Chapter 005　What Is Decision-Making in the Age of AI?

What is decision-making? Asked to name the executive’s job in a single word, most people choose this one. Yet the content of the word has been left remarkably vague. AI gathers the information. It lays out the options. It attaches probabilities to each. So how much of decision-making does AI carry, and where does the human part begin? In this chapter we break the act of deciding into five layers. Then we redraw its structure for the Age of AI.

1 The question — why it arises now

Decision-making sits at the center of management vocabulary. It is also the least analyzed word in it. We say every day that a decision was good, or that a judgment was wrong. We can rarely say afterward at which stage the quality was settled. Was the information thin? Was the option set narrow? Was the axis of evaluation wrong? Was the decision late? In most companies that separation is never made. Not making it used to cost nothing. Every stage of a decision was carried by the same person. Gathering the information, generating the options, deciding, and owning the result all belonged to the executive. There was no need to decompose the act, because decomposing it changed nothing about who performed it. AI broke that assumption. Information gathering now runs faster and wider through AI than through us. Option generation is broader through AI as well, and less biased. Parts of evaluation, quantitative comparison above all, are more accurate through AI. Yet the act of deciding, and the act of carrying the consequence, remain on the human side. An activity that was once single has split into several activities with different owners. Treat a split thing as though it were still whole, and management falls into confusion. That confusion is what we see in practice. AI produces a refined analysis. The meeting pack gets thicker. Every point is covered. Still nothing is decided. Or something is decided and nobody moves. Only the first half of decision-making got faster. The second half — choice and responsibility — did not speed up at all. If anything, the richer first half makes the weight of the second half more visible. So the question to ask now is this. What kind of structure does the act of deciding actually have? And which parts of that structure can be mechanized, and which cannot. Introducing AI into decision-making without answering that is like knocking out a wall without looking at the drawings.

2 Conventional answers and their limits

Three answers dominate practice. Each is partly right. Each stops working in the Age of AI. The first answer: “Decision-making is choosing the best available option” This is the most widely shared understanding. List the options, evaluate them against a criterion, and take the top one. Management textbooks rest on this frame, and so does much of the research on decisions. Its weakness is that the option set is treated as given. The frame only describes what happens after somebody has prepared the options. Yet the quality of management is often settled at the stage where options are born, not at the stage where one is picked. Picture a meeting agonizing between A and B. The real problem is that C is not on the agenda. An option that is not on the agenda is never chosen, however precisely the others are evaluated. Skill in choosing cannot exceed the quality of the choices. AI fills part of this gap. It enumerates more options than we do. But AI enumerates inside the world it has learned. An option that moves the frame itself enters the set only when a person decides that such a future ought to exist. The second answer: “Decision-making is judgment based on data” The second answer travels under the word data-driven. Decide by data, not by instinct and experience. The claim held for two decades. It still partly holds. And yet we face a strange fact. Companies with plenty of data are not making fewer bad judgments. In some settings, the errors become more conspicuous as the data grows. There are four reasons. First, all data is a record of the past. Where the future extends the past, data is powerful. Where the assumptions change, data points confidently in the wrong direction. A market’s indicators look healthy right up to the moment the market disappears. Second, if the question is wrong, data will confirm the error. Data accumulates around “is there demand for this product?” There is no data for “will this product category exist in ten years?” A question with no data behind it never becomes an object of analysis. Third, the measurable drives out the unmeasurable. Revenue, cost, and utilization can be measured. The accumulation of trust, the speed of learning, and the capability to redefine the enterprise are hard to measure. Meetings proceed on the measurable, and the unmeasurable disappears quietly from the discussion. Fourth, data disperses responsibility. “The data indicated it” blurs the outline of the person who decided. A decision with no clear location of responsibility produces no learning afterward. When an error belongs to nobody, nobody learns. Missing data, then, is only part of the problem. Most errors arise when data is present and question design and responsibility design are absent. The third answer: “Decision-making is speed” The third answer is the speed argument. Change is fast, so decide fast. This is also right. Slow decisions lose opportunities, and the proof arrives daily. But speed contains no direction. Deciding fast means moving fast inside the information currently visible. That information is visible to competitors too. Read by AI, it is even more nearly identical across firms. From here many executives fall into a false alternative. Decide fast, or think deeply. The received view that these two trade off against each other is wrong. We demonstrate that later in the chapter. The three conventional answers share an assumption. All three treat a decision as a reaction to a given situation. There is a situation, there are options, and one selects among them. Management in the Age of AI asks the question one level above. Who defines the situation itself, over what span of time, and how?

3 Redefinition — decision-making is choosing a future

and owning the result Future Value Theory defines decision-making as follows. A decision is the act of selecting one among several possible futures, allocating capital toward it, and carrying the result. Three elements sit inside that definition. The selection of a future, the allocation of capital, and the assumption of responsibility. All three lie either before or beneath the act of comparing options. The five layers of a decision To bring the definition into practice we break a decision into five layers. This decomposition is the core of the chapter. Layer 1, information gathering. Establish what is happening. Market, customers, technology, competitors, regulation, and the internal reality. Layer 2, option generation. Lay out the available roads. What to do, what not to do, when to start. Layer 3, evaluation. Estimate the consequence of each option. Effect, cost, probability, side effects, time frame. Layer 4, choice. Take one. Discard the rest. Layer 5, responsibility. Carry the result. Explain it. Admit the error and connect it to learning. These five layers have always been handled as a single mass. In the Age of AI the owner changes layer by layer. Layer 1 can be carried by AI almost entirely. The volume of information a person can gather is no longer comparable. Spending human time here is, as of August 2026, a misallocation of resources. Layer 2 can be carried mostly by AI. There is a limit. The options AI generates lie inside the world it has learned. An option that answers a future nobody has yet described appears only when a person decides to bring that future into existence. Layer 3 can be carried by AI conditionally. AI is better at estimating quantitative consequences. But evaluation always requires an axis of evaluation. What counts as good cannot be derived from data. A person sets the axis. Layer 4 cannot be carried by AI. Choice means discarding. AI can maximize an expected value. Computing an expected value requires an objective function, and setting the objective function is itself the choice. The circularity here is logical, not technical. Layer 5 cannot be carried by AI structurally. Responsibility means there is a subject whose position is at stake when things go wrong. AI has no position to stake. A subject that cannot bear responsibility has not decided anything. First Principle 4 states the boundary in one line. AI Optimizes. Humans Define. AI optimizes; humans define value, purpose, and direction. Layers 1 through 3 are the domain of optimization. Layers 4 and 5 are the domain of definition. The boundary does not move The observation that “what AI cannot do moves every year” is correct. The boundary just described does not move. AI cannot carry Layers 4 and 5 for a reason that has nothing to do with performance. Deciding a purpose is not a question of capability. It is a question of meaning and responsibility. A more capable model will arrive. The right to decide what things mean does not transfer with it. The executive’s role is therefore not defined by subtraction. It is not the remainder left after AI’s share. It sits at the center of the two deepest layers of the act of deciding. Choice is capital allocation In practice, the choice at Layer 4 always appears as capital allocation. Budget, headcount, time, and executive attention. Where these are directed is the substance of the choice. First Principle 3 fixes what that means. Capital Exists to Create Possibility. Capital exists to create possibility, not merely to maximize return. Adopt that principle and the capital allocation question changes. It is no longer “where does the highest return lie?” It becomes “which possibility do we want to bring into existence?” The first is a question of optimization, and AI can handle it. The second is a question of definition, and only people can handle it. A budget table is the record of which possibilities a company selected. Put the other way around: a budget table that adds no new possibility is evidence that no decision was made. Human-on-the-Loop Management The five-layer decomposition, restated as a form of management, is Human-on-the-Loop Management. The familiar arrangement is Human-in-the-Loop, in which humans supervise AI from inside the operational loop. AI produces output, a person checks it one item at a time, and approves or corrects. The quality of each judgment is preserved. But as AI’s throughput rises, the person becomes the rate-limiting step and the whole flow jams. Human-on-the-Loop is different. Human beings sit above the system. They do not intervene in individual outputs. They design people, AI, capital, the organization, and society as one system. What goes into the objective function. Which axis of evaluation is adopted. How far AI is trusted to run. Under what conditions a person takes the matter back. These are designed in advance, and the design itself is continuously updated. This point must not be mistaken. Human-on-the-Loop is not a mechanism for supervising AI. It is a management practice in which people, AI, capital, the organization, and society are designed as a single system. The objective is better design rather than better control. In terms of the five layers: design the operating rules for Layers 1 through 3, and keep Layers 4 and 5 with people. In an organization where that separation is explicit, management gets stronger as AI’s throughput grows. In an organization where it is vague, management gets weaker as AI’s throughput grows.

4 Structure — the Future Horizon decides the quality of

decisions We now give the redefinition a frame. Two concepts and one equation carry it.

4.1 Future Horizon

Future Horizon is the distance over which an organization plans — the span of future actually taken into account in the decisions an executive is making today. In most companies that span is short. The medium-term plan runs three years, the investment payback standard runs five, and executive tenure is shorter still. When institutions enforce a short horizon, lengthening individual awareness changes nothing. In a company with a short Future Horizon, every decision becomes an extension of the present. If only proposals that pay back in three years can pass, only futures that pay back in three years can be selected. The option set has been cut down in advance, by the institution. Extending the Future Horizon is not a matter of becoming leisurely. It is widening, as an institution, the span of time over which things are evaluated.

4.2 The Future Time Equation

The fifth equation of Future Value Theory fixes this structure. Future Value = Future Time × Future Capability Future Value is the product of Future Time and Future Capability. This is not addition. It is multiplication. If either term is zero, the whole is zero. No term compensates for another. Future Time is time intentionally invested in creating the future. The executive’s thinking time, the organization’s learning time, the time permitted for experiments, and the Future Horizon itself. Future Capability is the capability that converts that time into value. The equation explains two kinds of failure at once. The first failure is the company with capability and no time. It holds excellent engineers and adequate funds, and every hour of management goes to processing what is immediately in front of it. Future Time is near zero, so the product is near zero. The second failure is the company with time and no capability. It speaks of long-term visions and paints a distant future. No learning and no redefinition take place. Future Capability is near zero, so again the product is near zero. AI raises Future Time directly, because it cuts the working hours in Layers 1 through 3. But released hours do not become Future Time automatically. In a company that fills the freed time with more meetings, nothing changes. What the company does with the time AI returned is itself the first decision of the Age of AI.

4.3 Speed and horizon are not a trade-off

Here we answer the point left open by the third conventional answer. Can a company decide fast and see far at the same time? It can. There are three reasons. First, the two belong to different layers of the decision. Speed is at issue in Layers 1 through 3. Horizon is at issue in Layers 4 and 5. Things that live in different layers cannot stand in exchange with one another. A slow company is not seeing further. Its first half is simply slow. Second, the longer the horizon, the faster each individual decision becomes. When the shape of the company twenty years out is settled, any proposal can be judged immediately by whether it moves toward that shape. A company with a short horizon reopens the argument about purpose for every proposal. That is why it is slow. The longer the service life of a criterion, the faster the judgment. Third, because the Future Time Equation is multiplicative, time and capability amplify each other. A company that secures time learns more. A company that learns more judges faster. A company that judges faster frees more time. Speed and horizon are two faces of one cycle. They look like a trade-off because the work of extending the horizon and the work of handling the immediate are made to compete for the same hours. That is a problem of structure, and structure can be designed.

4.4 The procedure of Future Back Planning

The method that converts the Future Horizon into practice is Future Back Planning. Rather than building up from the present, it designs from the future back toward the present. The procedure has six stages. Stage one: describe the future society. What society should exist twenty years from now? This is not a forecast. It is the selection of a future that ought to exist. Population, energy, health care, education, cities. Do not start from your own businesses. Start from society. Stage two: fix the enterprise’s role in that society. When that future has arrived, what role is this enterprise playing? Describe it by the function performed, not by the goods sold. This is where Enterprise Redefinition begins (→ Vol. V, Ch. 041). Stage three: enumerate the capabilities that shape requires. What must a company be able to do in order to play that role? Technology, people, data, trust, and partners. Write them all out as components of Future Capital. Stage four: measure the gap against the present. Of the capabilities enumerated, which do we hold today? Which do we not hold? Which can be bought? Which cannot be bought and must be grown? How many years does growing one take? Stage five: convert the gap into a calendar. A capability that takes ten years to grow had to be started ten years ago. So start it now. This backward calculation is the core of Future Back Planning. The capabilities needed in twenty years determine today’s work uniquely. Stage six: bring it down to today’s capital allocation. How will next year’s budget, headcount, and executive attention be allocated to close the gap? Only when a plan has come down this far has it become a decision. The pivot of the procedure is stage five. Long-term visions end as pictures because no bridge has been built between the vision and the budget, and that bridge is the number of years a capability takes to grow. Without the bridge, the vision stays a thing to be done someday. AI works powerfully in stages three and four. Enumerating required capabilities and measuring gaps both belong to the domain of optimization. Stages one and two — which future to select — are decided by people.

4.5 Management that designs decision-making

We connect the above to the equation of leadership. Leadership = Purpose × Question Design × Capital Allocation × System Architecture × Trust The five layers map onto this equation cleanly. Purpose supplies the axis of evaluation. Question Design sets the width of the options at Layer 2. Capital Allocation gives Layer 4’s choice its substance. System Architecture carries the design of Layers 1 through 3. And Trust is the ground on which the organization accepts the responsibility of Layer 5. This equation is multiplicative as well. If Question Design is zero, the organization cannot test the answers AI produces. If Trust is zero, a decision that admits an error becomes impossible inside the organization. No strength in the remaining terms recovers either loss.

5 What it looks like in practice — why companies cannot

stop We now lay the theory over the working life of a company. What happens in a meeting where the data is complete Picture a meeting. It exists to decide whether a new business continues. The materials are thorough. Market size estimates, competitor movements, an analysis of customer interviews, three scenarios with probabilities. What AI assembled far exceeds the human output of a few years ago. Still no conclusion arrives. Or the conclusion is “let us watch it one more year.” The materials answer only up to Layer 3. The choice at Layer 4 requires an axis of evaluation. What was this business started for? Is that purpose consistent with the future the company is now selecting? In a meeting where this has never been put into words, no thickness of material makes a choice possible. Then there is Layer 5. Concluding that the business should stop means admitting that the decision to start it was wrong. In an organization where nobody will admit that, the decision to stop cannot structurally be produced. This is not a shortage of data. It is the absence of a responsibility design. Why exit becomes harder in the Age of AI Exit and cancellation were always the hardest decisions. In the Age of AI they get harder. There are four reasons. First, the improvement proposals never run out. Exit used to be decided when there was nothing left to try. AI now supplies the next improvement continuously. Change the price. Change the pitch. Change the customer segment. Because the moves do not run out, the natural trigger for exit has disappeared. Second, a company can now keep climbing a local optimum. AI improves continuously inside the objective function it was given. The business gets slightly better every month. A decision to stop a business that is getting better is extremely hard to produce. But AI does not point out that the hill being climbed may be a low hill. The question of changing hills lies outside the objective function. Third, the point of comparison is hard to see. The essence of exit is moving capital to another possibility. The possibility it moves to does not yet exist. Data does not support the work of comparing a thing that exists with a thing that does not. Fourth, responsibility has been dispersed. “The AI analysis also supported continuing” has become an available explanation. The vaguer the deciding subject, the further exit is postponed. In the Age of AI, therefore, exit has to be handled as a problem of design. That means setting the conditions of ending at the moment of starting. What has to happen for us to stop. When we judge. Who judges. Document those three at the outset, and put the judgment automatically on the agenda on the judgment date. This is a concrete instance of Human-on-the-Loop. We do not try harder at each individual exit decision. We design the system so that exit reaches the agenda structurally. What can be read across industries Within what can be read from public information, we lay the theory over the shape of several industries. In diversified electronics, companies holding many business areas have been seen to swap the areas themselves. What deserves attention is that many of the businesses swapped out were still profitable at the time. They did not stop after things went bad. They moved capital to another possibility before things went bad. That decision is impossible unless people hold Layers 4 and 5. In pharmaceuticals, discontinuation during research is built in as an institution. Each stage has criteria, and candidates that miss them are stopped. Discontinuation is not a failure. It is the procedure by which capital moves to the next candidate. This is the most mature example we have of exit treated as design. There is a contrasting picture. Companies that held the technology and still failed to meet a market turn did not lack capability. They could not produce the decision to stop as an institution. When the Future Horizon is fixed to the life of the existing business, the option set is closed from the start. What is common to all three? The results were separated not by skill in deciding, but by skill in designing how decisions are made. What a company that plans backward looks like Companies where Future Back Planning is working show common signs. The budget table carries line items that contribute nothing to this year’s revenue. Hiring requirements include capabilities the current businesses have no use for. The executive agenda holds items that cannot be judged within five years, and those items carry the name of a responsible person. None of these appear unless the company has calculated backward from the future. A company showing none of them is selecting only an extension of the present, however long the horizon of its stated vision.

6 Questions for the executive

The argument, in one line. Decision-making in the Age of AI is the practice of leaving optimization to AI, keeping purpose and responsibility with people, and allocating capital by continuously calculating backward from the future. It is not deciding fast. It is not gathering more data. It is not picking the right option. All of those are parts of the practice, not the practice. If we adopt this definition, how does management change tomorrow? Here are three questions. Each can be answered at your next executive meeting. Question 1 — In your most recent important decision, which of the five layers did your time go into? Time spent reading materials, comparing, and confirming belongs to Layers 1 through 3, and can be moved to AI. Once it has moved, is the freed time going into Layers 4 and 5? If it is not, AI is only making your meeting packs thicker. Question 2 — How many years is your Future Horizon? Measure it institutionally. The payback standard for investments, the period of the plan, the appraisal cycle. The shortest of these is the company’s actual Future Horizon. Futures outside that span are never selected, however often they are discussed. Only the executive has the authority to extend it. Question 3 — Of the initiatives now running, what share have a defined condition of ending? Agreement gathers easily around a decision to start. It does not gather around a decision to stop. The conditions for stopping can therefore only be set at the moment of starting. In a company where that share is low, capital quietly becomes fixed to the past. None of the three questions asks how you decide. All three ask how you have designed the act of deciding. AI widens the options. People fix the purpose. Capital creates possibility. Responsibility belongs to whoever decided. In a company where that division is explicit, management gets stronger as AI gets stronger. In a company where it is vague, less and less gets decided as AI gets stronger. The difference is not model performance. It is the structure of decision-making. And only the executive can design that structure. AI will not tell us which future we ought to choose. That it will not is exactly why the office of the executive remains. Decision-making means choosing a future. Choosing means discarding the rest. Only when a subject carries the weight of what was discarded does the act deserve the name of a decision.

In brief

A decision is the act of selecting one among several possible futures, allocating capital toward it, and carrying the result.
Layers 1 through 3 of the five can move to AI. The choice at Layer 4 and the responsibility at Layer 5 do not move.
The Future Horizon is set by institutions. The shortest span among the payback standard, the plan period, and the appraisal cycle is the real horizon.
Conditions of ending can only be set at the moment of starting. Where they are not set, capital becomes quietly fixed to the past.

Key concepts

Human-on-the-Loop Management / Human-in-the-Loop / Future Horizon / Future Back Planning / Future Time Equation / Capital Allocation

The chain of ideas

Future Horizon → Future Back Planning → Capital Allocation → Future Value → Enterprise Value

Related first principles

Principle 3 — Capital Exists to Create Possibility. Principle 4 — AI Optimizes. Humans Define. Principle 9 — Leadership Means Designing the Future.

Related chapters

Vol. I, Ch. 002 “Can AI Become a CEO?” — argues in three layers why responsibility does not transfer to AI
Vol. I, Ch. 009 “How Should Work Be Divided Between AI and People?” — brings the five layers down to verbs and a responsibility ledger
Vol. I, Ch. 008 “What Is Management Strategy in the Age of AI?” — how to build a vision by calculating backward from the future
Vol. IV, Ch. 032 “How Do Investors Assess Future Value?” — the path by which the length of the horizon reaches external evaluation

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, #013 “Should We Let AI Make Management Judgments?” / #020 “Why Do We Get It Wrong When the Data Is There?”

Chapter 005 What Is Decision-Making in the Age of AI?