The New Participant at the Executive Table

What Executives Need to Know About AI and Decision Authority

By Dr Danie Adendorff

For generations, executive decision-making was a human enterprise. Around the board table sat directors, executives, advisers, legal counsel, financial specialists, and subject-matter experts. Every voice represented human judgement, experience, authority, and accountability. Today, another participant has effectively taken a seat at that table. Artificial intelligence now analyses, recommends, predicts, prioritises, filters, and increasingly shapes the information on which executive decisions are based.

Yet this raises a fundamental governance question that few organisations have seriously examined: what authority has AI been given, what power does it actually exercise, and who remains accountable when its influence extends beyond advice into the architecture of decision-making itself?

This is not a question about whether AI is useful. It plainly is. Nor is it a question about whether AI is accurate, biased, or well-aligned — those are important but separate concerns, already the subject of extensive technical and regulatory attention. This is a question about authority: about who, or what, actually shapes the space within which a decision gets made, long before anyone in the room casts a formal vote.

Answering it requires borrowing from a body of knowledge that most executives have never formally studied, yet which increasingly governs how AI changes organisational decisions. That body of knowledge sits at the intersection of social choice theory, coalition decision-making, and a small but significant strand of recent research into how AI systems can erode collective human agency. None of it was written for the boardroom. All of it now applies there.

Why Executives Need Hidden Decision Science

AI governance is usually discussed in the language of technology management, compliance, or risk: model accuracy, data privacy, bias audits, regulatory exposure. That language is necessary, but it is incomplete. It treats AI as a tool to be controlled, rather than as a participant in a decision process to be understood.

The decision sciences ask a different question. They do not ask whether a tool works. They ask how groups of people — boards, committees, cabinets, executive teams, military staffs — actually convert competing views into a single collective decision. This is old territory in political science, economics, and organisational theory, but it has rarely been taught to executives, because until recently the machinery of group decision-making was almost entirely human. Now that AI systems sit inside that machinery — filtering data, ranking options, drafting recommendations — the old questions about how collective decisions form have acquired a new and urgent relevance.

Understanding this does not require executives to become mathematicians or political theorists. It requires understanding a handful of clear concepts well enough to ask sharper questions in the boardroom. That is the purpose of what follows.

Social Choice Theory: How Groups Turn Preferences into Decisions

Social choice theory is the study of how individual preferences are converted into a single collective decision. It emerged from economics and political science, largely to answer a deceptively simple question: if a group of people each has their own view, how should — and how does — the group arrive at one outcome?

The reason this matters to executives is straightforward. Very few consequential organisational decisions are made by one person acting alone. They are made by boards, investment committees, risk committees, procurement panels, or executive teams — bodies that must somehow aggregate a range of views, evidence, and preferences into a single approved course of action. The mechanics of how that aggregation happens are not neutral. Who speaks first, what information is presented, how options are framed, and in what order alternatives are considered can all shape the outcome, independent of the underlying merits.

This becomes directly relevant to AI governance once an AI system enters that aggregation process. If AI drafts the board paper, ranks the risk register, or produces the pre-read summary that frames the discussion, it is not merely supplying information. It is shaping how preferences form in the first place — before the group has even begun to deliberate. Social choice theory gives executives a vocabulary for noticing this: the outcome of a collective decision is never simply "what the evidence showed." It is also a function of how the group's preferences were elicited, ordered, and combined — and AI is now frequently doing some of that ordering.

Coalition Decision-Making: Who Really Controls the Outcome?

A closely related field studies coalition decision-making: how smaller groups within a larger decision-making body accumulate enough influence to determine the outcome, regardless of what the formal structure suggests.

In practice, many organisations discover that the formal decision-maker — the full board, the investment committee, the cabinet — is not always where the real decision is made. The decisive group may in fact be a smaller cluster: a chair and two influential directors, a CFO and a trusted external adviser, a small team that controls which figures make it into the papers. Political scientists call this the "decisive coalition" — the minimum group whose agreement is sufficient to determine the outcome, whatever the formal voting structure implies.

This concept becomes newly important once AI enters the picture, because AI systems, data pipelines, and recommendation engines can quietly become part of that decisive coalition without ever being named as members of it. If a procurement panel's decision is substantially shaped by an AI-generated vendor comparison that nobody on the panel independently re-derives, the panel's formal authority to decide has not changed — but the practical location of decision-making power may have shifted, invisibly, toward the system that produced the comparison. The question executives should be asking is not whether AI has a vote. It does not. The question is whether AI has become part of the small group whose inputs are, in practice, decisive — while carrying none of the formal accountability that comes with that role.

RAND's Warning: AI and the Erosion of Human Agency

This is not merely a theoretical concern. In April 2026, RAND's Global and Emerging Risks division published a research report by Alvin Moon and Benjamin Boudreaux, conducted through RAND's Center for the Geopolitics of Artificial General Intelligence, titled A Formal Model of How Artificial Intelligence Erodes Human Agency. It is worth being precise about what the report does and does not claim, because its contribution is specific and should not be overstated.

The authors draw directly on social choice theory to build a formal model of how AI can erode collective human agency over time, and to model decision-making in terms of coalitions — the same "decisive coalition" concept described above, now applied specifically to AI. Their central concern is that once human decision-making authority erodes past a certain point, the skills, institutions, and standing needed to reclaim it may simply no longer exist. The report does not argue that this outcome is inevitable, nor does it claim to measure how far any particular organisation currently is from that point — it proposes a framework and metrics for doing so, precisely because no such shared measurement previously existed.

The report identifies three distinct mechanisms through which this erosion can occur. The first is human disenfranchisement — simply, fewer humans occupying decision-making roles as AI takes on tasks previously performed by people. The second is AI enfranchisement — AI entities themselves gaining a form of decision-making power, changing the composition of the groups that actually determine outcomes. The third, and arguably the most immediately relevant to executive governance, is AI agenda control — AI systems shaping which alternatives ever reach human decision-makers in the first place, which can consolidate decision-making power in ways nobody explicitly authorised.

The authors go further and identify a theoretical terminal state within their formal model: a single minimal coalition that becomes decisive for every choice — the mathematical end-point of agency erosion, useful not as a prediction but as a reference point for measuring distance from irreversibility. Their recommendations are accordingly practical rather than alarmist: organisations should develop evaluations that measure structural effects on decision-making, not just AI capability or safety; consider minimum thresholds for human presence in decisive coalitions in high-stakes domains; monitor the composition of decisive coalitions over time; and assess whether human decision-making capacity could actually be restored if AI-driven agency loss accelerated.

None of this means RAND has proven that any particular organisation, industry, or governance framework is right or wrong. It means RAND has given the field something it previously lacked: a rigorous, published, peer-reviewed way of thinking about where decision authority actually sits, and a formal vocabulary — disenfranchisement, enfranchisement, agenda control, decisive coalition — for describing how it can move without anyone deciding that it should.

Agenda Control: When AI Controls the Menu

Of RAND's three mechanisms, agenda control deserves particular attention from executives, because it is the one most likely to be operating quietly inside their organisations right now, unnoticed precisely because nothing about it looks like a decision being taken away from anyone.

Agenda control means shaping which options are considered before the formal decision is made. Consider a simple, realistic case: an executive committee is presented with two strategic options, generated and pre-filtered by an AI-assisted planning process out of what was originally a wider set of possibilities. The committee deliberates carefully, weighs the two options on their merits, and votes freely. By every visible measure, this looks like sound, deliberate, human-led decision-making.

But the committee never saw the other options. They were not rejected by human judgement after consideration — they were simply never presented. The executives remain formally accountable for the decision they made. What they no longer fully control is the space within which that decision was made. This is the essence of the governance risk: agenda control does not require AI to argue, persuade, or override anyone. It only requires AI to narrow the menu before the human ever picks up the fork.

This same dynamic shows why poor-quality AI output — sometimes informally called "AI slop" — is dangerous not primarily as a content-quality problem, but as a governance problem. When defective, unverified, or shallow AI-generated material enters the decision pipeline and is not recognised as defective, it can quietly shape the summaries, assumptions, risk assessments, or option sets that later reach the committee. The damage is not that the content is poor. The damage is that poor content, once inside the pipeline, can perform agenda control just as effectively as good content — narrowing what the decision-makers ever get to see, without anyone flagging that the narrowing has happened.

From Technology Governance to Decision Architecture

Taken together, these ideas point to a reframing that most organisations have not yet made. AI governance is usually treated as a subset of technology governance: model risk, data governance, cyber-security, regulatory compliance. Each of those remains necessary. But none of them, on its own, asks the question this article has been building toward: who controls the architecture through which the organisation forms its options, gathers its evidence, and reaches its decisions?

Decision architecture is the connective structure joining authority, information flow, option formation, evidence validation, and accountability. Historically, that architecture was almost entirely human — built from committee structures, reporting lines, sign-off procedures, and professional judgement. AI now sits inside large parts of that architecture: drafting the papers, ranking the risks, summarising the evidence, sometimes even proposing the recommendation the committee will discuss. Governing AI well, in this light, is not chiefly about auditing a model. It is about deliberately governing the decision architecture the model now partly inhabits — making sure that authority, evidence, and accountability have not quietly migrated away from the people who are formally answerable for the outcome.

Governance Responses: Zero Trust, the Human Return Point, and Judgement

None of the diagnosis above is useful without a practical response. Three governance ideas, compatible with one another and with the analysis so far, offer executives a way to act on it. None of them is presented here as proven or validated by RAND's research — they are practical governance implications drawn from the wider synthesis, and their standing rests on their own logic and utility, not on borrowed authority.

Zero-trust AI governance is worth defining carefully, because the term is borrowed from cyber-security and can be misread. In network security, "zero trust" means no device or user is trusted by default, regardless of location, until verified. Applied to AI governance, the same logic applies to outputs rather than devices: trust is not granted to AI because it produced fluent, confident, or fast output. Trust is earned by evidence, constrained by governance, and withdrawn the moment verification fails. A system that produces an output and then declares that output correct is not governance — it is self-certifying risk, and no closed loop of that kind should be trusted to sign off on itself. A second, related discipline follows directly from recent research on AI sycophancy: agreement is an output pattern, not evidence of understanding or verification, and a system that consistently affirms the executive, the transformation team, or the preferred strategic direction must be treated with the same suspicion as one that is simply wrong. The point of zero-trust AI governance is not distrust of AI as a tool. As the doctrine states plainly: "AI may assist production. It must not certify completion."

The Human Return Point addresses a different, related question: not whether a human is vaguely "in the loop" somewhere in the process — a phrase that sounds reassuring but specifies nothing — but the precise point in an AI-assisted workflow where responsibility must return to a competent, authorised human before the output becomes a decision, action, exposure, or consequence. A genuine Human Return Point rests on four elements, not one. It has location: the organisation can name the exact point at which AI assistance stops and human authority must intervene. It has timing: the return happens before consequence, not after it — a post-incident review is not governance. It has authority: the person at that point must actually be able to say no, amend the output, demand further validation, or escalate, not merely comment on a decision already made. And it has competence: the person must understand enough about the task, the system, and the possible failure modes to recognise error, overreach, or misalignment when it appears. A human sitting silently in a meeting where AI-shaped options have already narrowed the agenda satisfies none of these four conditions, however "in the loop" it may look on paper.

The third idea is best captured in a single sentence from Logic Is Not Judgement: "logic can point to inconsistency; it remains our choice how to deal with it." Logic is a discipline of validation — it can expose contradiction, invalid inference, and claims that cannot all be true at once. What it cannot do is decide what should be sacrificed, revised, or defended once that contradiction is visible. That is not a logical operation; it is an exercise of judgement, which must weigh incomplete information, competing goods, and genuine accountability for the outcome — closer to what Aristotle called practical wisdom than to formal inference. The danger this creates in AI-assisted organisations is specific: it is possible to know that an inconsistency exists, document it carefully, and still fail to assign anyone the authority to resolve it. A validated warning that is never escalated into a decision is not a safety margin. It is a liability with a paper trail.

Conclusion: The Executive Test

The organisations that will govern AI well are not the ones that ask only whether their AI systems are accurate. Accuracy is necessary but is no longer sufficient. The deeper question this article has tried to open up is whether AI has changed who — or what — actually controls the decision, quietly and without anyone formally deciding that it should.

Part of the difficulty is that AI-generated material is easily mistaken for something it is not. Information is raw material — a summary, a ranked list, a flagged risk. Intelligence is different: information that has been checked, interpreted, placed in context, and connected to an actual decision. An AI system can produce a great deal of the former without ever producing the latter, and a board that treats a fluent AI summary as intelligence, rather than as an unvalidated signal awaiting interpretation, has quietly let the machine do part of the deciding.

This is also why timing matters as much as accuracy. Meaningful choices reduce as a situation develops — what the Adendorff Doctrine of Decision-Making calls option compression. Early in a problem, an organisation typically has many live options: it can investigate, prepare, contain, or delay at limited cost. Later, those same options may still exist on paper but are no longer practical, credible, or affordable. An AI system that quietly narrows the option set before anyone deliberates does not just risk a bad decision. It risks a well-reasoned decision made too late, inside a space that has already closed.

The Adendorff Doctrine offers a single governing question that draws the whole argument together: "did decision-relevant understanding reach accountable authority while meaningful options still remained?" Everything this article has examined — social choice theory, coalition decision-making, RAND's account of agenda control, the Human Return Point, zero-trust AI governance, the distinction between logic and judgement — is ultimately in service of that one test. It is not enough that the organisation eventually understood the problem. It is not enough that the board eventually saw the warning. Understanding and warning have value only if they reach the people with real authority to act while the options that matter are still open.

Turning that test into practice means executives should be able to answer, specifically and without hedging, before any consequential AI-shaped decision is finalised:

● Who shaped the options that were ultimately presented?

● Who filtered the evidence before it reached the room?

● Who ranked the alternatives, and on what basis?

● Who had genuine authority to challenge the recommendation — not merely the standing to ask a question, but the power to change the outcome?

● Where, precisely, was the Human Return Point, and was it exercised or merely present?

● What evidence was independently validated before the organisation committed to a course of action?

● Who remains accountable for the consequence, once it can no longer be reversed?

None of these questions requires an executive to distrust AI, reject its use, or slow the organisation to a crawl. They require something more disciplined: the recognition that decision-making before consequence — the moment before a course of action hardens into something that can no longer be undone — is precisely where authority must be protected, not assumed. AI governance, understood this way, is not primarily about controlling a tool. It is about governing the architecture through which an organisation forms its options, validates its evidence, builds its coalitions, exercises its judgement, and finally commits to consequences it must live with. That architecture now has a new participant sitting at the table. The task of governance is to make sure it never quietly takes the chair.

Author Workflow Disclosure

This article was produced through an AI-assisted but human-directed workflow. AI support was used for accessibility assistance, structuring, language refinement, source-discovery prompts, revision planning, and conversion of editorial comments into amendments. The author retained responsibility for the argument, accepted or rejected changes, checked the logic of claims, assessed source credibility, and remains accountable for the final text. AI-generated material was not treated as empirical evidence.

Sources and Notes

1. Moon, A. and Boudreaux, B. (2026) A Formal Model of How Artificial Intelligence Erodes Human Agency. RAND Corporation, RR-A4817-1. Santa Monica, CA: RAND. Available at: https://www.rand.org/pubs/research_reports/RRA4817-1.html (Accessed: 9 July 2026).

2. Adendorff, D. (2026) 'The Human Return Point: Managing Trust, Simplicity and Accountability in the Enterprise AI Transformation'. rumbls.com.

3. Adendorff, D. (2026) 'Logic Is Not Judgement'. rumbls.com.

4. Adendorff, D. (2026) 'Zero-Trust AI Governance for the C-Suite'. rumbls.com.

5. Adendorff, D. (2026) 'The Adendorff Doctrine of Decision-Making: Making Better Decisions Before Consequences Take Over'. rumbls.com.

contact@rumbls.com

Terms and Conditions

Privacy Policy

Cookie Policy

© 2026 WDA Publishing. All rights reserved.

Disclaimer

Contact .