Levers and Leverage: How Real People Shape AI Governance

A company announces an AI governance committee. There is a policy document, a risk register, a review workflow, a few tasteful slides, and perhaps a new Slack channel with “responsible” in the name. Everyone feels governed. Excellent. The bureaucracy has successfully acquired stationery.

The harder question is not whether the institution has an AI governance process. It is who can actually move it.

That is the useful question behind Levers of Power in the Field of AI, a short ethnographic paper on how decision makers in academia, government, business, and civil society experience their ability to influence AI-related institutional change.¹ The paper is not a statistical survey of AI elites. It is not a rankings table of who has the most power. It is not another sermon about “stakeholder engagement,” that most elastic of governance phrases.

Its value is more specific: it treats AI governance as an institutional field made of people, norms, events, labels, access routes, professional logics, and partially built infrastructure. In other words, it studies the machinery beneath the memo.

AI governance moves through mechanisms, not just mandates

The paper’s core framework is the idea of “levers of power”: the mechanisms through which individuals can shape how an institution responds to technological change. These levers range from formal instruments, such as governance structures and regulators, to softer mechanisms, such as norms, status markers, field events, relational channels, and the mobility of ideas across sectors.

That range matters. AI governance is often discussed as though the decisive action happens in law, compliance, procurement, or technical standards. Sometimes it does. But the paper’s institutionalist lens shows why that view is too tidy. Before a rule becomes formal, someone has to define the problem, name the risk, decide who counts as a relevant stakeholder, invite the right people into the room, translate a concern into institutional language, and make it legible to budget-holders. Glamorous? No. Decisive? Often.

This is the paper’s first useful correction: governance does not begin when the policy is signed. It begins when a problem becomes recognisable inside an institution.

A university may frame generative AI as academic integrity, student support, research productivity, labour substitution, or reputational risk. A company may frame the same technology as product acceleration, liability exposure, customer service automation, data leakage, or employee enablement. A regulator may frame it as consumer protection, industrial competitiveness, democratic accountability, or national security. Each frame activates different people and different tools. The same model, very different machinery.

That is why the paper’s emphasis on mechanisms is better than a persona-by-persona summary. The personas are evidence; the levers are the argument.

What the paper actually did

The study used personalised questionnaires for 12 high-level decision makers working across academia, government, business, and civil society. The paper says the questionnaires combined demographic questions, Likert-scale items, and open-ended prompts about respondents’ roles, institutions, influence, external interactions, and concerns about AI. The questionnaire was adapted from institutional governance concepts in social institutionalism and extended with “idea mobility,” a concept associated with Scandinavian institutionalism.

The responses were anonymised into persona-style portraits. The accessible paper text presents examples such as an AI policy professional in academia, a political scientist, a senator, a responsible AI architect, a senior religious leader, a university researcher, a data science executive, a public servant in AI policy, a responsible AI think tank director, and a scientific director of a technology institute. The personas are not decorative case studies. They are the paper’s main qualitative evidence.

The study then cross-references the questionnaire responses with a documentary review of respondents’ professional work. From that, it builds a table of the current status of different levers of power in AI governance and proposes five hypotheses for future institutional and social movement research.

There are no model benchmarks, experiments, ablations, or performance charts here. That should not be held against the paper; it is not that genre of research. But it does affect interpretation.

That distinction is important because this paper is easy to overread. Its best use is diagnostic, not predictive. It helps readers ask better questions about influence. It does not tell them, with statistical authority, how AI governance works everywhere. We have enough fake certainty in this sector already; no need to add artisanal sociology-flavoured certainty to the pile.

The lever map shows a field that is formal on paper and informal in practice

The paper’s synthesis table is the densest part of the work. It identifies the status of multiple levers in the AI institutional field: logics, institutional infrastructure, governance, collective interest organisations, regulators, informal governance bodies, field-configuring events, status differentiators, organisational templates, categories and labels, norms, relational channels, and idea mobility.

Several patterns matter for business readers.

First, formal governance is incomplete. The paper characterises governance coverage as incomplete and regulators as lacking full purview. That does not mean regulation is irrelevant. It means the field is still forming faster than its formal institutions can stabilise. Companies should not confuse regulatory silence with strategic safety. A risk can become institutionally important before it becomes legally compulsory.

Second, institutional infrastructure is described as emergent and translated. That phrase is awkward but useful. It means AI governance is not simply being installed from a central blueprint. It is being translated across sectors, roles, and local organisational constraints. A responsible AI framework in a global technology company does not mean the same thing as AI oversight in a government department, a university, or a religious community. The vocabulary may travel. The institutional function changes.

Third, informal mechanisms appear increasingly important. The paper’s first hypothesis is that formal methods are being outpaced, with growing leverage for informal methods. That is plausible given the field’s speed and fragmentation. When official rules lag, influence moves through relationships, conferences, working groups, professional networks, advisory roles, and elite access. This is governance by corridor, committee dinner, expert panel, and inbox.

Nobody prints that on the org chart. Perhaps they should, but then it would be less useful to the people using it.

Sector does not predict logic

One of the sharper observations in the paper is that a decision maker’s sector does not automatically reveal their operating logic. A business executive may reason from social good. An academic may reason from cost-effectiveness. A senator may draw heavily on entrepreneurial experience. A civil society leader may be focused less on abstract rights language and more on practical community continuity.

This matters because many AI governance strategies are built around sector stereotypes. Companies assume regulators think like regulators. NGOs assume companies think only in market terms. Academics assume business actors are allergic to complexity until proven otherwise. These assumptions are often directionally convenient and operationally lazy.

The paper suggests a better unit of analysis: not the sector, but the logic.

A logic is the internal rationale that makes an action feel legitimate to a person or institution. In the paper’s synthesis, AI governance is shaped by market logic, social justice logic, and technosaviourism. One could add others in practice: professional duty, reputational defence, national competitiveness, theological stewardship, academic integrity, public service neutrality.

For business, this changes stakeholder engagement from messaging to diagnosis. The question is not “How do we persuade government?” or “How do we manage civil society?” It is “Which logic is this person using, and what kind of evidence becomes legitimate inside that logic?”

A compliance team may need legal exposure. A product team may need customer friction data. A public servant may need procedural defensibility. A community organisation may need distributive impact. A board may need reputational and fiduciary framing. Same AI system, different persuasion grammar.

Access is not the same as accessibility

One of the paper’s most useful tensions concerns access. Many respondents appear, in principle, reachable. Yet the paper notes a contradiction: people may say stakeholders can contact them while remaining practically inaccessible to the public because of institutional position, knowledge barriers, social confidence, political constraints, or security rules.

The public servant persona is the cleanest example. The respondent works in AI policy, attends government meetings as an observer, and emphasises anonymity for ethical, political, safety, and security reasons. Yet this coexists with the idea that stakeholders can reach out. The paper interprets this as a kind of institutional decoupling: formal expectations of democratic openness collide with the protective anonymity and procedural constraints of bureaucracy.

This is not hypocrisy. It is institutional physics.

For companies, civil society groups, and AI vendors, the distinction is practical. A stakeholder map that lists names and offices is not a power map. It is a directory. A power map asks:

The real leverage is often in the middle layers: people who cannot officially decide but can translate, escalate, delay, legitimise, or quietly kill an initiative by making it administratively exhausting. Every organisation has these people. They rarely have the best titles. They often have the best institutional memory.

Field events are not networking fluff

The paper also treats field-configuring events as a lever. Conferences, workshops, policy meetings, expert gatherings, and cross-sector forums can shape what counts as a legitimate AI problem and who is seen as qualified to address it.

One persona describes a small workshop where participants read each other’s work before engaging. Another identifies a major international AI policy conference as important because it brought together scientific expertise, policy exchange, and diplomatic tension. These are not merely calendar items. They are sites where categories harden, alliances form, and ideas acquire institutional passports.

For business readers, this is easy to underestimate. Events are often treated as brand exposure or BD theatre. Some are exactly that, with lanyards. But in emerging fields, events can perform institutional work. They decide which phrases travel, which risks become fashionable, which policy templates become respectable, and which actors become “the people to call.”

That makes participation strategy more important than attendance strategy. Showing up is not the same as shaping the field. The latter requires knowing which logic the event rewards: technical authority, democratic legitimacy, ethical seriousness, venture-scale ambition, regulatory caution, or geopolitical urgency. If the room rewards one logic and the company speaks another, the company has not engaged. It has merely emitted.

Categories and labels are cheap, powerful, and dangerous

The paper marks categories and labels as having high potential impact. That is one of the most business-relevant points in the study.

AI governance runs on labels. “High-risk AI.” “Responsible AI.” “Human oversight.” “Frontier model.” “Open model.” “Safety.” “Alignment.” “Productivity tool.” “Decision support.” “Autonomous agent.” Each label changes the institutional treatment of the system. It can move a tool into compliance review, procurement scrutiny, board attention, union negotiation, academic misconduct policy, or public controversy.

Labels are powerful because they compress complexity into administrative action. They are dangerous because the compression can be wrong.

A company deploying AI in customer support may call it “automation.” Customers may call it “denial of service by chatbot.” Regulators may call it “consumer risk.” Employees may call it “headcount strategy.” Each label recruits a different coalition. The technical system did not change. Its institutional identity did.

This is where the paper’s institutionalist approach gives business leaders something more useful than moral theatre. The strategic question is not “What do we want to call this?” It is “Which labels will other actors plausibly attach to this, and what institutional machinery will those labels activate?”

That question belongs at the beginning of AI product planning, not after launch, when the communications team is asked to produce sincerity at speed.

The paper’s hypotheses are a research agenda, not a verdict

The study ends with five hypotheses: formal methods are being outpaced by informal methods; institutional dynamics are dominated by business marketing and community disengagement; big tech is strategically addressing formal and informal power mechanisms; informal influence remains elitist and siloed; and collective action is growing through large-scale initiatives.

These are not proven general laws. They are testable propositions inferred from a preliminary qualitative study. Their importance lies in the pattern they sketch: AI governance is not only a legal or technical contest. It is also a contest over attention, access, status, language, and coalition-building.

For business, the hypotheses should be read as prompts for institutional due diligence:

• Are our AI governance processes formal only where external auditors can see them?
• Which informal channels actually decide whether a system is slowed down, approved, reframed, or ignored?
• Are we mistaking elite consultation for community engagement?
• Are our labels imported from big tech marketing because they are convenient, or because they are analytically correct?
• Which external groups can organise faster than our internal policy cycle?

The last question is becoming less optional. The paper points to collective action as a growing force. Even if that claim needs broader testing, the business implication is already visible from the mechanism: when institutions fail to create credible channels for participation, pressure moves elsewhere. Sometimes to advocacy networks. Sometimes to media. Sometimes to litigation. Sometimes to internal employee resistance. Institutions may prefer neat consultation. The world, inconsiderately, does not always provide it.

How to use this paper without abusing it

The right business use of the paper is not to create a “levers of power dashboard” and pretend governance has been solved by a matrix. Please spare the interns.

A more serious use is to improve stakeholder mapping for AI decisions. Before deploying, buying, governing, or contesting an AI system, organisations can map five layers of influence:

1. Formal authority: Who signs off, regulates, audits, funds, or blocks?
2. Operational translation: Who turns broad policy into day-to-day practice?
3. Norm formation: Who decides what is acceptable, embarrassing, risky, or prestigious?
4. Label control: Who names the system, risk, affected group, or policy category?
5. Relational access: Who can reach whom informally, and who cannot?

This turns the paper’s academic contribution into a practical diagnostic. It also keeps the boundary clear. The study does not tell a company what its stakeholders believe. It tells the company where to look before assuming it knows.

That is already a meaningful advance. Many AI failures are not caused by the absence of principles. They are caused by organisations discovering too late that the people affected by the system were never inside the influence map. Governance then arrives as crisis response, wearing the slightly stunned expression of a department that has just learned society exists.

The boundary: this is rich evidence, not representative evidence

The paper is careful about its own limits. It does not seek statistical correlations or completeness. It focuses on individual context and experience. All respondents currently work in North America or Europe, even though they come from several cultural backgrounds. The authors explicitly frame the study as idiographic: designed to deepen understanding of specific lived experiences, not to generalise across all AI governance settings.

That limitation is not a footnote. It changes how the paper should be used.

A multinational company should not use this study to infer how AI governance power works in Southeast Asia, Africa, Latin America, China, or the Gulf. A regulator should not use it to claim that decision makers everywhere feel personally empowered. A vendor should not use it as evidence that informal engagement is always more effective than formal consultation.

The safer conclusion is narrower and stronger: in fast-moving AI governance environments, formal authority is only one part of institutional influence. Informal channels, labels, norms, field events, and personal logics can shape what institutions see, ignore, accelerate, or resist.

There is also an ethical boundary. Power maps can help communities gain access to institutions. They can also help incumbents manage, neutralise, or bypass opposition. The paper itself warns that power is not automatically benevolent. That warning deserves more than a polite nod. Any framework that improves influence can improve manipulation. Governance, inconveniently, has users too.

Conclusion: the policy is not the institution

The paper’s contribution is not that AI governance needs stakeholders. That sentence has been typed so many times it now qualifies as ambient noise. The contribution is more practical: it shows that stakeholders do not meet institutions as abstract entities. They meet people operating through roles, constraints, beliefs, professional norms, reputational incentives, informal routes, and institutional vocabulary.

For business leaders, this reframes AI governance from a compliance asset into an institutional capability. The question is not simply whether the company has a policy. The question is whether it understands how decisions actually move through the organisation and the wider field around it.

That means mapping not only authority, but leverage. Not only rules, but translation. Not only risk categories, but who benefits when a category sticks. Not only engagement, but who is structurally unable to enter the conversation despite being “welcome” in theory.

AI governance will not be shaped only by the people with the cleanest frameworks. It will be shaped by the people who can define the issue, convene the room, translate the logic, control the label, and keep the practice alive after the announcement has aged into another PDF.

The policy is the visible part. The institution is where the power moves.

Cognaptus: Automate the Present, Incubate the Future.