Chapter 009　How Should Work Be Divided Between AI and People?

How should work be divided between AI and people? Most companies answer with a table. AI’s tasks on the left, human tasks on the right. Six months later the table no longer matches the company. The problem is not what the table contains. The problem is the criterion used to build it. In this chapter we recast the division of roles. It is not an allocation of tasks. It is the design of where responsibility sits. We set out the design principle, and then the implementation, layer by layer.

1 The question — why it arises now

Dividing work between people and machines is not a new question for management. It has been asked since the industrial revolution. Standing in front of a conveyor line, a plant manager decided which steps went to the machine and which kept a person. The criterion was clear. Machines repeat identical motions with precision. People do everything else. The line could be drawn by the nature of the motion. Computers did not change that picture much. Calculation, record-keeping, and retrieval moved to the machine. Interpreting the output and deciding what to do with it stayed with the person. Machines took the work; humans took the judgment. One sentence explained the whole of twentieth-century role design. AI broke that sentence. What AI took over was not the work. It was the judgment. AI reads markets. It infers causes. It lays out options. It ranks them. It writes prose, writes code, and proposes designs. Most of the territory once called human judgment now sits inside AI’s output range. That range does not hold still. Tasks AI handled poorly three years ago it handles well today. Tasks it handles poorly today may look different in a few years. The line we used as our criterion moves faster than the table built on it. The common response is to update the table more often. An annual review becomes a quarterly one. This does not solve anything. Raising the frequency of revision while keeping the criterion of revision means holding a slightly obsolete table forever. The question therefore has to be restated. It is not “which tasks do we assign to AI and which to people.” It is “what criterion can we divide by, such that the criterion itself does not go stale?” This looks like a question of work design. It is a question of management. The criterion for dividing work is the criterion for who is answerable for what. An organization with unclear responsibility decides more slowly as AI gets stronger. An organization with clear responsibility decides faster as AI gets stronger. Two companies deploy the same AI. One speeds up. One slows down. The difference is not model performance. It is the design principle behind the division.

2 Conventional answers and their limits

Three answers circulate today. All three are in active use. All three stop working within a few years. The first answer: “Draw the line at what AI can and cannot do” This is the most widely adopted method. Inventory what AI can do, list the tasks it covers, and assign the remainder to people. It looks rational. In the first phase of adoption it works. It breaks down structurally. The reason is not simply that the boundary of capability moves. It is that the speed of the move and the speed of institutional revision are nowhere near each other. A division table is an institution. Institutions have revision procedures. Agreement across departments. Amendment of authority rules. Alignment with the appraisal system. Configuration changes in the systems themselves. Companies that need several months to rewrite one line of a division table are not unusual. Meanwhile AI capability is refreshed every few months. By the time the revision closes, the assumption behind it has already moved. Under this structure a company always operates on a stale table. The staleness also runs in one direction. Tasks AI can already do keep sitting on the human side. The organization spends human time against an obsolete criterion, without noticing. There is a deeper failure. If you define the human role by what AI cannot do, that role is cut back every time AI gets stronger. This is design by subtraction. What is impossible this year may be routine next year. A job built on that assumption is permanently provisional. A person given a provisional job cannot define their own work. Work that cannot be defined cannot be owned. What is lost here is not efficiency. It is responsibility. And in an organization where responsibility has been lost, nobody picks up what AI produces. The second answer: “Humans take the important work, AI takes the routine” The second method divides by importance. It matches intuition. In practice it fails twice. First, importance is a continuous quantity, not a quantity you can draw a line through. Every task becomes important under some condition. Expense processing is routine on a normal day and the most important process in the company the moment fraud appears. To divide by importance is to divide by importance on a normal day. Accidents do not happen on normal days. Second, routine work carries responsibility too. Approving credit. Passing or failing a quality check. Halting a shipment. The procedure in each case is routine. The consequence reaches customers and society. The logic “it is routine, so AI can have it” confuses the nature of a procedure with the weight of a responsibility. Division by importance works quietly in normal conditions and fails without exception in abnormal ones. A role design must be judged by how it fails. The third answer: “A human signs off at the end, so the division holds” The third answer looks safest. AI processes; a human approves. Responsibility rests with the approver. It is easy to explain to a governance committee. This is the Human-in-the-Loop arrangement, in which humans supervise AI from inside the operational loop. A person sits inside the loop and checks output one item at a time. Early in adoption this is reasonable. The arrangement is weak against volume. AI’s throughput keeps rising. The number of approvals rises in proportion. Human time available does not rise. At some point, therefore, approval necessarily becomes a formality. An approval signed without reading is not the acceptance of responsibility. It is the appearance of responsibility. Worse is the case where the approver has not been told what to look at. Without a criterion for judging whether AI’s output is sound, approval is a wager. Approval without a criterion creates a state that looks like accountability and is in fact accountability by nobody. The three conventional answers share one assumption. All three treat the division of roles as an allocation of tasks. Who performs which task. That framing is already out of date. What has to be allocated in the Age of AI is not tasks.

3 Redefinition — the division of roles is the design of

where responsibility sits Future Value Theory defines the division of roles between AI and people as follows. The division of roles is not an allocation of tasks. It is the design of where responsibility sits. Why adopt this definition? For one reason. The boundary of capability moves. The location of responsibility does not. However advanced AI becomes, AI does not take on responsibility. This is not a limitation of performance. Responsibility means carrying an outcome as your own, explaining it to society, and paying the price when a price is due. That belongs to the domain of meaning. AI can compute meaning. It cannot assume meaning. First Principle 4 states the structure in one line. AI Optimizes. Humans Define. AI optimizes; humans define value, purpose, and direction. Optimization happens inside a given objective function. Definition is the act of setting the objective function itself. The criterion for dividing, then, is not capability. It is verbs. Three verbs to people, three verbs to AI Three verbs stay with people. Define, choose, and own. Define means deciding what this is being done for. Framing the question, setting the purpose, selecting the axis of evaluation. Deciding what will be called success and what will be called failure. Without that decision, AI cannot even begin. Choose means deciding which of several futures to take. AI can lay out the options and estimate the consequence of each. Which consequence is desirable is a judgment of value. A judgment of value requires a subject. Answer for means carrying the outcome. When something goes wrong, who explains, who absorbs it, and who decides the next move. A task with a blank at that point is incomplete as management, however thoroughly it has been automated. Three verbs go to AI. Analyze, generate, and execute. Analysis extracts structure from data. Generation produces options, prose, designs, and code. Execution advances processing under defined conditions. AI does all three faster, more broadly, and without fatigue. The six verbs are not a ranking of ability. AI analyzes better than we do. It generates better too. Definition stays with people anyway — not because people are better at it. Definition requires a subject, and a subject carries responsibility. Three exceptions, built into the design from the start The basic structure has three practical exceptions. Discovering an exception later breaks the design. Build them in from the beginning. Exception 1. AI touches definition. AI can propose the axis of evaluation itself. “This business should be measured by retention rather than gross margin” is a proposal AI can produce well. The boundary to protect here is the separation of proposal from adoption. AI may propose. A human decides whether to adopt. The better the proposal, the more easily that separation blurs. So separate it institutionally, not by good intentions. Exception 2. People stay in execution. Domains where law requires human execution. Domains where trust is established only through physical contact. Domains so new that AI has no learning to draw on. People execute in these. But they do not stay because only a person can do the work. They stay because responsibility and execution cannot be separated in that structure. A different reason produces a different design. Exception 3. Responsibility for execution rests with the definer, not the executor. When AI processes something automatically, responsibility for the result is not AI’s. It belongs to the person who defined the scope of that automatic processing. The scope of automation is therefore always defined in advance, by a person. Automatic execution outside a defined scope creates a blank in responsibility. Most accidents in the Age of AI happen in that blank. Human Capital and AI Capital are joined by multiplication Look at the division from the standpoint of capital. Future Value Theory treats the capital of an enterprise this way. Future Capital = Financial × Human × Learning × Trust × AI × Knowledge × Ecosystem × Purpose This is multiplication, not addition. If any single term is zero, the whole product is zero however large the others are. An abundance of financial capital cannot compensate for absent purpose, and advanced AI cannot compensate for absent trust. One conclusion about capital in the Age of AI follows immediately. AI Capital produces no value on its own. Consider a company whose Human Capital term is close to zero. However much AI it deploys, no Future Capital appears. The converse holds as well. The role of AI Capital is to amplify the other terms. It accelerates learning, extends the reach of knowledge, and multiplies the contact points of the ecosystem. But an amplifier with zero input produces zero output. The inputs here are Purpose and Human Capital. This gives the design of roles one consequence. Designing the division is designing the circuit that connects AI Capital to Human Capital. From people to AI flow purpose, evaluation axes, and decision criteria. From AI to people flow analysis, options, and execution results. Where that loop is broken, AI remains an isolated asset. Most of the “our AI deployment is not working” conversations we are asked into can be explained as a break in this circuit. AI is running. People are working. The two are not connected. When AI Capital and Human Capital exist separately, there is no product. Human-on-the-Loop Management — design the generating rule, not the table Restate the structure as a form of management. That form is Human-on-the-Loop Management. Under Human-on-the-Loop Management, human beings sit above the system. They are not intervening in individual outputs. They design people, AI, capital, the organization, and society as one system. This is not a doctrine of supervision. It is a doctrine of design. The objective is better design rather than better control. The implication for role design is direct. What the executive should build is not the division table. It is the generating rule from which the division table updates itself. The generating rule can be written as three questions. Who decides the purpose of this task? Who holds the axis of evaluation? Who is accountable for explaining the result? With those three answers filled in, the division survives even when the work itself moves to AI. Task allocation shifts naturally with AI capability. The location of responsibility does not shift. A company that divides by boundary lines rebuilds its design every time a boundary moves. A company that divides by responsibility does not rebuild when a boundary moves. Over five years that difference becomes large.

4 Structure — three frames that hold the division

together Three structures carry the redefinition into practice.

4.1 The responsibility ledger — what replaces the division table

What most companies hold is an assignment table, task by task. What they should build is a responsibility ledger, task by task. The responsibility ledger does not record the task name and the assignee. It records four items. The purpose of the task. The axis of evaluation in use. The name of the person who set that axis. The name of the person accountable for the result. The executor field can stay empty. Whoever is the best subject at the time goes in it, human or AI. The format has a secondary effect. The moment you try to build the ledger, you find tasks whose axis of evaluation nobody ever set. The more a task has run on custom, the more likely that is. This is exactly where the first accident after an AI deployment occurs. Put AI into a task whose purpose was never put into words, and AI will optimize a substitute purpose. The responsibility ledger is not a document for AI. It is a document for people.

4.2 The Future Time Equation — where the returned time goes

decides the outcome The success of a role design cannot be measured in hours removed. It is measured by this equation. Future Value = Future Time × Future Capability Future Time is time intentionally invested in creating the future. Future Capability is the capability that converts that time into value. This too is multiplication. If time is zero, no Future Value appears however high the capability. AI can raise Future Time directly, because it cuts the time spent on analysis, generation, and execution. But time released does not become Future Time automatically. Returned time has three destinations. The first is a return to supervision. The freed hours go into checking AI’s output one item at a time. This is a regression to Human-in-the-Loop. The more volume grows, the more this choice squeezes the organization. The second is more throughput. The same work, more of it, faster. Short-term productivity rises. The company is now doing last year’s work quickly. Only the third becomes Future Time: moving hours into definition and choice. Which market do we want to bring into existence? Which of our assumptions is going obsolete? Toward which future do we allocate capital? Only when that time increases can we say the role design succeeded. The measure of a role design is therefore not “hours saved.” It is the share of time that moved into definition and choice. In a company that measures only the first, AI ends as an expensive efficiency device.

4.3 Designing the division is part of management itself

Many companies treat the division of roles as a process-improvement item. It is a core act of management. This equation shows why. Leadership = Purpose × Question Design × Capital Allocation × System Architecture × Trust The fourth term, System Architecture, is precisely the design of the division between people and AI. Who defines, who chooses, who is accountable, and how far processing runs automatically. That design fixes the speed and the safety of the enterprise at the same time. This equation is multiplicative as well. Consider leadership with System Architecture at zero. However clear the Purpose, it cannot be converted into organizational motion. Consider the reverse: an exquisite architecture with Purpose at zero. A precisely meaningless optimization runs. A company that has delegated the whole of this design to the front line or to the IT function has left the term blank. Some designs have no one to whom they can be delegated.

5 What it looks like in practice — how the division

changes by layer The division of roles is not one design for the whole company. What must be divided differs by layer. The executive layer — move time toward definition and choice What remains with the executive team is the three verbs: define, choose, and own. Analysis moves to AI. Market structure analysis, competitor tracking, scenario enumeration, financial sensitivity analysis. Work that once took a corporate planning team weeks now finishes in far less time. The agenda of the executive meeting is therefore rebuilt. Reporting and confirmation can be read in advance, prepared by AI. The human hours inside the meeting go to reselecting the axis of evaluation and to adopting or rejecting options. The output of the meeting is the number of items the minutes can record as decided. One failure is common here. The more refined AI’s options are, the more the executive team is put in the position of being made to choose. Three options appear, two are clearly inferior, and the remaining one is decided by default. That is not choice. The power of definition is the ability to doubt the set of options itself. Middle management — from a junction of information to a junction of responsibility The most unsettled layer is middle management. The traditional function had three parts: relaying information up and down, allocating work, and evaluating results. Relaying is replaced by AI and the information platform. Conditions on the front line are aggregated without passing through a manager. Much of allocation can be automated as well. The traditional manager loses most of the traditional functions. Two functions are not lost. Holding a criterion, and growing people. Holding a criterion means putting into words what good work is for this unit, and deciding when judgments split. AI can apply a criterion. Whether that criterion fits this particular front line is something a person has to own. The manager becomes the named signatory on the unit’s responsibility ledger. Growing people means designing the occasions on which subordinates experience definition and choice. The more AI takes over execution, the fewer chances junior people get to practice judgment. Left alone, this produces a company with nobody able to define anything in ten years. Few companies have noticed. The redefinition of middle management fits in one line. From a junction of information to a junction of responsibility. The front line — from executor to finder of exceptions When AI takes over execution at the front line, the human role becomes heavier, not lighter. Processing inside the defined scope goes to AI. What remains for people is finding the situations that were never defined. A customer reaction nobody anticipated. Something wrong with a machine that the numbers do not show. An event that fits nowhere in the procedure manual. For AI these sit outside the training data. Only the people on the floor notice them first. That discovery is an input to definition. Because exceptions come up, management can update the axis of evaluation. In an organization where exceptions do not come up, the definition stays frozen. When the front line goes quiet, learning stops, however smart AI becomes. The most important part of role design at the front line is therefore not the allocation of execution. It is the route by which exceptions travel upward, and an appraisal system that does not penalize the person who raises one. Deploy AI without designing this, and the front line starts saying “AI decided it.” The blank in responsibility spreads upward from below. Two flows that connect the three layers The three layers are connected by two flows. Downward flow purpose and the axis of evaluation. Upward flow exceptions and learning. That round trip is the Learning stage of the Future Value Chain. Purpose → Learning → Redefinition → Creation → Enterprise Value Designing the division of roles is the work of rebuilding the Learning circuit in that chain, using both people and AI. What the division looks like across industries In manufacturing, visual inspection has moved to AI in a growing number of plants. Success and failure separate here. In the plants that succeeded, inspectors moved into the meeting that sets the pass/fail criterion itself. In the plants that failed, inspectors keep watching AI’s verdicts with their own eyes. Same technology. Only the destination of the time differs. In financial services, models have judged credit for a long time. The crux of the division is not model accuracy. It is the definition of which segment we should extend credit to, given our Purpose. Where a lender lacks that definition, the model effectively sets policy. In software, it is public knowledge that major providers have built generative AI into their development processes. What can be observed is that the organizations pulling ahead are the ones where the hours removed from writing code were replaced by argument about what to build. What can be read from public information extends that far and no further. What failure looks like — the company that makes a table and stops Finally, the most common failure. A company builds a division table, distributes it company-wide, and treats the job as done. Six months later the table is unused. The reason is simple. The table records tasks and not responsibility. Tasks migrate as AI advances. Nothing records who sets the purpose at the destination, or who picks up the result. The organization then grows a particular kind of task. AI proposes it, nobody has decided whether to adopt it, and the front line runs it as proposed. Nobody notices until something goes wrong. Afterward, the search for where responsibility sat comes back empty. Failure in the division of roles is not the failure to raise efficiency. It is the appearance of a blank where responsibility should be.

6 Questions for the executive

The argument, in one line. The division of roles between AI and people is not the splitting of tasks along a boundary of capability. It is the design of where definition, choice, and responsibility sit — and, through that design, the continuous connection of AI Capital to Human Capital. Divide by boundary lines, and the design breaks every year. Divide by responsibility, and the design stands when the boundaries move. Only the company that divides by responsibility can send the time AI returned into the future. Three questions to close. Each can be answered at your next executive meeting. Question 1 — For each of your major tasks, can you name the person who set its purpose and its axis of evaluation? Where you cannot, responsibility is blank. Put AI into a blank and AI optimizes a substitute purpose. In most cases the substitute is a short-term number. An optimization nobody wanted proceeds quietly. Question 2 — Where is the time AI returned going right now? Most companies measure hours saved. Few measure the destination. Back into supervision, into more throughput, or into definition and choice? If Future Time has not increased, the deployment is still only an efficiency program. Question 3 — When the front line finds a situation that was never defined, how many days does it take to reach management? That number is the learning speed of the enterprise. Where it never arrives, the axis of evaluation is never updated. Keep running AI under an axis that is never updated, and the company grows obsolete with precision. None of the three questions asks what AI should be made to do. All three ask what human beings will take on. AI analyzes. People define. AI generates. People choose. AI executes. People own the result. The arrangement of those six verbs is the whole of role design in the Age of AI. As technology advances, AI’s three grow stronger. The human three do not grow stronger. They grow heavier. The more AI can do, the more weight sits on the act of deciding what it is all for. AI Optimizes. Humans Define. That line is not the conclusion of the division. It is the starting point from which the division is designed.

In brief

The division of roles between AI and people is not an allocation of tasks; it is the design of where responsibility sits.
The boundary of capability moves every year. The location of responsibility does not. Only a design divided by responsibility survives.
AI Capital is an amplifier. Where Human Capital is thin, no Future Capital appears.
What the executive builds is not the division table but the generating rule that keeps the table current.

Key concepts

Human-on-the-Loop Management / Human-in-the-Loop / Future Capital / Future Time / Future Time Equation

The chain of ideas

Purpose → Human-on-the-Loop Management → Future Capital → Future Time → Future Value

Related first principles

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

Related chapters

Vol. I, Ch. 002 “Can AI Become a CEO?” — shows in three layers why responsibility does not transfer to AI
Vol. I, Ch. 005 “What Is Decision-Making in the Age of AI?” — the five-layer breakdown and the fixing of boundaries
Vol. II, Ch. 016 “What Kind of People Does the Age of AI Need?” — the people who carry the three human verbs
Vol. II, Ch. 015 “Will AI Take Our Jobs?” — how the reallocation of roles reaches employment

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, #005 “The Real Division of Roles Between AI and People” / #041 “The Day Your Reports Become AI: What Happens to the Manager’s Job” / #047 “What You Do with the Time AI Frees Up Decides the Company”

Chapter 009 How Should Work Be Divided Between AI and People?