and every performance conversation is read through this lens. #localisation is therefore not only an HR compliance task. It is a daily leadership challenge in fairness and communication. 2.8 Virtual and hybrid dispersion Finally, hub organisations are rarely all in one room. Regional headquarters coordinate teams across several countries, and #hybrid_work has become normal for professional roles. Leading in this mode has its own literature. #trust in dispersed teams must form quickly, on the basis of role and reputation rather than long acquaintance, a process described as #swift_trust (Zakaria & Mohd Yusof, 2020). #virtual_teams and hybrid arrangements introduce #proximity_bias, the tendency to favour those who are physically present, which can quietly reproduce cultural hierarchies if the people in the room are systematically from one group and the people on the screen from another. Reviews of hybrid leadership stress that inclusion in this mode must be designed, through explicit norms for meetings, deliberate rotation of inconvenient #time_zones, and attention to who is being heard (da Silva, Castello-Sirvent, & Canos-Daros, 2022; Detjen & Webber, 2023; Ding, Ren, & Lin, 2024). 2.9 The gap Each stream contributes something. None of them, alone, describes the hub condition. The cross-cultural literature explains difference but assumes relative stability. The talent literature explains systems but often assumes a settled institutional environment. The nationalisation literature explains policy but stops at the office door. The virtual teams literature explains distance but not high turnover. The gap is at the intersection: leadership under simultaneous cultural heterogeneity, high churn, and institutional pressure to change workforce composition. That intersection is where this article works. 3. Conceptual Framework and Propositions 3.1 The four-capability framework The synthesis suggests that effective leadership in high-mobility hubs rests on four capabilities. They are not personality traits. They are practices that can be designed, taught, and audited. Capability 1: Cultural agility. The leader's ability to read a situation and vary the response, choosing consciously between adapting, adopting, integrating, and holding firm. In hubs this capability must operate at speed, because the leader will meet more novel cultural situations in a month than a leader in a homogeneous setting meets in a year. Cultural agility also includes the discipline of not over-attributing behaviour to culture. Many problems that look cultural are actually structural: a person who does not respond to emails at night is not necessarily showing a cultural difference in commitment, but may simply be enforcing a reasonable boundary. Capability 2: Structural fairness. The leader's ability to make the rules of the game visible and defensible. Where nationality correlates with pay, visa status, and promotion prospects, perceptions of #procedural_justice and #distributive_justice become the single largest driver of engagement and voice. Fairness in this context is not primarily about equal outcomes, which local law and market structure may not permit. It is about consistent, explained, non-arbitrary process. Capability 3: Communication design. The leader's ability to build explicit norms for how the team talks, decides, disagrees, and records. In a mono-cultural, low-turnover team, communication norms are absorbed tacitly. In a hub, tacit absorption fails, because there is no shared tacit base and no time for slow socialisation. Norms must be made explicit and re-taught with each intake. Capability 4: Knowledge continuity. The leader's ability to ensure that what the team knows survives the departure of the people who know it. This includes documentation, deliberate overlap between leavers and joiners, distributed rather than concentrated expertise, and cultures in which asking a question is not treated as evidence of incompetence. 3.2 Propositions Cultural agility: P1. In high-mobility hubs, leader cultural agility is more strongly associated with team performance than leader knowledge of specific national cultures. P2. The relationship between leader cultural agility and team performance is stronger where team cultural heterogeneity is high and team tenure is short, because the demand for rapid situational reading rises with both. P3. International experience improves leader cultural agility only when the experience involved sustained interaction with, and dependence on, culturally different colleagues, and was followed by structured reflection. Structural fairness: P4. Perceived procedural justice mediates the relationship between visible differentiation in reward or status and employee outcomes such as voice, engagement, and intention to stay. P5. Strong demographic faultlines that align nationality with hierarchical level weaken team performance, and this effect is reduced when leaders deliberately create cross-cutting roles, project teams, and mentoring pairs that break the alignment. P6. Where nationalisation targets are communicated as quotas without explanation, they increase perceived unfairness among both citizen and non-citizen employees; where they are communicated as part of a transparent capability-building strategy with clear criteria, the negative effect is significantly reduced. Communication design: P7. Explicit, jointly agreed communication norms increase psychological safety in culturally heterogeneous teams more than general leader friendliness or approachability. P8. In teams where the working language is a second language for most members, deliberate practices that reduce the fluency penalty, such as circulating agendas in advance, permitting written contributions, and slowing turn-taking, increase the participation of non-fluent members without lowering decision quality. P9. Psychological safety mediates the relationship between inclusive leadership and employee voice more strongly in high-mobility settings than in low-mobility settings, because the baseline barriers to voice are higher. Knowledge continuity: P10. Organisational performance in high-turnover environments depends less on the retention rate itself than on the effectiveness of knowledge capture and transfer at the point of exit. P11. Teams with distributed expertise, in which critical knowledge is held by at least two people, absorb departures with a smaller performance loss than teams with concentrated expertise, and this difference grows as turnover rises. P12. Structured, culturally sensitive #onboarding that makes implicit norms explicit shortens the time to productivity for new arrivals and reduces early voluntary exit. These propositions are stated to be testable. Several could be examined with existing survey instruments; others would require longitudinal or multi-level designs. 4. Method This article is an integrative conceptual review. It does not report new empirical data. Its aim is to synthesise dispersed literatures into a framework and a set of testable propositions, which is an accepted contribution type in management journals when a field is fragmented across sub-disciplines. The review followed four steps. First, five literature streams were identified as relevant to the research question: cross-cultural leadership; multicultural team dynamics; cultural intelligence and cultural agility; global talent management and international human resource management; and workforce nationalisation in the Gulf Cooperation Council. A sixth stream, virtual and hybrid team leadership, was added during the review because it recurred consistently in recent work on global talent. Second, within each stream, priority was given to review articles, meta-analyses, and conceptual papers published in the last five years, because these consolidate earlier findings and identify open questions. Older foundational work is acknowledged where it remains the reference point for current debate, but the evidence base for the argument rests on recent sources. Third, the material was coded thematically. Rather than organising by discipline, the analysis asked what each stream implies for a leader operating under conditions of simultaneous cultural heterogeneity, high turnover, and institutional pressure. Themes were retained when they appeared in at least two independent streams, which produced the seven themes reported in Section 5. Fourth, the themes were organised into the four-capability framework described above, and propositions were derived from the points at which the literature makes a claim that has not yet been tested in the hub context. Three limitations of the method should be stated at the outset. The review is selective rather than exhaustive; it is not a systematic review conducted under a formal protocol with reported search strings and inclusion counts. The empirical base for the Gulf context is thinner than for Western contexts, and some of it is qualitative and small-scale, so conclusions about that setting are correspondingly tentative. And the framework, being conceptual, awaits testing. 5. Analysis: Seven Themes 5.1 Theme one: churn changes the mathematics of relationship building Leadership models assume that a leader and a follower will have time. Transformational leadership assumes that a leader can develop followers over a period long enough for development to occur. Leader-member exchange assumes that a relationship deepens through repeated interaction. In a hub where the median professional stays two to three years, and where the first six months are spent adjusting to a new country, the effective window for a developmental relationship may be eighteen months. This does not make relationship building pointless. It changes its economics. Three adjustments follow. First, the front end matters disproportionately. If the relationship has eighteen useful months, the first six weeks are worth far more than they would be in a five-year relationship. Onboarding stops being an administrative task and becomes the highest-leverage leadership activity available. Yet in most hub organisations onboarding is treated as paperwork, delegated to human resources, and completed in a day. Second, trust must be front-loaded rather than accumulated. Work on dispersed teams describes swift trust, in which people extend provisional trust based on role, reputation, and clear structure, and then confirm or withdraw it based on early behaviour (Zakaria & Mohd Yusof, 2020). Leaders can support this by being explicit about competence and reliability early: stating clearly who is responsible for what, meeting small commitments visibly, and closing loops quickly. Warmth helps but is not sufficient. In cross-cultural settings, warmth signals are read differently, whereas reliability is read similarly almost everywhere. Third, exit must be normalised. When leaving is treated as betrayal, people conceal their plans, leave abruptly, and take their knowledge with them. When leaving is treated as normal and departures are handled gracefully, people give notice earlier, hand over properly, and often become sources of business and referrals afterwards. The leader who reacts badly to resignations is, in effect, training the whole team to leave badly. 5.2 Theme two: language is a structure of power, not just a medium In most hub organisations, English is the working language. Very few of the people using it grew up speaking it. The consequence is a quiet but powerful distortion: fluency is systematically mistaken for competence. The mechanism is easy to observe in any meeting. The person who speaks the corporate #lingua_franca most fluently gets to shape the frame of the discussion, interrupts more successfully, is quoted more often, and is remembered as having contributed more. The person who is thinking hard in a second language speaks later, more slowly, and less often, and is remembered as quiet. Neither outcome has anything to do with the quality of their ideas. Over time, promotion decisions accumulate this bias. Leaders can counteract this without lowering standards. Circulating an agenda in advance allows people to prepare. Allowing contributions in writing, before or after the meeting, gives slower speakers a channel. #language_diversity therefore has to be managed openly. Explicitly slowing turn-taking, and asking direct questions to named individuals rather than throwing questions open to the room, redistributes airtime. Summarising and checking understanding, rather than assuming it, catches errors early. #code_switching, where two colleagues briefly move to a shared first language to clarify a technical point, should be treated as a tool rather than a discourtesy, provided the conclusion is brought back into the shared language. Leaders should also be careful about idiom. Sporting metaphors, cultural references, humour, and irony are the parts of language that travel worst. A leader who says a project is "on the back burner" may be understood to have cancelled it. Plain #communication_norms are not a lowering of sophistication. In a multilingual team they are a precondition of accuracy. 5.3 Theme three: faultlines are built by systems, not by people Teams in hubs rarely split along cultural lines because individuals are prejudiced. They split because organisational systems align attributes. #recruitment agencies specialise in particular countries. Visa categories map to job grades. #wage_stratification is built in when pay bands are benchmarked against the home country of the employee rather than the job. Housing allowances, schooling support, and flight benefits differ by contract type. The result is that nationality, seniority, pay, and benefit package can end up correlated, which is the textbook definition of a strong faultline. Once the faultline exists, ordinary human behaviour does the rest. People form in-groups, attribute mistakes by out-group members to character and mistakes by in-group members to circumstance, and interpret ambiguous acts in ways that confirm existing beliefs. Communication across the line becomes formal and thin. Information stops flowing. The leadership response is structural rather than attitudinal. Diversity training that asks people to be more tolerant has limited effect when the structure keeps generating the division. What works is cross-cutting: forming project teams that deliberately mix the groups, rotating who represents the team to senior management, creating #mentoring and #reverse_mentoring pairs across the line, and reviewing whether any function has become mono-national by accident. Leaders should audit their own team composition by asking a simple question: if I drew my team on paper and coloured in nationality, seniority, and pay band, would the three colours overlap? If they would, the faultline is real, and it is the organisation's design, not the employees' attitudes, that created it. 5.4 Theme four: fairness is a communication problem before it is a pay problem In many hub economies, differences in reward across nationality groups are a legal and market reality that a line manager cannot unilaterally change. That does not release the leader from responsibility for fairness. It relocates the responsibility. Justice research distinguishes between what people receive and how decisions are made. In hubs, the second matters enormously, because the first is often visibly unequal and cannot be quickly fixed. Employees can tolerate a great deal of inequality when they understand the basis of it, when the criteria are applied consistently, when they have a voice in the process, and when they are treated with dignity. They tolerate very little when the rules appear to be invented case by case. Practical consequences follow. Criteria for promotion should be written down and applied to everyone. Where a decision was constrained by policy, such as a nationalisation target, saying so plainly is better than pretending the decision was purely meritocratic; employees generally know the constraint exists, and pretending otherwise destroys credibility on everything else. #pay_transparency in the full sense may not be available, but transparency of process almost always is. Leaders should be especially careful with discretionary rewards, project allocation, and access to visible work, because these are the informal currencies through which favouritism, real or perceived, is expressed. The #status_hierarchy of a hub organisation is often invisible to those at the top of it and painfully visible to those below. Leaders who never eat in the staff canteen, never learn the names of contract workers, and never ask how the night shift is managed will consistently underestimate how their organisation is experienced. 5.5 Theme five: nationalisation must be led, not merely implemented Where nationalisation policies apply, they land on the desk of the line manager as a number. The number, on its own, produces predictable dysfunctions. One is the ceremonial hire, in which a citizen is recruited to satisfy the quota and then given work of little consequence. This is corrosive twice over. The citizen employee learns that the role is symbolic and disengages or leaves. The rest of the team observes that the citizen employee does little and concludes that the stereotype is confirmed. Both groups end up with reinforced negative beliefs, and the organisation has spent money to make its culture worse. Research on citizen job seekers in the Gulf shows precisely this dynamic of mutual stereotype reinforcement between employers and nationals (Elsharnouby et al., 2024). Another is the resentful expatriate, in which experienced foreign staff perceive that their careers have been capped and their knowledge is being extracted for the benefit of people who will replace them. If a senior manager believes she is training her own replacement without recognition or protection, she will train that replacement badly. This is not malice. It is rational self-protection. Leadership can break both dynamics, and the review suggests three levers. First, give citizen employees real work with real accountability from the start, including the possibility of failure. Development without stakes is not development. Second, make the transfer of knowledge an explicitly rewarded part of the experienced employee's role, evaluated and paid for, rather than an unspoken expectation. #knowledge_transfer that threatens the transferrer will not happen. Third, communicate the strategic logic of the policy rather than the quota. Employees understand that a country cannot build a durable economy on a workforce that has no long-term stake in it. What they resist is being handed a number with no explanation. The four-level analysis of nationalisation, covering legal, organisational, human development, and socio-cultural dimensions, makes the same point at the policy level: interventions that operate at only one level fail (Elbanna, 2022). 5.6 Theme six: hybrid dispersion multiplies existing cultural asymmetries Hybrid and remote arrangements are neutral in principle and rarely neutral in practice. In a regional hub, the pattern is often that senior staff are in the head office and junior or delivery staff are elsewhere, and that the geographic split correlates with nationality. Under those conditions, proximity bias is not a minor inconvenience. It is a mechanism that converts a physical arrangement into a career disadvantage for an identifiable group. Recent work on inclusive leadership in hybrid settings identifies the levers available (Detjen & Webber, 2023; Ding et al., 2024). Meeting design is the most powerful. If some participants are remote, treating everyone as remote, with each person on their own screen, removes the room advantage. Rotating meeting times so that the same region is not always inconvenienced distributes the cost of #time_zones. Written asynchronous updates give slower speakers and non-native speakers a fair channel. Explicitly assigning visible tasks to remote members, and naming their contribution in front of senior stakeholders, counteracts the invisibility effect. The leadership question to ask regularly is simple: when I picture the people I have spoken to informally this week, who is on the list, and who is missing? The answer usually reveals the bias. 5.7 Theme seven: knowledge continuity is the true measure of hub leadership The final theme reframes the retention problem. Organisations in hubs often set themselves the goal of reducing turnover. This is worth pursuing, but it has limits. Some turnover is structural and cannot be removed. People leave because their spouse's visa expires, because their children reach an age where they want them schooled at home, because a competitor in another country offers a route to permanent residence. No amount of engagement work changes those facts. The more useful question is not how to stop people leaving but how to ensure that their leaving costs less. This turns #organizational_memory into a leadership responsibility. Several practices follow. Documentation must be treated as part of the work, not as an optional extra done when there is time, because there is never time. Critical knowledge should be deliberately duplicated, so that no single person is the only one who understands a system, a client, or a regulatory requirement. Handover periods should overlap where possible, since a week of shadowing transfers more than a hundred-page document. Exit interviews should be treated as knowledge capture rather than as a formality, and conducted by someone who can act on what they learn. And the leader should assume, at all times, that any member of the team might give notice next month, and ask what would break. This is also where #succession_planning and the #leadership_pipeline meet nationalisation policy. A hub organisation that has systematically built local capability has, in effect, solved its continuity problem, because citizen employees are the part of the workforce least likely to leave the country. Framed this way, localisation ceases to be a compliance burden and becomes a #retention strategy. That reframing is available to any leader willing to take it seriously. 6. Discussion 6.1 Theoretical implications The framework makes three contributions. First, it identifies the high-mobility hub as a boundary condition for existing leadership theory. Most leadership models contain an unstated assumption of temporal stability: that the leader and the led will be together long enough for influence, development, and relationship quality to operate. Where that assumption fails, the models do not become wrong, but their emphasis shifts. Speed of trust formation becomes more important than depth. Explicit structure becomes more important than tacit culture. The article therefore invites a re-examination of familiar constructs under conditions of high churn, which is a research setting that has been largely absent from mainstream leadership journals. Second, it connects two literatures that have developed separately: the micro literature on inclusive leadership and psychological safety, and the macro literature on workforce nationalisation. The connection is not decorative. Nationalisation policy determines the composition of the team, the implicit contracts under which its members work, and the justice perceptions that flow from them. Those perceptions are precisely the variables that inclusive leadership research treats as outcomes. A study of inclusive leadership in a Gulf organisation that ignores nationalisation policy is missing the largest single influence on the fairness perceptions it is trying to explain. Third, it argues that #cultural_distance, the standard variable in cross-cultural research, may be less useful in hubs than variables describing structural alignment. In a team of twelve nationalities, aggregate cultural distance is a nearly meaningless number. Faultline strength, tenure dispersion, language fluency dispersion, and the correlation between nationality and hierarchical level are more informative and more actionable. This is a measurement argument as much as a theoretical one, and it points toward a research agenda that treats the hub as a distinctive configuration rather than as a noisy version of an ordinary subsidiary. 6.2 The paradox at the centre The framework contains a tension that leaders must live with rather than resolve. High-mobility hubs demand both flexibility and consistency. Flexibility, because a leader who cannot vary behaviour across cultural situations will fail. Consistency, because a leader who varies too much will be seen as arbitrary, and arbitrariness destroys the perception of fairness on which everything else rests. The resolution is to be flexible about style and rigid about principle. How feedback is delivered, how a meeting is run, how closeness is expressed, how disagreement is invited: these can and should vary. What counts as good work, how promotion decisions are made, what behaviour is unacceptable, and how the leader responds to a mistake: these must not vary. Leaders who invert this, holding rigidly to a communication style while allowing their standards to shift according to who is in front of them, produce the worst of both worlds. A #paradox_mindset, which treats such tensions as permanent conditions to be managed rather than problems to be solved, is a useful stance here. 6.3 What the evidence does not yet support Intellectual honesty requires stating what remains uncertain. The claim that culturally diverse teams outperform homogeneous ones is not supported as a general proposition; the evidence points to a conditional relationship that depends on process and context (Stahl & Maznevski, 2021). The claim that cultural intelligence training reliably improves performance rests on a body of evidence with acknowledged construct overlap and measurement problems (Yari et al., 2020). The claim that nationalisation policies improve long-term productivity remains contested, and most of the available evidence describes policy design rather than outcomes. The framework offered here should be read as a structured argument built on the best current evidence, not as a settled finding. 7. Practical Implications for Managers and HR Professionals The following practices follow from the analysis. They are stated in operational terms because that is how they must be used. Design onboarding as a leadership act. Give every new joiner a named colleague responsible for them for ninety days. Make implicit norms explicit in writing: how decisions are made here, how disagreement is expressed here, what a deadline means here, when it is acceptable to contact someone outside working hours. Do not assume that norms which are obvious to you are obvious to someone who arrived last week from a different working culture. Audit your faultlines annually. Map nationality against seniority, function, pay band, and contract type. Where these align, act on the structure. Break the alignment through cross-cutting project teams, mixed mentoring pairs, and deliberate variation in who gets visible assignments. Do this before problems appear, not after. Write down the rules of promotion and reward. Publish the criteria. Apply them consistently. Where a decision is constrained by policy or law, say so. Do not use a language of pure meritocracy to describe a system that everyone can see is constrained; the credibility cost is greater than the honesty cost. Redesign meetings for multilingual participation. Agenda in advance. Written contributions permitted. Named questions rather than open ones. Explicit summarising and checking. Slow the pace. Ban idioms in decision-critical communication. Assume misunderstanding until confirmed otherwise. Reward knowledge transfer explicitly. Make it a documented objective for senior staff, evaluated and paid. Never ask someone to train their replacement without recognising what you have asked of them. Give localisation hires real work. Real accountability, real deadlines, real feedback, real consequences. Anything less is a slow-motion failure that damages the individual, the team, and the policy. Treat departures as normal. Conduct handovers seriously. Keep alumni relationships, since they protect #employer_branding in a market where reputation travels fast. Never punish a resignation, because everyone watching learns from how you react. Distribute critical knowledge. No single point of failure. Ask regularly which departures would break the operation, and fix those dependencies before they are tested. Check who you talk to informally. Proximity, language fluency, and cultural similarity all shape who a leader speaks to without thinking. Audit it. Correct it. Measure what matters. Standard engagement scores hide too much. Break results down by nationality group, contract type, tenure, and location. If the average is fine but one group is far below it, the average is not fine. Track early exit rates, participation in meetings, and the distribution of visible assignments. #performance_management systems that cannot detect group-level differences will not detect the problems described in this article. 8. Implications for Students and Early-Career Professionals Many readers of this article will enter one of these labour markets within a few years, either as a graduate hire in a hub economy or as a member of a multicultural team elsewhere. Four points are directly useful. Build cultural agility deliberately, not passively. Living abroad does not automatically develop it. What develops it is depending on people who are different from you to accomplish something that matters, and then thinking honestly about what happened. Seek projects where you are in the minority. Ask for feedback from #expatriates and local colleagues from different backgrounds, and treat the discomfort of receiving it as the actual learning. Keep a short record of situations where you misread someone, and what the correct reading would have been. Learn to read the structure, not just the people. When a team is not working, ask what system produced the pattern. Who was hired through which channel, who reports to whom, who is paid on which contract, who is in the room and who is on the screen. Most conflicts that look personal or cultural have a structural cause, and structural causes can be changed. Take the fluency effect seriously in both directions. If English is your first language, you are being over-credited for clarity and you have an obligation to create space for others. If it is not, you are being under-credited, and you should use the tools that redress this: prepare, write, ask for the agenda, follow up in writing, and do not accept the internal story that you are less capable because you are less fluent. Understand the institutional environment before you accept a role. Ask how the residence permit works, what happens if the job ends, how nationalisation targets affect the function you are joining, and what career path realistically exists for someone in your category. These are legitimate questions. A good employer will answer them. #future_of_work in hub economies rewards those who plan with open eyes, and #reskilling or #upskilling in areas that are portable across countries is a rational hedge against a mobility system you do not control. 9. Limitations and Future Research The limitations of this article are the limitations of a conceptual review. No new data are presented, so the framework is an argument rather than a finding. The concept of the high-mobility hub is defined here in a way that captures a real pattern but has not yet been operationalised with thresholds; future work will need to specify measurable criteria, such as the proportion of foreign-born workers, the median tenure of professional staff, and the growth rate of the relevant sectors. The evidence base is uneven, with far more research from North America and Europe than from the Gulf, and the Gulf research that exists is concentrated in a small number of journals and research groups. The article also treats the hub as a single category, when in reality Dubai, Singapore, and Toronto differ enormously in their legal treatment of migrants, their pathways to permanent residence, and their social structures. Those differences will matter, and they are not modelled here. A #research_agenda follows. Measurement. Develop and validate a hub-context index that combines cultural heterogeneity, tenure dispersion, faultline strength, and institutional mobility constraints. Existing cultural distance measures are inadequate for these settings, and a better instrument would open the field to comparative work. Causal work on churn. Test directly whether the relationships between leadership behaviour and team outcomes are moderated by turnover rate. This is a straightforward moderation hypothesis that has, surprisingly, rarely been tested, and it is central to the argument made here. Nationalisation from the inside. Almost all research on nationalisation looks at policy design or at aggregate labour market outcomes. Very little looks at what happens inside teams. Qualitative work following citizen hires and experienced foreign staff through the first two years of a localisation programme would be highly valuable, and would speak directly to both the policy and the leadership literatures. Knowledge continuity. Test the proposition that performance in high-turnover settings depends more on transfer effectiveness than on retention rate. If it holds, it would justify a reallocation of resources in many organisations from retention initiatives toward knowledge systems. The wellbeing dimension. Mobility has costs. #burnout, family strain, and the psychological load of repeated adaptation are under-studied in hub populations, particularly among lower-paid migrant workers who are almost invisible in the management literature. Research that includes them would be both scientifically and ethically valuable. #wellbeing in these settings is not a soft topic; it is directly connected to safety, error rates, and turnover. Technology. Artificial intelligence tools are already being used for translation, meeting summarisation, and candidate screening in hub organisations. Whether these tools reduce the fluency penalty or entrench it, and whether automated screening reproduces the nationality-based patterns described in this article, is an open and urgent question. 10. Conclusion High-mobility hubs are not simply diverse workplaces. They are workplaces where diversity, turnover, and institutional change arrive together, and where the leadership task is therefore different in kind rather than merely in degree. The leader in such a setting cannot rely on time to build understanding, cannot rely on shared tacit norms, and cannot rely on the people who know things being there next year. What the leader can do is build capability, design structure, make communication explicit, and protect knowledge. The four capabilities set out here, cultural agility, structural fairness, communication design, and knowledge continuity, are demanding but learnable. None of them requires charisma. All of them require attention, honesty, and a willingness to look at the organisation as it is experienced by the people with the least power in it rather than as it appears from the top. For the cities themselves, the stakes are high. #global_cities that have grown quickly on the strength of imported talent are now trying to convert temporary presence into durable capability, through #Vision_2030 style diversification programmes and through localisation policy. Whether that conversion succeeds will depend less on legislation than on the quality of leadership inside thousands of ordinary teams, where the policy is either made real or quietly hollowed out. That is an unglamorous conclusion, but it is the correct one. The future of these #labour_market systems is being decided in meeting rooms, one handover, one promotion decision, and one difficult conversation at a time. Hashtags #CrossCulturalLeadership #HighMobilityHubs #DiverseTalentPools #ManagingGlobalTeams #InternationalHRM #GlobalTalentStrategy #LeadingAcrossCultures #TalentMobilityResearch #MulticulturalWorkplace #EmergingMarketLeadership #GulfWorkforceStudies #InclusiveGlobalLeadership #CultureAndManagement #WorkforceLocalisation #StudentResearchArticle References Backmann, J., Kanitz, R., Tian, A. W., Hoffmann, P., & Hoegl, M. (2020). Cultural gap bridging in multinational teams. Journal of International Business Studies, 51(8), 1283-1311. Caligiuri, P. (2021). Build your cultural agility: The nine competencies you need to be a successful global professional. London: Kogan Page. Caligiuri, P. (2023). The development of cultural agility competencies through global mobility. Journal of Global Mobility, 11(2), 145-158. Caligiuri, P., & Caprar, D. V. (2022). Becoming culturally agile: Effectively varying contextual responses through international experience and cross-cultural competencies. The International Journal of Human Resource Management (advance online publication). Caligiuri, P., Collings, D. G., De Cieri, H., & Lazarova, M. B. (2024). Global talent management: A critical review and research agenda for the new organizational reality. Annual Review of Organizational Psychology and Organizational Behavior, 11, 393-421. Caligiuri, P., De Cieri, H., Minbaeva, D., Verbeke, A., & Zimmermann, A. (2020). International HRM insights for navigating the COVID-19 pandemic: Implications for future research and practice. Journal of International Business Studies, 51(5), 697-713. Collings, D. G., McMackin, J., Nyberg, A. J., & Wright, P. M. (2021). Strategic human resource management and COVID-19: Emerging challenges and research opportunities. Journal of Management Studies, 58(5), 1378-1382. da Silva, A. B., Castello-Sirvent, F., & Canos-Daros, L. (2022). Sensible leaders and hybrid working: Challenges for talent management. Sustainability, 14(24), 16883. Den Hartog, D. N., & De Hoogh, A. H. B. (2024). Cross-cultural leadership: What we know, what we need to know, and where we need to go. Annual Review of Organizational Psychology and Organizational Behavior, 11, 535-566. Detjen, J., & Webber, S. S. (2023). Leading hybrid teams in a transition to the future knowledge workplace. Strategy and Leadership, 51(6), 16-21. Ding, G., Ren, X., & Lin, F. (2024). Adopting active or passive leadership in project-based organizations? The role of inclusive leadership in remote work environment. International Journal of Project Management, 42(6), 102623. Elbanna, S. (2022). Policy and practical implications for workforce nationalization in the Gulf Cooperation Council (GCC) countries. Personnel Review, 51(4), 1248-1261. Elsharnouby, T. H., Elbanna, S., Obeidat, S. M., & Mauji, N. I. (2024). Exploring local job seekers perception toward employment in the private and public sectors in Qatar: Implications for workforce nationalization policies. Personnel Review, 53(6), 1425-1442. Gallardo-Gallardo, E., Thunnissen, M., & Scullion, H. (2020). Talent management: Context matters. The International Journal of Human Resource Management, 31(4), 457-473. Jiang, J., Ding, W., Wang, R., & Li, S. (2022). Inclusive leadership and employees' voice behavior: A moderated mediation model. Current Psychology, 41(9), 6395-6405. Korkmaz, A. V., van Engen, M. L., Knappert, L., & Schalk, R. (2022). About and beyond leading uniqueness and belongingness: A systematic review of inclusive leadership research. Human Resource Management Review, 32(4), 100894. Lanvin, B., & Monteiro, F. (2022). The global talent competitiveness index 2022: The tectonics of talent. Fontainebleau: INSEAD, Portulans Institute and Human Capital Leadership Institute. Lazarova, M., Caligiuri, P., Collings, D. G., & De Cieri, H. (2023). Global work in a rapidly changing world: Implications for MNEs and individuals. Journal of World Business, 58(1), 101365. Michailova, S., Fee, A., & DeNisi, A. (2023). Research on host-country nationals in multinational enterprises: The last five decades and ways forward. Journal of World Business, 58(1). Mohase, K., Donald, F., & Israel, N. (2025). Inclusive leadership, psychological safety, and employee voice in remote and hybrid work employees. South African Journal of Psychology (advance online publication). Pereira, V., Collings, D. G., Wood, G., & Mellahi, K. (2022). Evaluating talent management in emerging market economies: Societal, firm and individual perspectives. The International Journal of Human Resource Management, 33(11), 2171-2191. Sahu, M. (2025). Impact of the nationalisation of the labour force on South Asian migrants in the Gulf Cooperation Council states. Journal of Asian and African Studies (advance online publication). Seijts, G. H., & Milani, K. Y. (2022). The application of leader character to building cultures of equity, diversity, and inclusion. Business Horizons, 65(5), 573-590. Stahl, G. K., & Maznevski, M. L. (2021). Unraveling the effects of cultural diversity in teams: A retrospective of research on multicultural work groups and an agenda for future research. Journal of International Business Studies, 52(1), 4-22. Taras, V., Steel, P., & Stackhouse, M. (2023). A comparative evaluation of seven instruments for measuring values comprising Hofstede's model of culture. Journal of World Business, 58(1). van Knippenberg, D., & van Ginkel, W. P. (2022). A diversity mindset perspective on inclusive leadership. Group and Organization Management, 47(4), 779-797. Vardi, S., & Collings, D. G. (2023). What's in a name? Talent: A review and research agenda. Human Resource Management Journal, 33(3), 660-682. Yari, N., Lankut, E., Alon, I., & Richter, N. F. (2020). Cultural intelligence, global mindset, and cross-cultural competencies: A systematic review using bibliometric methods. Journal of International Entrepreneurship, 18(2), 210-250. Zakaria, N., & Mohd Yusof, S. A. (2020). Crossing cultural boundaries using the internet: Toward building a model of swift trust formation in global virtual teams. Journal of International Management, 26(1).

Generative artificial intelligence has moved from research laboratories into the daily working life of founders, students, marketers, and software teams in a very short period of time. Text, images, code, and video can now be produced at a speed and a scale that no human team can match. This change has created a problem that sits at the meeting point of business practice and moral judgement. When a machine can write a business plan, a pitch deck, a research paper, or an advertising campaign, what does it still mean for a venture to claim that its work is original, honest, and its own? This article studies the relationship between #entrepreneurial_ethics and #AI_detection, and asks how new ventures can protect #integrity and #originality inside #tech_driven_business_models. The study uses an integrative review of scholarly work published mainly between 2020 and 2025, combined with a conceptual synthesis that produces a practical governance framework. The review shows three consistent findings. First, current detection tools are useful signals but weak proofs, because they produce both #false_positives and #false_negatives at rates that are too high for high stakes decisions. Second, detection performance is unevenly distributed, and writers who learned English as an additional language are penalised more often than others, which turns a technical weakness into a question of #fairness. Third, evasion is cheap, because light editing and simple rewriting can defeat most classifiers. From these findings the article argues that detection alone cannot carry the ethical load. It proposes an integrity model built on four pillars, namely #disclosure, #provenance, #verification, and #accountability, and it translates the model into concrete practices for founders, teams, educators, and student entrepreneurs. The article concludes that integrity in the age of generative systems is not a matter of catching cheaters. It is a matter of designing work processes in which honest practice is the easiest path to take. Keywords: entrepreneurial ethics, artificial intelligence detection, generative artificial intelligence, academic integrity, originality, business governance, responsible innovation, startups Introduction 1.1 The setting of the problem A student in a business incubator writes a market analysis in twenty minutes with the help of a chatbot. A founder produces sixty variations of a landing page in one afternoon. A consultant delivers a report that reads well, cites confidently, and contains three references that do not exist. A university flags a first year assignment as machine written, and the student, who wrote every word by hand, cannot prove otherwise. Each of these situations is now ordinary. Together they describe a working environment in which the old signals of effort, skill, and honesty have quietly stopped working. For most of modern commercial history, the link between effort and output was reliable enough to be used as a proxy for character. A polished document suggested a careful author. A working prototype suggested a competent engineer. A well argued strategy suggested a thoughtful mind. That link has weakened. #generative_AI systems, built on #large_language_models, can produce fluent output without understanding, without effort, and without any claim to authorship in the older sense of the word. The output looks like work. It may even be good work. But the signal it once carried about the person behind it has become unreliable. This creates a specific problem for entrepreneurship. New ventures live on trust that they have not yet earned. A young company has no long record of delivery, no decades of audited accounts, and no established brand. What it has instead is a story, a team, a set of documents, and a promise. Investors, customers, partners, and regulators must decide whether to believe that promise on thin evidence. Anything that weakens the credibility of documents and claims therefore strikes directly at the foundation of entrepreneurial life. If a pitch deck can be generated in minutes, its persuasive power tells the reader less than it once did. If customer reviews can be manufactured at scale, they inform less. If a technical white paper can be assembled by a model that has never built anything, it proves less. 1.2 Why detection became the first response The first institutional reaction to this shift was defensive. Universities, publishers, and some employers turned to #detection_tools, which are software systems that try to decide whether a text was written by a human or by a machine. The logic was familiar, because it repeated the earlier response to #plagiarism, which was similarity checking. If machines create the problem, machines can find it. That logic is understandable but incomplete. Similarity checking worked reasonably well because it compared a document against a corpus of existing documents, and a match was a fact that could be shown and examined. Machine generated text has no source document to match. It is new text, produced probabilistically, and so detection must rely on statistical properties of the writing itself, such as predictability, sentence rhythm, vocabulary distribution, and the smoothness of word choice. These are indirect clues, not evidence in the ordinary sense. A person can write predictably. A machine can be made to write unpredictably. The clue and the conclusion are separated by a large gap of uncertainty. Empirical studies of detection performance have made this gap visible. Independent testing has repeatedly found accuracy that falls far short of what an institution would need before accusing a person of misconduct, and the same testing has shown that light editing degrades detection sharply. This does not mean detection is worthless. It means detection is a smoke alarm rather than a court of law, and treating it as the second thing rather than the first has caused real harm to real people. 1.3 The ethical shift this article proposes The central argument of this article is that the problem has been framed too narrowly. Institutions and firms have asked a technical question, which is how do we detect machine writing. The more useful question is an ethical and organisational one, which is how do we build ways of working in which honest, transparent, and genuinely valuable contributions are recognised, rewarded, and easy to demonstrate. This reframing moves the discussion from policing to design. It shifts attention from the finished artefact, which is now easy to fake, towards the process that produced it, which is much harder to fake. It replaces the binary question of human or machine with the more honest question of what did the person actually contribute, understand, and take responsibility for. And it treats #transparency not as a punishment but as a professional standard, in the same way that a scientist declares funding sources or an accountant declares a conflict of interest. For entrepreneurs this reframing has commercial as well as moral value. A venture that can show how its work was made, who checked it, and what a human being is willing to defend under questioning holds an asset that a competitor using undisclosed automation does not hold. In a market flooded with cheap synthetic content, verifiable honesty becomes scarce, and scarce things acquire value. 1.4 Research questions The study is guided by four questions. The first question asks how the arrival of generative systems changes the meaning of #originality and #authorship in entrepreneurial work, where documents are instruments of persuasion rather than contributions to knowledge. The second question asks what the empirical literature actually establishes about the reliability, bias, and robustness of #AI_detection, and what limits those findings place on the legitimate use of detection in academic and commercial decisions. The third question asks which ethical risks are specific to new ventures, as distinct from the risks already discussed in the education literature, and how those risks interact with investor relations, marketing, hiring, and product development. The fourth question asks what a workable governance model looks like for organisations that have very few people, very little time, and no compliance department, which is the normal condition of a startup. 1.5 Contribution and structure The contribution of the article is threefold. It offers a synthesis of a scattered literature that currently sits in separate fields, namely educational integrity, machine learning, business ethics, and entrepreneurship studies. It provides an evidence based assessment of what detection can and cannot support. And it converts that assessment into a governance framework that a team of three people in a rented room can actually operate. The remainder of the article proceeds as follows. Section two reviews the relevant literature. Section three develops the conceptual framework and states the propositions. Section four explains the method. Section five presents the findings. Section six discusses their meaning for entrepreneurship. Section seven sets out the practical framework and implementation guidance. Section eight considers limitations. Section nine proposes a research agenda. Section ten concludes. Literature Review 2.1 Entrepreneurial ethics as a field Business ethics has long distinguished between compliance and character. Compliance asks what the rules require. Character asks what a good actor would do when the rules are silent, ambiguous, or unenforced. Entrepreneurship lives mostly in the second space. New ventures operate in markets that regulators have not yet mapped, using technologies whose consequences are not yet understood, under time pressure that discourages reflection. The founder is often the only person in the room who can say no. Three features of the entrepreneurial context make ethical questions sharper than they are in established firms. The first is resource scarcity. When a venture has three months of cash left, the temptation to overstate a metric, to present a prototype as a product, or to describe a pilot customer as a paying client becomes strong. The second is concentrated authority. A founder frequently controls hiring, accounting, communication, and product decisions at once, which removes the internal checks that constrain managers in larger organisations. The third is uncertainty about consequences. A new technology may harm users in ways that nobody anticipated, and the venture that builds it usually lacks the resources to study those harms carefully. Research on artificial intelligence in entrepreneurship has mostly concentrated on opportunity, showing how machine learning improves opportunity recognition, resource assembly, and decision making under uncertainty. Studies have described how algorithmic tools reshape venture creation and how founders use data to reduce guesswork. This work is valuable, but its ethical dimension has remained thin. The literature tells us that entrepreneurs can now do more, faster, with fewer people. It says much less about what they owe to the people who read, buy, fund, and rely on the output of that acceleration. 2.2 Generative systems and the collapse of costly signals Economic reasoning about trust often relies on the idea of a costly signal. A signal is informative when it is expensive to produce for those who lack the underlying quality and cheap for those who possess it. A degree signals persistence. A detailed technical document signals expertise. A responsive support team signals commitment. The signal works because faking it would cost more than it is worth. #generative_AI attacks this mechanism directly by lowering the cost of the signal without changing the underlying quality. A person with no expertise can now produce a document that reads as though it were written by an expert. The document is no longer costly, and so it is no longer informative. This is not merely an academic observation. It explains why so many institutions feel that something important has broken. What broke is the inference from artefact to ability. The consequences appear in several places at once. Recruiters can no longer read a cover letter as evidence of writing skill. Investors can no longer read a market section as evidence of market understanding. Teachers can no longer read an essay as evidence of learning. Customers can no longer read a review as evidence of experience. Each of these breakdowns is a small collapse of trust, and together they impose a real cost on honest actors, because honest actors are now forced to prove things that used to be taken for granted. 2.3 How AI detection works and why it struggles Detection systems generally exploit the statistical fingerprints of machine text. Language models produce text by choosing likely next tokens, which tends to yield writing that is smoother and more predictable than human writing. Detectors measure this predictability, often through measures related to #perplexity, which captures how surprised a model is by a given sequence, and burstiness, which captures how much variation exists between sentences. Human writing tends to lurch, digress, and vary. Machine writing tends to glide. More sophisticated approaches include curvature based methods, which test whether small perturbations of a text raise or lower its likelihood under a reference model, and supervised classifiers trained on labelled human and machine samples. A separate approach is #watermarking, in which the generating model deliberately biases its token choices according to a secret pattern that can later be recovered statistically. Watermarking is technically elegant because it embeds the signal at the point of creation rather than trying to recover it afterwards. Each approach carries a structural weakness. Predictability based detection punishes clear, conventional, simple writing, which is exactly the writing that careful professionals and second language users often produce. Classifier based detection depends on training data that ages quickly as models change. Watermarking works only if the model provider applies it, and it can be weakened by rewriting, translation, or passing the text through a second model. None of these are minor engineering problems that will be solved next year. They follow from the nature of the task, because the space of good human writing and the space of good machine writing overlap and the overlap is growing. 2.4 Empirical evidence on detector reliability Independent evaluations have converged on a discouraging picture. Comparative testing of widely used tools has found that accuracy on unedited machine text is moderate at best and that performance collapses when the text is lightly obfuscated through paraphrasing, machine translation, or human editing. Studies of detection in higher education have found that combining software output with human academic judgement improves results, but that neither alone reaches the confidence level that a disciplinary process requires. Two findings deserve special attention because they carry ethical weight rather than merely technical weight. The first is #detector_bias. Research has shown that classifiers frequently misclassify writing by #non_native_speakers as machine generated, apparently because such writing often uses a narrower vocabulary and more conventional constructions, which raises its predictability score. The result is a systematic penalty falling on people who are already disadvantaged in English dominated academic and commercial systems. A tool that is wrong in a random way is a nuisance. A tool that is wrong in a patterned way, against an identifiable group, is a justice problem. The second is the fragility of detection under #adversarial_attacks. Simple interventions defeat most detectors. Running text through a paraphrasing service, asking a model to write in a more irregular style, introducing minor errors, or editing a few sentences by hand can move a document from flagged to clean. Because these interventions are trivial and freely available, detection reliably catches the careless and reliably misses the deliberate. This is close to the worst possible distribution of enforcement, since it punishes those who did not try to hide and excuses those who did. 2.5 Originality, authorship, and ownership Behind the technical debate sits an older conceptual question. What is #originality, and who is the author of a text that a machine helped produce. Scholarly and legal traditions have generally located authorship in intention, judgement, and responsibility rather than in the mechanical act of producing words. A researcher who dictates to a typist is still the author. An architect who uses drafting software is still the designer. On this view, a founder who directs a model, evaluates its output, corrects its errors, verifies its claims, and takes public responsibility for the result has a genuine claim to authorship, while a founder who copies unread output and signs their name to it does not. The difficulty is that these two cases look identical from the outside. The finished document does not reveal how it was made. This is precisely why the ethical centre of gravity must move from the artefact to the process. Norms in scholarly publishing have already moved in this direction, since major publishers now hold that a machine cannot be listed as an author because it cannot take #accountability for the work, while requiring human authors to declare where and how such tools were used. The principle is sound and it exports well to commercial settings. Responsibility cannot be delegated to a system that cannot be held responsible. The related question of #intellectual_property remains unsettled in many jurisdictions. Ventures that build products on generated text, images, or code face genuine uncertainty about what they own, what they may license, and what liability they carry if generated material reproduces protected work from training data. This uncertainty is a business risk as much as an ethical one, and prudent founders treat it as such. 2.6 Ethics washing and its commercial cousin The critical literature on artificial intelligence has warned about #ethics_washing, which is the practice of adopting the vocabulary of responsibility while avoiding its substance. Principles are published, committees are formed, statements are issued, and behaviour does not change. The commercial cousin of this practice is #AI_washing, which is the practice of describing ordinary software as artificial intelligence in order to attract capital and attention. Both are forms of misrepresentation, and both are especially tempting for early stage ventures, where the gap between the story and the substance is naturally largest and hardest for outsiders to inspect. The connection to detection is direct. A firm that automates its content, conceals that it has done so, and simultaneously markets itself as a champion of responsible technology has committed two related deceptions. The first concerns the origin of its words. The second concerns the character of its practice. Regulators in several markets have begun to treat exaggerated claims about artificial intelligence as ordinary misrepresentation, which suggests that the reputational risk is becoming a legal risk. 2.7 The research gap Three bodies of work therefore run in parallel without meeting. The education literature studies #academic_integrity and detection but rarely considers commercial incentives. The machine learning literature studies detection and watermarking but treats the institutional question as somebody else's problem. The business ethics and entrepreneurship literature studies governance and responsibility but has not yet absorbed the specific mechanics of generative systems. The gap this article addresses is the space where all three meet, which is the working life of the student entrepreneur and the small venture, where academic norms, technical constraints, and commercial pressure collide inside the same person. Conceptual Framework 3.1 From artefact to process The framework proposed here rests on a single move. Judgement should attach to the process that produced a piece of work, not only to the work itself. This move is not new. Auditing, clinical practice, engineering certification, and scientific peer review all rest on it. What is new is the necessity of applying it to ordinary written and creative work, which previously carried its own evidence of effort. Once the move is made, four requirements follow naturally. They form the pillars of the framework. 3.2 Pillar one, disclosure #disclosure means stating, in a form appropriate to the context, where and how automated systems contributed to a piece of work. Disclosure is not confession, and it should not be treated as an admission of weakness. It is the same category of statement as a footnote, a funding declaration, or a materials and methods section. Good disclosure has three properties. It is specific, naming the stage of work at which the tool was used, such as brainstorming, drafting, translation, code generation, or data cleaning. It is proportionate, so that a memo does not require the same detail as a regulatory filing. And it is standardised within an organisation, so that readers know what a given phrase means. Disclosure solves a coordination problem. In the absence of a shared norm, honest people who disclose appear less capable than dishonest people who do not, which punishes honesty and drives disclosure out of the market. A clear institutional standard removes that penalty by making disclosure normal. 3.3 Pillar two, provenance #provenance means retaining evidence of how a piece of work came into existence. In practice this means drafts, notes, version histories, prompt records, editing trails, and the sources consulted. The value of provenance is that it is expensive to fabricate after the fact and cheap to keep at the time. Provenance turns an unanswerable question into an answerable one. A detector can only tell you that a document looks machine written, which is a guess. A version history can show that a document was written across nine sessions over two weeks, with sources gathered before the argument formed, which is a record. Where detection generates suspicion, provenance can dissolve it. This asymmetry is the practical heart of the framework, because it gives the accused person something to stand on, and it gives the institution something to examine. 3.4 Pillar three, verification #verification means that a competent human being has checked the substantive claims of the work before it leaves the organisation. Language models produce confident text without any commitment to truth, and the resulting errors, often called #hallucination, include invented citations, misattributed statistics, fabricated case law, and plausible but wrong technical detail. For a venture, verification is where ethics and self interest align most tightly. A false claim in a pitch deck is not only dishonest, it is dangerous, because it will be discovered during #due_diligence, and the discovery will destroy the trust that the entire relationship depends on. Verification should therefore be treated as a hard gate, not as an optional review. Nothing that asserts a fact about the world should reach an investor, a customer, or a regulator without a named person who has checked it and is prepared to defend it. 3.5 Pillar four, accountability #accountability means that a specific human being owns the output and its consequences. This pillar exists because the other three can be performed cosmetically. A team can disclose vaguely, retain messy files, and conduct a shallow review. Accountability closes the loop by attaching a name. The practical test is simple and can be applied in any meeting. If a claim in a document is challenged, is there a person in the room who can explain how it was reached, why it is believed to be true, and what would change their mind. If there is no such person, the document should not have been sent, regardless of how it was produced. 3.6 The integrity equation and its propositions Taken together, the pillars support a small set of propositions that guide the analysis that follows. Proposition one holds that in generative environments, the evidential value of a finished artefact declines while the evidential value of process records rises. Institutions that continue to judge only artefacts will therefore make increasingly poor judgements. Proposition two holds that detection technologies can support integrity only as triggers for inquiry, and that using them as proof produces injustice which is concentrated among already disadvantaged groups. Proposition three holds that voluntary, specific, and normalised disclosure increases perceived trustworthiness in commercial relationships, because it signals that the discloser has nothing to hide and understands their own process. Proposition four holds that ventures which invest early in lightweight governance covering disclosure, provenance, verification, and accountability incur a small ongoing cost and avoid a large tail risk, and that this trade is favourable under most plausible assumptions. Proposition five holds that #human_oversight is the irreducible element of integrity, because responsibility can only be borne by an entity capable of understanding consequences and of being sanctioned. Methodology 4.1 Design This study uses an integrative literature review combined with conceptual development. An integrative review is appropriate when a topic is emerging, when the relevant evidence is spread across disciplines that use different vocabularies, and when the goal is synthesis and framework building rather than statistical aggregation. The subject here satisfies all three conditions, since the evidence sits in education research, machine learning, information systems, business ethics, and law. 4.2 Sources and selection The review draws on peer reviewed journal articles, peer reviewed conference proceedings, scholarly books, and official regulatory documents. Priority was given to work published between 2020 and 2025, because the technical and institutional landscape before that period differs so substantially that older findings about detection are no longer reliable guides. Material was included when it addressed at least one of the following. The empirical performance of AI text detection. The ethical implications of generative systems in education or business. The governance, regulation, or policy response to generative systems. The role of artificial intelligence in entrepreneurship and venture creation. Material was excluded when it consisted of vendor marketing, when its methods were not described sufficiently to be assessed, or when it made empirical claims about detector accuracy without reporting how those claims were tested. 4.3 Analysis Sources were read and coded thematically. Three clusters emerged. The first cluster concerns technical capability and its limits. The second concerns harm, fairness, and the distribution of error. The third concerns institutional response, covering policy, regulation, and organisational practice. These clusters structure the findings section that follows. The conceptual framework was then developed by asking, for each documented limitation of detection, what organisational practice would compensate for it. This produced the four pillars, which were checked against the constraints that actually bind small ventures, namely limited time, limited money, limited staff, and the absence of any dedicated compliance function. 4.4 Limitations of the method An integrative review reflects the literature that exists, and the literature on this topic is young, fast moving, and unevenly distributed. Detection research in particular ages quickly, since a finding about a model released two years ago may not hold for a model released this year. Findings about detector accuracy should therefore be read as evidence about a class of approaches rather than as fixed measurements. The conceptual contribution has not yet been tested empirically, and that limitation is discussed further in section eight. Findings 5.1 Detection is a signal, not a verdict The most consistent finding across independent evaluations is that AI text detection performs far below the standard required for consequential decisions. Tools that report high accuracy in vendor materials perform substantially worse under independent testing, particularly on text that has been edited, paraphrased, or translated. Reported accuracy varies widely across tools and across text types, and no tool has demonstrated the kind of stable, reproducible performance that would justify treating its output as proof. This has a direct implication that institutions have been slow to accept. A detection score is admissible as a reason to look more closely. It is not admissible as a reason to conclude. The difference matters because the two uses have very different consequences for the person on the receiving end. Used as a trigger, a detector starts a conversation in which the person can show their drafts, explain their argument, and answer questions about their sources. Used as a verdict, it ends the conversation before the person has spoken. 5.2 The asymmetry of harm #false_positives and #false_negatives are not symmetric in their consequences, and this asymmetry is often ignored in technical evaluations that report only overall accuracy. A false negative means that a person who used a machine without disclosure is not caught. The cost is a modest unfairness, a mark awarded that was not earned, a document circulated that was not really written. This is a genuine harm but it is diffuse and recoverable. A false positive means that an innocent person is accused of dishonesty on the basis of a statistical guess. The cost includes reputational damage, psychological distress, disciplinary proceedings, and in serious cases the loss of a degree, a job, or a visa. It is concentrated, severe, and often irreversible. Because the person cannot prove a negative, they are placed in the position of having to disprove a machine, which is an impossible task without process records. Institutions that weigh these two errors equally are therefore making a moral error, not merely a statistical one. Any policy that relies on detection must be built around the principle that a false accusation is far worse than a missed offence, and #proportionality in sanctions must follow from that. 5.3 Bias falls on the already disadvantaged Research on classifier behaviour has found that texts written by people who learned English as an additional language are disproportionately flagged as machine generated. The mechanism is intelligible. Such writers often use a more limited and more conventional set of words and structures, which produces exactly the low variability that detectors read as a machine signature. The implication for #equity is serious. International students, migrant founders, and non Anglophone research teams already carry a disadvantage in English dominated systems. A detection regime that adds an elevated risk of false accusation on top of that disadvantage compounds an existing injustice with a technical one. Any institution deploying detection without addressing this pattern is, in effect, running a policy whose burden falls hardest on those least able to contest it. The same logic extends to neurodivergent writers, to writers trained in highly formulaic professional registers, and to anyone whose style is regular by habit or by training. Regularity of expression is not evidence of dishonesty, and treating it as such confuses a stylistic property with a moral one. 5.4 Evasion is cheap and detection is therefore regressive Studies of #adversarial_attacks show that the barrier to defeating detection is very low. Paraphrasing tools, style instructions, round trip translation, and modest manual editing all reduce detector confidence substantially. Access to these methods requires no technical skill and often no money. The consequence is a regressive enforcement pattern. The person who uses a model openly and clumsily is flagged. The person who uses it strategically and hides it is not. Enforcement therefore selects for concealment rather than for honesty, which is the precise opposite of what an integrity system should do. This is a strong argument against building any serious policy on detection as its foundation, because a system that rewards sophistication in hiding is a system that trains people to hide. 5.5 Provenance and watermarking as partial answers Two approaches attempt to solve the problem at the point of creation rather than after the fact. #watermarking embeds a statistical signature into generated text by biasing the model's token choices in a way that a holder of the key can later detect. It is technically promising, but it depends on voluntary adoption by model providers, it does not cover open models that users can run themselves, and it degrades under rewriting and translation. It is therefore useful as part of a broader system and useless as a sole defence. Provenance, by contrast, works from the human side. Version histories, drafts, editing timelines, and structured records of tool use create an #audit_trail that is easy to produce honestly and awkward to fabricate. Provenance also has a property that detection lacks, which is that it protects the innocent rather than merely accusing the guilty. A student or founder with a full #version_history can answer an accusation immediately. This is why the framework in this article places provenance ahead of detection in importance. 5.6 Risks specific to new ventures The education literature has documented risks to learning. The entrepreneurial context adds risks of a different kind, which are less discussed and in some cases more consequential. Investor communications carry the sharpest risk. A generated market section that invents a figure, misstates a competitor's funding, or cites a report that does not exist will be found during #due_diligence. The discovery will not be interpreted as carelessness. It will be interpreted as dishonesty, because the investor cannot distinguish the two from the outside, and the rational response to that uncertainty is to walk away. The asymmetry here is brutal. Generation saves a founder a few hours. Discovery costs the round. Marketing content carries a slower risk. Mass produced copy is cheap, and cheap copy is now abundant, which means it no longer differentiates. Worse, generated claims about product performance, health effects, safety, or environmental impact can cross into regulated territory, where the firm is liable regardless of who or what wrote the sentence. The machine does not share the liability. Product and code carry a technical and legal risk. Generated code may carry security weaknesses, licence obligations, or reproduced fragments from training data. A venture that ships such code without review has inherited a liability it cannot see. Customer interaction carries a trust risk. Synthetic reviews, undisclosed automated agents posing as human staff, and manufactured testimonials are straightforward deception, and they are increasingly treated as such by consumer protection authorities. The reputational damage of exposure is out of all proportion to the saving. Hiring carries a symmetrical risk that founders often overlook. Candidates now use generation to produce applications, and firms use automation to screen them, which produces a strange situation in which machines write to machines and no human learns anything about any other human. Ventures that redesign hiring around demonstrated work, live problem solving, and conversation recover information that automated screening has destroyed. 5.7 The regulatory environment Regulation has begun to catch up, and it has generally chosen transparency rather than prohibition as its main instrument. The European Union has adopted a risk based statute that imposes obligations proportionate to the level of risk a system creates, and that requires, among other things, that people be informed when they are interacting with a machine and that certain synthetic content be marked as such. Whatever a founder thinks of the details, the direction is clear. #transparency about machine involvement is moving from a courtesy to a duty. In the United States, the National Institute of Standards and Technology has published a voluntary risk management framework organised around governing, mapping, measuring, and managing risk. It is not law, but it functions as a reference point that larger customers increasingly expect suppliers to follow. The practical lesson for a small venture is straightforward. Disclosure practices adopted now, voluntarily and cheaply, are likely to become requirements later, and it is far less expensive to build them into the culture of a five person team than to retrofit them into a fifty person organisation under regulatory pressure. Discussion 6.1 Integrity as a scarce asset Standard economic reasoning suggests that when the supply of something rises sharply, its price falls. The supply of plausible content has risen to the point where it is effectively unlimited. Its price has therefore collapsed. What has not become abundant is verified content, which is content whose claims a named human being has checked and will defend. This suggests that integrity is moving from the cost side of the ledger to the value side. A venture that can say, with evidence, that a human being reviewed every factual claim in its documentation is offering something that its competitors are not offering, and that its customers increasingly need. The same holds for research, journalism, consulting, and education. Where synthetic output is cheap, the guarantee of human judgement becomes the product. This is not a sentimental argument. It is a claim about #competitive_advantage under changed conditions of scarcity, and it is testable. If it is right, we should observe firms beginning to advertise verification and provenance, and we should observe buyers beginning to ask for them. Early signs in professional services suggest that this is already happening. 6.2 Signalling under uncertainty If the artefact no longer signals ability, ventures must find new signals. Three candidates emerge from this analysis. The first is process evidence. Showing how the work was made, including the drafts, the dead ends, and the corrections, is now more persuasive than showing the polished result, because the polished result is cheap and the record of struggle is not. The second is live demonstration. Answering unscripted questions, working through a problem in real time, and being examined on the reasoning behind a claim cannot be delegated to a model in the room. This is why interviews, oral defences, and working sessions are regaining importance, and why #authentic_assessment in education is converging on the same logic. The third is voluntary disclosure. Stating plainly which tools were used and for what purpose signals confidence rather than weakness. A founder who says that a model drafted the first version of a section, that two team members verified every figure, and that the argument is their own is telling the reader something useful about how the venture works. A founder who says nothing leaves the reader to guess, and under current conditions the guess will not be generous. 6.3 Governance for organisations that have no time Most governance advice fails in startups because it assumes resources that startups do not have. A framework requiring committees, quarterly audits, and policy documents will be ignored, and being ignored is worse than being absent, because it creates the appearance of control without the substance. The framework proposed here is therefore deliberately minimal. It can be operated by a team of three. It requires one page of written #policy, a habit of keeping drafts, a rule that factual claims are checked by a named person, and a short disclosure convention. None of this requires software purchases, external consultants, or dedicated staff. The deeper point is cultural. In a small team, #culture is set by what the founder does when a deadline is close and nobody is watching. If the founder ships unverified generated claims under pressure, the team learns that the rule is decorative. If the founder holds the line, the team learns that it is real. No document can substitute for that demonstration, and no policy survives a founder who breaks it. 6.4 The educational implications For universities and business schools, the findings suggest a shift in how work is assessed. If the artefact is no longer a reliable indicator of learning, then assessment must reach for other indicators. Options include work that requires personal or local data that no model has seen, tasks that require the student to explain their reasoning aloud, portfolios that show development over time, and assignments that require the student to critique, correct, and improve machine output rather than to compete with it. This last option deserves emphasis, because it converts the threat into a teaching instrument. A student who is given a generated business plan and asked to find its errors, check its numbers, identify its unstated assumptions, and defend a revised version has learned more about analysis than a student who wrote a mediocre plan from scratch. #critical_thinking is exercised in the correction, and correction requires understanding. For #student_entrepreneurs specifically, the connection between the classroom and the venture is direct, and it should be made explicit. The habits that a student forms while writing coursework will be the habits they carry into investor communications. A student who learns to verify sources, keep drafts, and disclose assistance will find that these behaviours are exactly what #investors and enterprise customers expect. A student who learns to generate and submit without reading will carry that habit into a context where the consequences are financial and legal rather than academic. #business_education therefore has a duty to teach #AI_literacy as an ethical practice and not merely as a technical skill. 6.5 What should not be done It is worth stating the negative recommendations plainly, because several institutions have already made these mistakes. Detection scores should not be used as the sole basis for a misconduct finding. The evidence does not support it, and the harm from error falls disproportionately on the vulnerable. Blanket prohibition of generative tools should be avoided. It is unenforceable, it drives use underground, and it disadvantages the students and founders who follow the rules while doing nothing to constrain those who do not. It also fails to prepare people for a working world in which these tools are standard. Disclosure should not be punished. If a person who declares assistance is treated more harshly than a person who conceals it, the institution has built an incentive to conceal, and it will get what it has designed. Policies should not be written in language that only lawyers understand. A rule that a first year student or a first time founder cannot understand and apply on a Tuesday afternoon is not a rule, it is a decoration. A Practical Framework for Ventures and Student Teams 7.1 The four questions test Before any document leaves the organisation, four questions should be answerable. Who is responsible for this. There must be a named person, not a team and not a tool. What is the origin of the claims. Every factual assertion should be traceable to a source that a human being has actually looked at. What did the machine do. The stage and extent of automated assistance should be known internally, even where it is not disclosed externally. Could we defend this under questioning. If the answer is no, the document is not ready. These four questions take under a minute and they catch most of the serious failures. 7.2 A tiered disclosure standard Not every document needs the same treatment. A tiered standard keeps the cost of integrity proportional to the stakes. Tier one covers internal and low stakes material, such as notes, brainstorms, internal drafts, and routine correspondence. Use of generative tools is unrestricted and no external disclosure is required, but factual claims that will travel outward must be verified before they do. Tier two covers external non binding material, such as marketing copy, blog posts, and social content. Human review of all factual and comparative claims is required. A brief statement of tool use is recommended where the content is presented as expert commentary. Tier three covers external binding or high stakes material, such as investor communications, contracts, regulatory filings, safety information, financial statements, and academic submissions. Here the requirements are strict. Every claim is verified against a primary source. Drafts and version history are retained. A named human being signs off. Disclosure follows the standard of the receiving institution. The virtue of tiering is that it prevents integrity practices from collapsing under their own weight. If everything is treated as high risk, nothing will be, because the team will start ignoring the rules within a week. 7.3 The provenance habit The single most valuable habit a person can adopt is to keep the record. This costs almost nothing if done continuously and is impossible to reconstruct afterwards. Work in tools that preserve version history automatically. Keep the notes and the sources that were gathered before the writing began, since the sequence itself is informative. Keep the prompts and the raw outputs where they were used substantively. Keep the corrections, because the corrections are the evidence of judgement. The purpose of this record is not surveillance. It is protection. When a person is wrongly accused, the record is the only thing that will save them, and the moment of accusation is far too late to create it. 7.4 The verification gate A simple rule handles the majority of factual risk. No number, no citation, no quotation, and no comparative claim leaves the organisation until a human being has opened the source and confirmed it. This rule is boring, and it is the most valuable rule in the framework. Invented citations and confidently wrong figures are the characteristic failure mode of generative systems, and they are the failure mode most likely to destroy a venture's credibility with the people whose trust it most needs. #quality_control here is not bureaucracy. It is survival. 7.5 A one page policy A workable venture policy fits on a single page and answers six questions. Which tools may be used, and for what. Which categories of information may never be entered into an external system, covering personal data, confidential material, and unreleased product information, which protects #data_privacy and commercial secrecy. What must be verified before release. What must be disclosed, to whom, and in what form. Who signs off on each tier. What happens when the policy is broken. The last item matters more than teams expect. A policy with no consequence is an aspiration. But consequences should be proportionate and should distinguish sharply between an honest mistake that was disclosed and a deliberate deception that was concealed. The first should be met with correction. The second should be met with sanction. 7.6 Guidance for the individual student or founder Some rules are personal rather than organisational. Never submit anything you have not read. This sounds obvious and is violated constantly. Never assert anything you cannot explain. If a model produced an argument you do not understand, you have not made an argument, you have transported one. Never present machine output as evidence of your own ability when the purpose of the document is to demonstrate ability. This covers examinations, job applications, funding applications, and academic work. The deception here is not about words, it is about the self. Disclose when in doubt. The cost of unnecessary disclosure is small. The cost of concealment discovered later is not. Keep your drafts. It is the cheapest insurance available. Limitations This study has clear limits, and stating them plainly is part of the practice it recommends. The framework is conceptual and has not been tested empirically. Its propositions are plausible and grounded in the literature, but plausibility is not evidence, and the claim that disclosure increases perceived trustworthiness in investor settings, for example, is an empirical claim that has not yet been measured in that setting. The evidence base on detection ages quickly. Findings about detector performance reflect the models and tools available when the studies were conducted. A reader in three years should treat the specific accuracy figures in the underlying literature as historical rather than current, while the structural argument, which concerns the overlap between human and machine writing distributions, is likely to remain sound. The review is dominated by English language and Western institutional sources, which limits the generalisability of the findings to educational and commercial systems with different norms of authorship, collaboration, and text reuse. Finally, the framework assumes a degree of good faith and a degree of institutional autonomy that not every founder or student possesses. A person under severe economic pressure, or working inside an organisation that rewards deception, faces a harder problem than the one this article solves. Future Research Several lines of work follow directly from these limitations. Experimental work should test whether disclosure of machine assistance affects the judgements of investors, customers, and recruiters, and in which direction. The signalling argument in section six predicts a positive effect under conditions of general distrust and a negative effect where disclosure is rare and stigmatised, and this is a testable interaction. Longitudinal work should examine whether ventures with early governance practices experience fewer credibility failures during due diligence, and whether the effect survives controls for founder experience and sector. Comparative work should study how detection policies affect students and founders across linguistic and cultural groups, in order to quantify the equity costs identified in section five. Design work should develop and evaluate provenance systems that are lightweight enough for small teams, since the current tooling assumes either an enterprise budget or an unusual level of discipline. Finally, normative work should address the harder philosophical question that this article has treated pragmatically. If a person directs, reviews, corrects, and defends a piece of work that a machine drafted, in what sense is the work theirs, and does the answer change depending on whether the purpose of the work is to inform, to persuade, or to demonstrate capability. #future_research on this question will shape the norms that the next generation of founders inherit. Conclusion Generative systems have not created dishonesty. Dishonesty is old. What they have done is remove the friction that used to make dishonesty expensive, and in doing so they have exposed how much of our trust in documents rested on that friction rather than on anything more solid. The institutional reflex has been to look for a machine that can restore the lost certainty by identifying machine writing. The evidence assembled here suggests that this hope is misplaced. Detection is unreliable, it is biased in ways that harm people who are already disadvantaged, and it is trivially defeated by anyone who intends to defeat it. Built into policy as a source of proof, it produces injustice while failing to deter the behaviour it targets. The alternative is less dramatic and more durable. If the artefact can no longer be trusted, then the process must be made visible. Disclose what the machine did. Keep the record of how the work was made. Verify every claim before it leaves the building. Put a name on the result. These four practices cost little, they can be run by a team of three people with no budget, and they answer the question that detection cannot answer, which is not whether a machine was involved but whether a human being took responsibility. For entrepreneurs the argument is finally a commercial one as well as a moral one. In a world where anyone can produce a plausible document in a minute, the plausible document is worth nothing. What is worth something is a person who can be questioned, a claim that survives checking, and a record that shows the work was actually done. #responsible_innovation is not a constraint on building a valuable company. Under present conditions it is one of the few remaining ways to build one. The task, then, is not to catch the machine. It is to remain the kind of person, and to build the kind of organisation, whose word still means something when the words themselves have become free. Hashtags #entrepreneurial_ethics #AI_detection #integrity_in_business #originality #generative_AI #startup_governance #responsible_AI #academic_integrity #tech_ethics #founder_responsibility #AI_transparency #business_innovation #digital_trust #student_entrepreneurship #AI_and_society References Acemoglu, D., and Johnson, S. (2023). Power and progress: Our thousand year struggle over technology and prosperity. PublicAffairs. Bender, E. M., Gebru, T., McMillan-Major, A., and Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610-623. https://doi.org/10.1145/3442188.3445922 Bender, E. M., and Hanna, A. (2025). The AI con: How to fight big tech's hype and create the future we want. Harper. Casal, J. E., and Kessler, M. (2023). Can linguists distinguish between ChatGPT/AI and human writing? A study of research ethics and academic publishing. Research Methods in Applied Linguistics, 2(3), 100068. https://doi.org/10.1016/j.rmal.2023.100068 Chalmers, D., MacKenzie, N. G., and Carter, S. (2021). Artificial intelligence and entrepreneurship: Implications for venture creation in the fourth industrial revolution. Entrepreneurship Theory and Practice, 45(5), 1028-1053. https://doi.org/10.1177/1042258720934581 Chan, C. K. Y. (2023). A comprehensive AI policy education framework for university teaching and learning. International Journal of Educational Technology in Higher Education, 20, 38. https://doi.org/10.1186/s41239-023-00408-3 Coeckelbergh, M. (2020). AI ethics. MIT Press. Cotton, D. R. E., Cotton, P. A., and Shipway, J. R. (2024). Chatting and cheating: Ensuring academic integrity in the era of ChatGPT. Innovations in Education and Teaching International, 61(2), 228-239. https://doi.org/10.1080/14703297.2023.2190148 Du, S., and Xie, C. (2021). Paradoxes of artificial intelligence in consumer markets: Ethical challenges and opportunities. Journal of Business Research, 129, 961-974. https://doi.org/10.1016/j.jbusres.2020.08.024 Dwivedi, Y. K., Kshetri, N., Hughes, L., Slade, E. L., Jeyaraj, A., Kar, A. K., and others (2023). Opinion paper: So what if ChatGPT wrote it? Multidisciplinary perspectives on opportunities, challenges and implications of generative conversational AI for research, practice and policy. International Journal of Information Management, 71, 102642. https://doi.org/10.1016/j.ijinfomgt.2023.102642 Eke, D. O. (2023). ChatGPT and the rise of generative AI: Threat to academic integrity? Journal of Responsible Technology, 13, 100060. https://doi.org/10.1016/j.jrt.2023.100060 Elkhatat, A. M., Elsaid, K., and Almeer, S. (2023). Evaluating the efficacy of AI content detection tools in differentiating between human and AI-generated text. International Journal for Educational Integrity, 19, 17. https://doi.org/10.1007/s40979-023-00140-5 European Parliament and Council of the European Union (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union. Fleckenstein, J., Meyer, J., Jansen, T., Keller, S. D., Koller, O., and Moller, J. (2024). Do teachers spot AI? Evaluating the detectability of AI-generated texts among student essays. Computers and Education: Artificial Intelligence, 6, 100209. https://doi.org/10.1016/j.caeai.2024.100209 Floridi, L. (2023). AI as agency without intelligence: On ChatGPT, large language models, and other generative models. Philosophy and Technology, 36, 15. https://doi.org/10.1007/s13347-023-00621-y Fui-Hoon Nah, F., Zheng, R., Cai, J., Siau, K., and Chen, L. (2023). Generative AI and ChatGPT: Applications, challenges, and AI-human collaboration. Journal of Information Technology Case and Application Research, 25(3), 277-304. https://doi.org/10.1080/15228053.2023.2233814 Gao, C. A., Howard, F. M., Markov, N. S., Dyer, E. C., Ramesh, S., Luo, Y., and Pearson, A. T. (2023). Comparing scientific abstracts generated by ChatGPT to real abstracts with detectors and blinded human reviewers. npj Digital Medicine, 6, 75. https://doi.org/10.1038/s41746-023-00819-6 Giuggioli, G., and Pellegrini, M. M. (2023). Artificial intelligence as an enabler for entrepreneurs: A systematic literature review and an agenda for future research. International Journal of Entrepreneurial Behaviour and Research, 29(4), 816-837. https://doi.org/10.1108/IJEBR-05-2021-0426 Hagendorff, T. (2024). Mapping the ethics of generative AI: A comprehensive scoping review. Minds and Machines, 34, 39. https://doi.org/10.1007/s11023-024-09694-w Hermann, E. (2022). Leveraging artificial intelligence in marketing for social good: An ethical perspective. Journal of Business Ethics, 179, 43-61. https://doi.org/10.1007/s10551-021-04843-y Kasneci, E., Sessler, K., Kuchemann, S., Bannert, M., Dementieva, D., Fischer, F., and others (2023). ChatGPT for good? On opportunities and challenges of large language models for education. Learning and Individual Differences, 103, 102274. https://doi.org/10.1016/j.lindif.2023.102274 Kirchenbauer, J., Geiping, J., Wen, Y., Katz, J., Miers, I., and Goldstein, T. (2023). A watermark for large language models. Proceedings of the 40th International Conference on Machine Learning, PMLR 202, 17061-17084. Kshetri, N., Dwivedi, Y. K., Davenport, T. H., and Panteli, N. (2024). Generative artificial intelligence in marketing: Applications, opportunities, challenges, and research agenda. International Journal of Information Management, 75, 102716. https://doi.org/10.1016/j.ijinfomgt.2023.102716 Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., and Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779 Lund, B. D., Wang, T., Mannuru, N. R., Nie, B., Shimray, S., and Wang, Z. (2023). ChatGPT and a new academic reality: Artificial intelligence written research papers and the ethics of the large language models in scholarly publishing. Journal of the Association for Information Science and Technology, 74(5), 570-581. https://doi.org/10.1002/asi.24750 Mariani, M. M., Machado, I., and Nambisan, S. (2023). Types of innovation and artificial intelligence: A systematic quantitative literature review and research agenda. Journal of Business Research, 155, 113364. https://doi.org/10.1016/j.jbusres.2022.113364 Mitchell, E., Lee, Y., Khazatsky, A., Manning, C. D., and Finn, C. (2023). DetectGPT: Zero-shot machine-generated text detection using probability curvature. Proceedings of the 40th International Conference on Machine Learning, PMLR 202, 24950-24962. Mollick, E. (2024). Co-intelligence: Living and working with AI. Portfolio. Munoko, I., Brown-Liburd, H. L., and Vasarhelyi, M. (2020). The ethical implications of using artificial intelligence in auditing. Journal of Business Ethics, 167, 209-234. https://doi.org/10.1007/s10551-019-04407-1 Narayanan, A., and Kapoor, S. (2024). AI snake oil: What artificial intelligence can do, what it cannot, and how to tell the difference. Princeton University Press. National Institute of Standards and Technology (2023). Artificial intelligence risk management framework (AI RMF 1.0). NIST AI 100-1. U.S. Department of Commerce. Obschonka, M., and Audretsch, D. B. (2020). Artificial intelligence and big data in entrepreneurship: A new era has begun. Small Business Economics, 55, 529-539. https://doi.org/10.1007/s11187-019-00202-4 Perkins, M. (2023). Academic integrity considerations of AI large language models in the post-pandemic era: ChatGPT and beyond. Journal of University Teaching and Learning Practice, 20(2). https://doi.org/10.53761/1.20.02.07 Perkins, M., Roe, J., Postma, D., McGaughran, J., and Hickerson, D. (2024). Detection of GPT-4 generated text in higher education: Combining academic judgement and software to identify generative AI tool misuse. Journal of Academic Ethics, 22, 89-113. https://doi.org/10.1007/s10805-023-09492-6 Perkins, M., Roe, J., Vu, B. H., Postma, D., Hickerson, D., McGaughran, J., and Khuat, H. Q. (2024). Simple techniques to bypass GenAI text detectors: Implications for inclusive education. International Journal of Educational Technology in Higher Education, 21, 53. https://doi.org/10.1186/s41239-024-00487-w Ryan, M., and Stahl, B. C. (2021). Artificial intelligence ethics guidelines for developers and users: Clarifying their content and normative implications. Journal of Information, Communication and Ethics in Society, 19(1), 61-86. https://doi.org/10.1108/JICES-12-2019-0138 Stahl, B. C., and Eke, D. (2024). The ethics of ChatGPT: Exploring the ethical issues of an emerging technology. International Journal of Information Management, 74, 102700. https://doi.org/10.1016/j.ijinfomgt.2023.102700 Suleyman, M., and Bhaskar, M. (2023). The coming wave: Technology, power, and the twenty-first century's greatest dilemma. Crown. Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltynek, T., Guerrero-Dib, J., Popoola, O., Sigut, P., and Waddington, L. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19, 26. https://doi.org/10.1007/s40979-023-00146-z