Chapter 002　Can AI Become a CEO?

Can AI become a CEO? The question is asked half as a joke and half in earnest. AI analyzes markets, lays out strategic options, and grades investment cases. It is faster than a human executive, it does not tire, and it does not swing with emotion. So why does the role need a human being in it? Vol. I, Ch. 001 defined what management means in the Age of AI. This chapter moves to the subject who carries out that management. How far has AI come as a subject of decision, and where does it stop?

1 The question — why it arises now

Ten years ago the question did not stand up. The AI of that time was an instrument that returned a specific prediction under restricted conditions. Substituting for a management judgment belonged to science fiction. The situation is different now. As of August 2026, AI can read most of the information a company runs on. Financial disclosures, contracts, minutes, customer feedback, patents, regulatory filings, and competitors’ releases. Information that a corporate planning team once spent weeks integrating, AI binds together in a day. It also arranges the result in a form a human can read. Analysis is not the whole of it. AI generates options. Exit, consolidation, outsourcing, continued investment for a given business. It lays out the assumptions behind each, the consequences to expect, and the counterarguments. In many cases it already exceeds the coverage of points a human director reaches after a night of thought. AI is also consistent. Human decisions wobble. Last night’s sleep, the conversation just before, the memory of a recent failure. Show the same material to the same person twice and the same conclusion is not guaranteed. AI does not carry that wobble. And AI has no interests. It has no department to protect. It has no term of office to keep in mind. It has no wish to leave good numbers for a successor. The largest single distorter of management judgment is structurally absent. AI is not omnipotent, of course. In territory where past data does not exist, its inference grows uncertain quickly. An unprecedented regulatory change, an unfamiliar geopolitical rupture, an emotional conflict inside the company. In these situations AI produces plausible prose, and the soundness of that prose is hard to verify. We expect this limit to persist for some time. Even so, most everyday management judgments sit in territory with precedent. The practical effect is therefore large. Set out this way, the answer looks obvious. Better analysis, better coverage, better consistency, and no self-interest. Where is the human executive superior? We want to stop here. The four items above are all capabilities for answering a given question better. Integrate information. Lay out options. Keep judgment steady. Keep private feeling out. These bear on the quality of judgment. But the work of a CEO has never been measured by the quality of judgment alone. If it were, the most accurate analyst would be the best CEO. That is not what happens. We have watched the same thing repeatedly: an excellent staff strategist stops functioning the moment they become CEO. The question, then, is not “is AI smarter than a human.” It is “is the role of CEO made of smartness?” Restating the question in its correct form is where this chapter starts.

2 Conventional answers and their limits

Three answers to “can AI become a CEO” circulate today. All three are partly right. All three mistake the structure of the question. The first answer: “It will, eventually. It is a matter of time” This position extends technical progress in a straight line. What was impossible five years ago is done today. So in five more years, the work of a CEO will be done too. The weakness lies in what is being extrapolated. What is advancing is the capability to cognize and to generate. Reading, writing, reasoning, and planning. These have indeed grown fast. But the office of CEO contains elements that are not cognitive capability. Being called on to explain at a shareholders’ meeting. Carrying responsibility for employees’ livelihoods. Standing as a legal subject before a regulator. Losing position, pay, and reputation when a judgment proves wrong. These are not capabilities. They are a standing. Standing is not acquired through better performance. It arises only when society recognizes a subject as one that bears responsibility. The linear extrapolation looks at the axis of capability and never at the axis of standing. The second answer: “Never. Management requires what is human” This is the opposite position. Empathy, intuition, passion, and the ability to draw people. Management is the practice of a human being who holds these, and AI cannot. The sentiment is understandable. As an argument it is weak. As Ch. 001 noted, the boundary of “what AI cannot do” moves every year. Writing with empathy, inferring what another person feels, choosing words that motivate. AI already performs all of these at a considerable level. Ground the case in human qualities and the ground keeps eroding under technical progress. Three years from now, the same claim may no longer be available. What we need is an argument that better performance cannot overturn. The third answer: “AI stays an adviser. Humans decide” This is the most moderate answer and the most widely accepted one. In practice it is also correct. But it holds a dangerous ambiguity. The word “decide” is never defined. Consider the case. AI narrows ten options to three and recommends one of them as best. A human director reads the recommendation and signs the approval. Who decided? Formally, the human. Substantively, the outcome was settled when the field was cut to three. AI designed the criterion for that cut. If no human has examined the criterion, no human has decided. A human has approved. The conventional answer “humans decide” does not carry this distinction. So a state spreads quietly through the field, in which humans have not decided and are recorded as having decided. We call this hollowed-out approval. The state does not surface until an accident occurs. In normal conditions it looks efficient and orderly. Meetings get shorter, disagreement drops, and sign-offs move faster. It is indistinguishable from the signs of healthy governance. That is exactly what makes it dangerous. The defect the three conventional answers share is the same. All of them use the phrase “become a CEO” without taking it apart. Agree or disagree without taking it apart, and the argument does not move.

3 Redefinition — breaking “becoming a CEO” into three

layers We rebuild the question from the standpoint of Future Value Theory. The role of CEO is not a single function. It is made of three layers. The first layer is decision — evaluating options and choosing one. The second layer is responsibility — carrying the consequence of that choice in one’s own person. The third layer is accountability — telling others why the choice was made, answering their questions, and obtaining their acceptance. Ordinary usage often folds all three into one word. They are entirely different things. Without separating them, we cannot measure how far AI has come. The three layers are not independent. They have an order. Without decision, no responsibility arises. Without responsibility, accountability does not hold. A subject cannot hold an upper layer while a lower layer is missing. That asymmetry is the skeleton of the argument below. Layer one, decision — AI is already inside Confined to the layer of decision, AI is already deeply inside. This is not a statement about the future. Whether to extend credit. Where to place inventory. When to reprice. How to split the advertising budget. How to staff a site. Which capital projects come first. In many companies these are effectively decided by algorithms. Human beings handle the exceptions. And the quality of AI’s judgment is passing human judgment across many domains. Where data is sufficient, the objective function is clear, and results are measurable, AI is better. We should concede the fact. Claiming human superiority at the layer of decision is therefore no longer tenable. Take the “deciding” part of a CEO’s work in isolation, and most of it can move to AI. To this point, the first conventional answer is right. Layer two, responsibility — the structure changes here The moment we enter the second layer, the argument changes. What is responsibility? We define it strictly. Responsibility is the state of carrying the consequence of one’s own choice as one’s own loss. If the business fails, the CEO loses position. Loses pay. Loses reputation. In some cases faces legal action. That state — having something to lose — is the substance of responsibility. AI has nothing to lose. This is not an absence of capability. It is an absence of structure. AI has no interests. We listed the absence of interests earlier as one of AI’s advantages. The same property becomes, here, a decisive disqualification. Having no interests, its judgment does not bend. Having no interests, it cannot bear responsibility. These are the two faces of a single fact. Here is a wall that better performance cannot clear. However capable it becomes, AI does not become a subject that incurs loss. A subject that incurs no loss cannot take on responsibility. A subject that cannot take on responsibility is not a CEO. This is the central argument of the chapter. Objections are possible. Give the AI assets. Grant the AI legal personality. Technically we can imagine both. But in that case the party bearing responsibility is the human being or the legal person that gave the AI its assets and set up its legal personality. Responsibility has not transferred. It has only been mediated. Trace responsibility back and it always lands on a human being. Responsibility that does not land is not responsibility. Layer three, accountability — whether a thing can be said The third layer is more human still. Accountability is not simply stating reasons. Shareholders, employees, customers, local communities, regulators. It is telling the same choice to parties with different interests, each in their own context. And continuing the conversation with the party that is not persuaded. AI can generate an explanation. It can generate a very skillful one. But generating an explanation and holding accountability are different things. The difference lies in what stands behind the words. When a human CEO says “we are betting on this business,” the sentence has that person’s position posted as collateral. If it fails, the person who said it pays. That is where the weight of the words comes from. When AI outputs the same sentence, nothing stands behind it. Words without collateral, however logical, have no power to move an organization. First Principle 8 states it. Trust Compounds Faster Than Capital. Trust compounds faster than capital and becomes the last durable advantage. But trust accumulates only toward a subject capable of paying a price. Split into three layers, the answer is clean. AI enters layer one. It cannot structurally enter layer two. It cannot enter layer three without layer two. AI therefore cannot become a CEO. And AI will carry a substantial part of a CEO’s work. The two statements do not conflict. First Principle 4 puts it in one line. AI Optimizes. Humans Define. AI optimizes; humans define value, purpose, and direction. To define is to select a meaning and to take on the price of that selection.

4 Structure — testing it against the Leadership Formula

We check the three-layer breakdown against the equation. Future Value Theory defines management itself as follows. Leadership = Purpose × Question Design × Capital Allocation × System Architecture × Trust The relationship is multiplicative. It is not additive. If one term goes to zero, the whole is zero however large the rest are. So we test the five terms one at a time, asking whether AI can carry each. Purpose. AI can analyze a purpose. It can assess feasibility. It can refine the expression. But it cannot choose which societal challenge to take on. Choosing carries the resolve to discard what was not chosen. AI has nothing to discard. Question Design. AI answers questions. It answers them well. Which question should be asked is not something AI can settle. Question Design is a declaration of what counts as important. AI can compute importance. It cannot declare importance. Capital Allocation. AI optimizes allocation. It computes expected values, evaluates variance, and presents efficient combinations. But optimization requires an objective function. The objective function is supplied by a human being. The substance of the management judgment is finished at the moment someone decides what to maximize. System Architecture. AI can produce design proposals. Organizational structure, workflow, the division of roles between people and AI. But deciding to adopt a design means accepting that some people will lose their jobs to it. Drafting a design and deciding on a design are different acts. Trust. The clearest of the five. Trust holds only toward a subject capable of betraying it. The concept does not apply to an entity that cannot betray. AI is not trusted. AI is verified. The same structure appears at all five terms. AI can carry the work in each term. It cannot carry the decision in each term. And if the decision part is zero, the product being multiplicative, the whole of Leadership is zero. This is not a case where one term out of five would have been enough. The multiplicative structure has the reverse implication as well. A human may be deciding in four terms; hand one term over to AI in its entirety, and that enterprise’s Leadership is zero. Partial abdication produces total loss. That is the practical meaning of reading the equation as a product. This is the structural explanation for why AI cannot become a CEO. Human-on-the-Loop Management Where, then, does the human stand? Human-on-the-Loop Management is the answer. Place Human-in-the-Loop beside it for contrast. The human sits inside the loop, checking each of AI’s outputs, approving it, and correcting it. Human-on-the-Loop is different. The human sits above the loop. The human does not chase individual outputs. The human designs the whole system. What is entrusted to AI, what a human decides, under what conditions a case returns to a human, and by what measures the whole is evaluated. That design is the human’s work. Do not mistake this point. Human-on-the-Loop is not a doctrine of supervising AI. It is a doctrine of designing the whole system. Why must it be design? The reason is simple. AI’s throughput keeps rising. Management that checks one item at a time breaks down as throughput rises. Management that designs grows stronger as throughput rises. And a design always has the location of responsibility written into it. In whose name is this judgment made? Who explains it when it fails? Only enterprises that have embedded this in the design can use AI at scale and keep governance intact. First Principle 9 states it. Leadership Means Designing the Future. Leadership means designing the future — the right questions and systems rather than the right answers. To design is also to choose the courage not to intervene case by case.

5 What it looks like in practice — how AI enters the

boardroom AI does not become the CEO. It enters the center of management with certainty, and has begun to. We set out four realistic forms. First, as the generator of the agenda. In many companies the board agenda is settled by inertia. The same items as last time, with reports taking most of the space. AI can extract, from information inside and outside the company, the points that are important and are not being discussed. Combing out the assumptions that are going stale is precisely the first stage of Enterprise Redefinition: Recognize. The stage asks, “What assumptions about our enterprise are becoming obsolete?” In the role of confronting a board with the point it overlooked, AI is useful. Second, as the dissenter. The largest pathology of a board is the absence of dissent. Internal power dynamics, personal relationships, and awareness of one’s term suppress candid objection. None of that suppression acts on AI. It can present the strongest counter-evidence against a proposal without hesitation. We call this institutionalized dissent. Third, as auditor of the record and of consistency. How did we judge this point at a past board meeting? Which of the assumptions we held then have since collapsed? Human memory edits itself conveniently. Successful judgments are remembered as one’s own achievement, failed ones as the fault of the environment. AI does not edit. It can confront a board with the contradiction between its past judgments and its present one. This lifts the learning capability of the organization directly. As First Principle 5 says, learning is the ultimate competitive advantage, because knowledge and technology depreciate. Fourth, as a symbolic “seat.” In 2014 a Hong Kong investment firm announced that it had added an algorithm supporting investment judgment to its board. The announcement was widely reported at the time. What can be read from public information is that this was not a legal directorship. It was closer to a declaration that algorithmic assessment would be built into decision-making. What the example shows is not a milestone of capability. It is a problem in the use of words. The moment we say “an AI became a director,” the location of responsibility goes vague. Responsibility that has gone vague becomes responsibility nobody takes on. So we should keep our language strict. AI does not participate in a board. AI is an instrument a board uses. When the distinction between a participant and an instrument disappears, governance becomes a shell. Hollowed-out approval, in the field There is one more failure, and it is the one most likely to occur in practice. AI produces the analysis, narrows the options, and presents a recommendation. The board reads it, discusses it for twenty minutes, and approves. The minutes record unanimity. The form is perfect. But nobody has examined the assumptions behind the recommendation. Which constraints were imposed? Which options were removed at the first stage? What went into the objective function, and what did not? If no one in the room can ask these questions, the approval is hollow. This is where Question Design does its work. The capability to verify an answer AI produced is not the capability to read the answer. It is the capability to read how the question was framed. Evaluate the design of the input rather than the output. This is the most concrete skill required of a director in the Age of AI. And AI does not substitute for it, because the act is an interrogation of AI’s own design. The skill is trainable. The difficulty is that no place to train it has been set up. In most companies, directors sit on the receiving side of AI’s output and never on the side that gives AI its questions. A person trained only to receive cannot make the judgments involved in giving. This is the quietest spreading blank in management today. How the executive’s work is redistributed Taking all of this together, the office of CEO does not disappear. Its composition changes. Analysis and the generation of options move to AI. Approval or refusal on individual cases largely moves to AI as well. Human time is released from there. The released time goes to three places. Re-examining Purpose. Designing the questions. And writing the location of responsibility into the design. This is not a shrinking of the office. It is a shift in the center of gravity. The volume of operational work falls, and the density of decision rises. The more AI advances, the heavier and the less escapable the work left to a human becomes. The shift does not happen automatically. Deploy AI, and if the executive meeting’s agenda is the same as last year’s, the center of gravity has not moved. The freed time is absorbed into more reports and finer confirmation. What causes the shift is not technology. It is the executive’s own decision.

6 Questions for the executive

The argument, in one line. AI cannot become a CEO, because it cannot take on responsibility. But AI will carry much of a CEO’s work. So the executive must move from competing with AI to designing the decision system that includes AI. We connect the conclusion to tomorrow’s action. Four questions. Each can be started at your next executive meeting. Question 1 — Among your important decisions, which are substantively being made by AI? Take the inventory first. List the judgments that a human approves in form and an algorithm makes in substance. In most companies the list is longer than expected. Length is not itself the problem. Not knowing is the problem. Question 2 — For each one, who explains it when it fails? Write the location of responsibility as a person’s name. Not a department name — an individual name. Any judgment for which the name cannot be written is a blank in governance. A blank is invisible in normal conditions. It appears only in a crisis, and then it is fatal. This exercise has a secondary effect. The moment someone tries to write the name, the question of whether the judgment should be entrusted to AI at all gets serious consideration for the first time. Naming the responsible party is a governance procedure. It is also a design act that fixes the scope of delegation. Question 3 — At your last executive meeting, did anyone question an assumption AI had supplied? Not a remark about the answer. A remark that doubted an assumption. The objective function, the constraints, the excluded options. If the count of remarks cutting into these was zero, that meeting was an approving body, not a deciding body. Question 4 — Has the time allocation of the executive meeting changed since last year? If AI can now produce the reporting material, the time spent on reporting should have fallen. If it has not, AI was deployed and management did not change. What the freed time is spent on is the use of Future Time, and it separates one company’s future from another’s. None of the four questions asks about AI’s performance. All four ask about the design of responsibility. Can AI become a CEO? The answer is no. Not because AI is inferior. Because AI has nothing to lose. Management is the practice of a subject that has something to lose choosing a future anyway. Choice carries a price. Only the one who takes on the price is qualified to point out a direction to others. The smarter AI becomes, the less this structure fades. It sharpens. As the quality of judgment converges, the only question left is who bears responsibility for the judgment. The executive in the Age of AI is not the smartest person. It is the person who takes on the most. And as long as there are people willing to take it on, an enterprise can keep creating Future Value.

In brief

AI cannot become a CEO not because of a shortfall in capability, but because of an absence of structure: it has nothing to lose.
The office of CEO is made of three layers — decision, responsibility, and accountability. AI can enter only the first.
The executive must move from competing with AI to designing the decision system that includes AI.
Listing the judgments AI substantively makes, and naming the responsible person for each, is where governance begins.

Key concepts

Human-on-the-Loop Management / Human-in-the-Loop / Leadership Formula / Future Time / Future Value

The chain of ideas

AI Integration → Human-on-the-Loop Management → Trust → Leadership → Future Value

Related first principles

Principle 4 — AI Optimizes. Humans Define. Principle 8 — Trust Compounds Faster Than Capital. Principle 9 — Leadership Means Designing the Future.

Related chapters

Vol. I, Ch. 005 “What Is Decision-Making in the Age of AI?” — the five layers of decision and the boundary of responsibility in detail
Vol. I, Ch. 009 “How Should Work Be Divided Between AI and People?” — how to design the division from the location of responsibility
Vol. II, Ch. 012 “What Is Leadership in the Age of AI?” — how the sources of authority and trust move
Vol. VI, Ch. 052 “What Does It Mean to Redefine the Executive?” — rewriting the office of the executive itself

Papers and companion volumes

Kadowaki, N. (2026a). Future Value Theory: A Management Framework for Enterprise, Capital, and Society in the Age of AI. VURA Working Paper Series. SSRN: https://ssrn.com/abstract=7120980 / Zenodo: https://doi.org/10.5281/zenodo. 21255662
Kadowaki, N. (2026b). Enterprise Redefinition: Toward an Enterprise Evolution Theory for the Age of AI. VURA Working Paper Series. (Published on Zenodo; under review at SSRN)
100 Questions on Management in the Age of AI, #011 “Can AI Become a CEO?” / #013 “Should We Hand AI the Management Judgment?”

Chapter 002 Can AI Become a CEO?