Who Creates the Value and Who Gets It

Accounting can tell us where money went under the rules we have. That is useful. I would prefer the arithmetic not be vibes.

It cannot tell us, by itself, whether those rules produce a just claim on what remains. A ledger can close perfectly while the moral argument is still standing in the hallway asking why nobody invited it to the meeting.

Why does the legal residual claim establish the morally correct residual claim?

The fact that shareholders legally own the corporation tells me how the current institution distributes claims. It does not deduce from first principles that this distribution is just.

Imagine an apple cart.

One person supplies capital and owns the cart. Another person operates it every day, selects inventory, serves customers, maintains equipment, keeps records, and creates the relationships that make the business work. The law can assign the owner the residual after wages and expenses. That tells us what the contract permits. It does not mathematically prove that any particular split of the residual between owner and worker is ethically deserved.

Scale does not cure the philosophical problem. It gives the problem an ERP system, a tax department, and several excellent reasons not to call itself philosophical anymore.

Modern enterprise is a cooperative production system involving capital, labor, suppliers, customers, public infrastructure, legal institutions, education systems, communications networks, intellectual property, natural resources, and communities. Capital is indispensable. So is labor. So are the institutions that make property and contracts enforceable. The normative question is not whether investors deserve compensation for risk. Of course investment requires incentives. The question is why the ownership structure receives such presumptive moral priority that every distribution produced by it is treated as a natural fact rather than a political choice.

That question becomes urgent as automation increases.

Acemoglu and Restrepo’s task framework is helpful because it strips away the mythology. Production consists of tasks. Some tasks can be performed by labor, some by capital, some by both. When technology moves a task from labor to capital, a displacement effect occurs. When technology creates new tasks in which people have an advantage, a reinstatement effect occurs. [S114]S114 · Automation and New Tasks: How Technology Displaces and Reinstates LaborDaron Acemoglu; Pascual Restrepo · 2019NBER / peer-reviewed economics · labor mechanismClose

Automation is therefore not automatically anti-worker. The outcome depends on which effect dominates and on the institutions surrounding the transition.

Their empirical work is sobering. Acemoglu and Restrepo estimate that between 50% and 70% of changes in the US wage structure over four decades can be accounted for by relative wage declines among worker groups specialized in routine tasks in industries experiencing rapid automation, in the framework they develop. [S115]S115 · Tasks, Automation, and the Rise in US Wage InequalityDaron Acemoglu; Pascual Restrepo · 2021NBER / peer-reviewed economics · empirical labor evidenceClose They have also argued that US tax policy can favor capital over labor and induce more automation than would occur under an alternative tax system in their model. [S116]S116 · Does the U.S. Tax Code Favor Automation?Daron Acemoglu; Andrea Manera; Pascual Restrepo · 2020NBER economics · policy mechanismClose

So the distributional result is not simply “technology happened.” Technology is not weather. Somebody owns it, prices it, deploys it, bargains around it, taxes it, and decides which savings become whose gains.

Technology interacts with tax policy, labor bargaining, ownership, task creation, education, market concentration, mobility, and social insurance.

The distributional baseline is measurable. In Q2 2026 labor’s share of nonfarm business-sector output hit 52.8%, the lowest in the BLS series beginning in 1947. [S117]S117 · Productivity and Costs, Second Quarter 2026, RevisedU.S. Bureau of Labor Statistics · 2026official statistics · current empirical baselineClose The Federal Reserve’s Q1 2026 distributional accounts imply the top tenth of households owns roughly 68% of household net worth while the bottom half owns roughly 2.5%. [S118]S118 · Distributional Financial AccountsFederal Reserve Board · 2026official statistics · wealth-distribution baselineClose

That ownership distribution matters because many AI capabilities depend on assets and contractual access controlled by their owners. A purchased AI service may be an operating expense to its customer; that does not erase the ownership of the infrastructure supplying it.

A company buys compute, models, software, data, infrastructure, intellectual property, and automation capability. If the resulting system allows the company to produce the same output with fewer labor hours, the initial economic benefit appears where ownership and bargaining rules put it. Workers may receive higher wages because their productivity rises. Consumers may receive lower prices. New industries may emerge. Governments may tax the gains. Owners may capture much of the surplus. Some mixture will occur.

There is no silicon commandment specifying the mixture. The chip does not descend from the mountain carrying a compensation philosophy.

The IMF’s 2025 AI-inequality paper makes this explicit. Depending on displacement and complementarity, AI could reduce wage inequality in some scenarios because higher-paid cognitive jobs are unusually exposed. Yet the same model can increase wealth inequality because high-income households own more capital and firms invest aggressively where expensive labor can be automated. [S179]S179 · AI Adoption and InequalityEmma J. Rockall; Marina Mendes Tavares; Carlo Pizzinelli / IMF · 2025IMF working paper · distributional-model evidenceClose

This is why the simplistic automation argument fails in both directions. We have somehow managed to compress an economic transformation into two competing bumper stickers.

“AI will make everyone rich” assumes away distribution. “AI will make everyone unemployed” assumes away new tasks, complementarity, prices, demand, and adaptation.

The interesting question is institutional:

Who owns the machines, who owns the firms, who has bargaining power, who receives the productivity dividend, and who bears transition costs?

The ILO’s evidence is refreshingly unsensational. Its 2025 occupational analysis finds one in four workers worldwide is in an occupation with some exposure to generative AI, but most affected jobs are more likely to be transformed than eliminated outright. [S180]S180 · Generative AI and Jobs: A Refined Global Index of Occupational ExposurePaweł Gmyrek et al. / ILO · 2025ILO working paper · labor-exposure evidenceClose Its 2026 review finds productivity gains are real but uneven and economy-wide displacement remains limited so far. It also finds that worker-reported time savings have not automatically translated into higher measured output, earnings, or employment. [S181]S181 · The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidenceRossana Merola et al. / ILO · 2026ILO empirical review · labor/productivity evidenceClose

That is exactly the institutional gap I care about. The benchmark ends precisely where the bargain begins.

If a worker saves ten hours a week with AI, several futures are possible.

The worker can work thirty hours for forty hours of pay.

The worker can produce more in forty hours and receive some of the gain.

The employer can raise output expectations until forty hours feels exactly as pressured as before.

The employer can eliminate part of the team and distribute the same workload among fewer people using AI.

The employer can retain everyone, lower prices, gain market share, and grow.

The employer can capture most of the benefit as profit.

All are compatible with the same technological productivity improvement. Same tool. Six different bargains.

That is what I mean when I say aligned AI can accelerate a bad system. If the surrounding institution already treats productivity gains primarily as opportunities for cost reduction and capital return, making the optimization engine smarter does not spontaneously create a cooperative economy.

It makes the existing objective easier to achieve.

I am not asking every company to become a charity. Charity is not the point. I am asking why the phrase “fiduciary discipline” walks into the room wearing a suit and carrying presumptive moral authority while “the people creating this value should receive more of it” is treated like somebody brought politics to dinner.

Business Roundtable’s 2019 statement is useful here because it formally rejected a purely shareholder-only rhetoric and pledged commitments to customers, employees, suppliers, communities, and shareholders. [S186]S186 · Statement on the Purpose of a CorporationBusiness Roundtable · 2019corporate-governance statement · counterevidence / declared stakeholder postureClose The language of stakeholder capitalism has therefore entered the boardroom. The harder question is how much authority those stakeholders possess when their interests collide with earnings, margins, or capital returns.

The AI transition will answer that question in practice.

If millions of workers become dramatically more productive and the outcome is shorter workweeks, higher wages, stronger public services, cheaper essentials, new forms of ownership, and broad household wealth creation, AI could reduce many of the pathologies described in this essay.

If the outcome is thinner headcount, more intense surveillance, individualized productivity targets, greater capital concentration, weaker bargaining, and a growing population renting access to intelligence owned by a handful of firms, then the technology will have solved the wrong problem beautifully.

Three accounts of the same improvement

A productivity gain can appear in three different accounts: the production account, the distribution account, and the account of a person’s life. They are connected, but no single one contains the others. The first two fit beautifully in a quarterly deck. The third is where somebody discovers whether getting twice as much done means going home at three or getting assigned twice as much work.

The production account asks how much output is produced for a given input. The distribution account asks how income, cost savings, and ownership claims are divided. The human account asks what the change does to time, security, dignity, and the ability to make plans. A successful transition has to be understood across all three.

Imagine a small design business with ten employees. A new tool reduces the time required to prepare a routine draft. In one scenario, the firm wins more clients, hires another person, and shares the gain through pay. In another, it keeps revenue steady and reduces working hours. In a third, it cuts staff and uses the remaining team to maintain output. In a fourth, competitors adopt the same tool, prices fall, and most of the gain reaches clients rather than employees or owners.

These are possible arrangements, not forecasts. They show why a task-level speedup cannot be read directly as a wage increase, a profit increase, or a job loss. The tool changes what is feasible. Markets and institutions shape which feasible arrangement becomes actual.

The published study Generative AI at Work offers a concrete example of a real gain. Brynjolfsson, Li, and Raymond examine a staggered rollout among 5,172 customer-support agents and report an average increase of about 15% in issues resolved per hour, with larger benefits for less experienced and lower-skilled workers. The rollout provides evidence about that workplace and task. It is not a randomized measurement of the whole economy, and it does not establish that every minute saved became additional pay or leisure. [S201]S201 · Generative AI at WorkErik Brynjolfsson; Danielle Li; Lindsey Raymond · 2025peer-reviewed economics / Quarterly Journal of Economics · workplace productivity evidenceClose

The study is valuable precisely because it does not need to be turned into a universal promise. It shows that useful complementarity is possible. Good. Now the interesting part starts. Through what arrangement does the benefit reach the people whose work changed?

A flow is not a stock

Income and wealth are easy to mix up because both are measured in money. Their roles are different. Income is a flow over a period. Wealth is a stock of assets minus liabilities at a point in time. A household can receive a higher income and remain without a meaningful buffer. Another can receive modest current income while owning substantial assets.

The difference changes what a shock means. A person with a buffer can search longer for a job, repair a car without missing rent, or take time to learn a new skill. A person without one may have to accept the first available arrangement. The same wage offer therefore does not create the same practical freedom for every worker. Money buys things. Wealth also buys time to say no.

This is an analytical point about resources and options. It does not require assuming that everyone without wealth makes the same decision or that wealth removes every constraint. It means that the ability to wait is itself economically consequential.

AI can interact with both accounts. It can raise a worker’s income by making their skills more productive. It can also raise the expected returns to a firm that owns a useful model, a distribution channel, or complementary infrastructure. If access to those returns is uneven, income gains and wealth concentration can move in different directions.

The IMF’s distributional model is useful for making that possibility explicit. Its scenarios distinguish labor displacement, complementarity, and capital returns. A reduction in wage inequality in a scenario does not imply a reduction in wealth inequality. [S179]S179 · AI Adoption and InequalityEmma J. Rockall; Marina Mendes Tavares; Carlo Pizzinelli / IMF · 2025IMF working paper · distributional-model evidenceClose That is not a prediction that a particular distribution must occur. It is a warning against using one measure as a substitute for another.

The current Census result belongs here rather than in a footnote reluctantly added to protect an argument. Real median household money income rose in 2025. That is a positive change in the measure. It does not tell us that every household improved, and it does not measure the distribution of net worth. The release also reports that the change in the income Gini was not statistically significant. Those findings can coexist without contradiction. [S119]S119 · Income in the United States: 2025U.S. Census Bureau · 2026official statistics · credibility check / baselineClose

An argument about stratification should be able to survive good news because stratification is not defined as the absence of every improvement. It concerns differences in resources, security, access, and the capacity to benefit from future change. A rising median can be part of a better trajectory. It is not a complete description of that trajectory.

The value of work is larger than the record of tasks

Task-based analysis is more useful than treating each occupation as one indivisible object. A job can contain drafting, judgment, relationship management, checking, coordination, and responsibility for unusual cases. A tool may automate one component and increase the importance of another. [S114][S180]S114 · Automation and New Tasks: How Technology Displaces and Reinstates LaborDaron Acemoglu; Pascual Restrepo · 2019NBER / peer-reviewed economics · labor mechanismS180 · Generative AI and Jobs: A Refined Global Index of Occupational ExposurePaweł Gmyrek et al. / ILO · 2025ILO working paper · labor-exposure evidenceClose

But the task description can itself be incomplete. Some work is visible because it produces a document or closes a ticket. Other work prevents a failure that never appears. A colleague may notice a risky assumption before it becomes an incident. Someone may explain a process to a new hire without logging a training session. Someone may know which apparent exception is actually a warning sign.

These are illustrative examples of why an inventory of recorded tasks can understate contribution. They do not establish that every informal practice should be preserved. Some habits are wasteful. Some workarounds conceal a broken process. But a ticketing system is not an ontology of human usefulness. Eliminating a recorded task and eliminating a person’s contribution are different operations.

That matters when the worker’s knowledge helps train or improve a system. The firm may capture patterns from past work, customer interactions, or internal documents. A tool can then make some of that knowledge available to others. The resulting capability is real. The question of compensation for the people whose work helped create it is not answered merely by showing that the model can reproduce a useful pattern.

Nor is contribution always easy to assign. Knowledge accumulates through teams, institutions, tools, and earlier work. A simplistic calculation that credits every output to one person would be as misleading as one that credits everything to the owner of the final interface. Cooperative production makes the attribution problem difficult. It does not make the distributional question disappear.

The accounting still matters. Revenue is not distributable profit. Net income is not the same as cash available for immediate distribution. Capital spending and working-capital needs matter. If the arithmetic is wrong, fix the arithmetic.

Then ask the harder question anyway. A correct spreadsheet does not become a theory of justice merely because all the cells sum to zero.

None of those accounting distinctions supplies a complete theory of what each participant deserves. The books can be immaculate while the argument over ownership, bargaining power, risk, time, and contribution remains completely open.

Transition has a clock

A long-run gain can coexist with a severe short-run loss. That is not a paradox. A new industry may emerge after a person has exhausted savings. A lower price may arrive after a household has lost its income. A training opportunity may be useful but incompatible with the time and money available during the transition.

The relevant comparison is not only between two eventual states. It includes the path between them. Who can finance the interval? Who can move? Who can absorb an unsuccessful attempt? Who has care responsibilities that cannot be suspended while a market adjusts?

Those questions are often treated as details after the economic forecast. For a person living through the change, they may be the whole problem. “The economy adapts” describes an aggregate process. It can be completely correct and still arrive six months after somebody ran out of rent. It does not name the people who bear the costs of adaptation or the timing of the gains they may receive.

The same applies inside a firm. A tool may require new checking, new training, and a period of lower output before it becomes useful. A narrow comparison between the old workflow and an idealized mature workflow can omit the work needed to get from one to the other. A claimed saving should therefore be examined over an appropriate time horizon rather than only at the moment of a successful demonstration.

Useful tools still have transition costs. Count the work and risk where they actually occur. Otherwise the organization can describe a saving that depends on employees supplying unpaid transition labor or customers tolerating temporary failures. Costs do not evaporate when they cross the edge of a department. We covered the dimensional portal problem already.

The bargain remains a human question

There is no single distributional arrangement implied by intelligence. A more capable system can support higher output, better quality, lower prices, shorter hours, greater profits, or some combination. It can also change which participants have leverage over the combination.

That last point is easy to miss. AI is not only an input into production. It can become part of negotiation, monitoring, hiring, pricing, and the description of what work is worth. A tool that helps an employee perform a task may be paired with a tool that measures the employee against a moving target. The first capability does not settle the effects of the second.

I return here to the month I worked more than 300 hours because the work felt meaningful and I believed the sacrifice would come back to me somehow. A new tool might have relieved some of the pressure. It might also have raised expectations about what I could deliver. The difference would have depended on the relationship around the tool, not only on the tool’s speed.

That is the practical meaning of the distributional question. It is not an abstract dispute about whether technology creates value. Of course it can create value. The question is whether the people participating in that creation gain more control over their lives, or whether the institution simply acquires a more efficient way to ask for more. A faster engine does not negotiate the destination.