A leader stands up in a board meeting and reports that an AI deployment lifted productivity fifteen percent. The room accepts it. Someone asks about rollout to the next function. Nobody asks the question that determines whether the number means anything.
Where did the fifteen percent go?
Not rhetorically. Literally. Did it become margin, or did competitive pressure push it into price? Did the model vendor take it back in licensing? Did it fund headcount reduction, get absorbed by implementation cost, or turn into capacity the business spent on growth? Did any of it reach the people whose work the system learned from?
Most AI business cases cannot answer that, because they were never built to. They end at hours saved. That is the midpoint of the decision, not the end of it.
Productivity is not prosperity. That is not an argument against productivity — producing more useful output from the same resources creates real economic value, and the gains here are real. It is an argument about measurement. A productivity result tells you something changed in production. It does not tell you who received the benefit, whether the benefit became income or wealth, or whether the people whose work generated the underlying capability stayed economically secure.
Those are separate questions. Treating them as one is the most common analytical error in AI strategy right now, and it is being made at board level.
What the Evidence Actually Shows
The empirical record on AI and work does not tell one story, and anyone claiming it does is selling something.
Erik Brynjolfsson, Danielle Li, and Lindsey Raymond studied 5,172 customer-support agents and found that access to generative-AI assistance raised average productivity by 15%, with the largest gains concentrated among less-experienced and lower-performing workers. That is strong evidence of value creation. It is not evidence that wages rose fifteen percent, that workers received ownership, or that the gain was shared in any particular proportion among employees, the deploying firm, customers, and the AI supplier. The study measures production. It was never designed to settle distribution.
Other settings show why that distinction carries weight. Daron Acemoglu and Pascual Restrepo found that greater exposure to industrial robots reduced both employment and wages in affected U.S. labor markets. Xiang Hui, Oren Reshef, and Luofeng Zhou found that after generative-AI tools were released, freelancers in highly exposed occupations on a large online labor platform saw employment and earnings fall. In Denmark, Anders Humlum and Emilie Vestergaard found something different again — early chatbot adoption changed tasks and workflows with no detectable effect on earnings or recorded hours in the first two years after ChatGPT's launch.
Augmentation. Substitution. Loss. Short-run stability. Four credible studies, four different outcomes, because contract structure, labor-market institutions, bargaining power, task composition, and time all mattered.
The pattern worth extracting is not a prediction. It is that the production result never determined the distribution result. Something else did.
Three Questions, Not One
When an organization deploys AI, there are three distinct questions on the table, and almost every business case collapses them into the first.
What value was created? Did output, quality, speed, capacity, or willingness to pay improve relative to the resources consumed?
Who initially captured it? Did the gain show up as lower operating cost, wider margin, supplier revenue, lower prices, higher quality, or additional capacity?
How was it distributed? Did workers receive wages, bonuses, security, mobility, or ownership? Did customers get lower prices or better service? Did the vendor retain it as fees or rents? Did the public receive taxes or other returns?
None of these can be inferred from another, and the inference errors run in both directions.
A reduction in labor expense does not prove owners retained an equal amount — competition may pass some to customers, the model vendor may collect it in licensing, implementation cost may absorb it, and the business may spend the capacity on growth rather than banking it. Run the logic the other way and a faster worker does not become a better-paid worker. The employment contract determines what the worker is owed. The productivity metric has no view on it.
Value Does Not Travel Without a Mechanism
For created value to become a person's income, wealth, security, or economic power, something has to transmit it. A labor market. An employment contract. Collective bargaining. Profit sharing. Ownership. A license. A public benefit. A statutory claim.
This is not a new theory about AI. It is the ordinary operation of labor economics, contract, ownership, and public finance — the same machinery that governed every prior technology shift. The mistake is assuming the machinery runs automatically because the productivity number is large.
It matters most where AI substitutes for labor rather than complementing it. Task-based economics already explains how automation displaces some human tasks while new tasks reinstate demand for labor elsewhere; Acemoglu and Restrepo's account of automation and new tasks is the canonical treatment. But "new work will emerge" is a weaker claim than it sounds. It does not establish that comparable work emerges for the same people, in the same places, at the right time, with similar pay and security. Aggregate reinstatement is consistent with individual ruin.
Labor adaptation is one route to participation. It is not the only route, and it is not guaranteed.
The Channels That Actually Carry Value
There are five established ways people participate in economic gains. Each one works. Each one has boundaries that decide who it reaches.
Work and wages. People move into complementary or newly created tasks, raise their productivity, and bargain for part of the resulting value. This is where skills visibility, career pathways, internal mobility, and targeted development do real work — organizations cannot move people toward demand they cannot see or describe. But capability does not set compensation. Bargaining position does.
Profit sharing and employee ownership. Workers hold an explicit claim on enterprise results. The French evidence is unusually clean: Elise Nimier-David, David Sraer, and David Thesmar found that a mandatory profit-sharing rule raised labor's share by 1.8 percentage points and cut the profit share by 1.4 points, increasing total compensation for lower-skilled workers, with small to nonexistent effects on investment and productivity. The allocation changed because a rule created a claim — not because productivity found its own way to workers. Employee ownership is likewise associated with additional worker wealth. But firm-linked ownership carries boundaries: eligibility usually requires working for a participating employer, vesting ties the claim to tenure, and employer stock concentrates a person's job risk and investment risk in the same organization. Angelina Grigoryeva's recent U.S. work finds unequal access to stock-based compensation across jobs and demographic groups, and cautions against reading the associated wealth differences as automatically causal.
Broad capital ownership. Diversified retirement accounts, investment funds, public funds, and other collective structures let people participate without depending on one employer's equity program — though access to capital ownership is itself unequal. Benjamin Moll, Lukasz Rachel, and Pascual Restrepo's model of uneven growth shows why this becomes decisive: when automation raises returns to wealth while wages stagnate, ownership rather than work determines household outcomes.
Contribution-based claims. People could be compensated for the data, knowledge, creative material, and feedback used to build and improve reusable AI systems. Imanol Arrieta-Ibarra and colleagues, including Jaron Lanier and Glen Weyl, proposed treating data as labor in 2018 and named the monopsony and collective-action problems standing in the way. What is still missing is operational: systems that reliably trace material contributions, establish enforceable rights, value them at workable cost, and produce recurring returns. Treat this channel as a live hypothesis, not an available answer.
Public participation. Taxes, transfers, services, and returns from publicly held assets distribute value beyond current employees and owners. Alaska's permanent dividend is not AI policy, but it answers a question people keep asserting without evidence: Damon Jones and Ioana Marinescu found the dividend produced no significant reduction in aggregate employment and raised part-time work by 1.8 percentage points. Recurring non-wage income and continued work are not inherently incompatible. That does not settle how an AI-linked dividend would be funded or designed. It establishes that the design space is real.
Customers also benefit — lower prices, better quality, wider access, more variety. Those gains are genuine. They are not a substitute for income, wealth, security, or decision-making power, and they should not be allowed to quietly stand in for them in a business case.
The Participation Test
The useful question about a participation mechanism is never whether it sounds good. It is whether it creates a real claim, and whom that claim actually reaches.
Five tests make that visible:
- Eligibility. Who qualifies, and who is excluded?
- Enforceability. Is the benefit contractual, legally protected, market-based, discretionary, or merely promised?
- Portability. Does it survive job loss, separation, relocation, or a change in technology?
- Timing. Does it reach people during displacement, or only after a long adjustment?
- Concentration risk. Does it diversify a person's economic position, or bind more of it to a single employer, platform, or asset?
Run any mechanism through these and the coverage gaps surface immediately. Profit sharing performs well for retained employees and reaches no contractors or former workers. Employer equity builds wealth for vested employees and delivers nothing to people whose roles disappear before the cliff. Retraining works where adjacent demand exists and fails where the new opportunity is geographically or institutionally out of reach.
A mechanism can succeed completely for its recipients and still fail as an answer to AI-driven displacement. Average benefit is not coverage.
What It Costs to Skip This
The reason to run this analysis is not ethical positioning. It is that the alternative produces bad decisions with a delay.
You report a return you did not earn. If the business case books a productivity gain as margin without tracing where it landed, the board is being told a number that competitive pass-through, vendor pricing, or implementation drag already consumed. That error compounds across a portfolio and surfaces at exit, when a buyer diligences the value creation plan against actual EBITDA.
You lose the people who generated the gain. The Brynjolfsson result concentrated benefits among less-experienced workers. If the gain is captured entirely at the firm level while the workers carrying it see no change in pay, security, or path, you have built a retention problem in exactly the cohort that made the deployment work — and you find out at replacement cost.
You inherit a governance exposure you never priced. Compensation design, worker classification, and displacement handling are all regulated, and the regulation is moving. A deployment that quietly shifts task composition without touching the contract is a live exposure in labor negotiation, in litigation, and increasingly in disclosure.
You cannot defend the deployment when someone asks. Employees, regulators, customers, and acquirers all eventually ask where the value went. Not having an answer is itself an answer.
The Value Flow Statement
An AI business case should not end with hours saved and tasks automated. It should show a value flow, and it should name owners.
- What changed in production, measured against a real baseline — function lead
- Which costs, revenues, risks, or capabilities moved as a result — CFO
- Who holds a claim on the resulting value, and by what mechanism — CFO with CHRO
- Who is excluded from that claim — CHRO
- What survives separation, displacement, or a change in technology — CHRO
- Which of these outcomes is measured rather than assumed — operating partner or transformation lead
Three places to put it, in order of leverage. In the investment committee memo and value creation plan, so the productivity thesis carries its distribution assumption where it can still be interrogated. In the post-close hundred-day plan, where deployment decisions and compensation design are still simultaneously open — after that they diverge and rarely come back together. In the quarterly business review, as a standing line rather than a special topic, because the entire failure mode is that nobody asks.
Reading existing artifacts differently gets you most of the way there this quarter. Take any AI business case in front of you now and find the line where productivity becomes financial return. In most of them, that line is an assumption with no mechanism behind it. That is the gap.
One governance change makes it durable: give the compensation committee explicit standing over AI deployments that materially change task composition. Not approval rights over technology — visibility into deployments that change what people do, before the compensation consequences are locked.
Possibility Is Not a Recipient
AI productivity creates the possibility of prosperity. It does not specify who receives it.
That is not an argument against deploying AI. It is an argument against treating deployment as the end of the decision. The productivity number is the beginning of a value question, and organizations that stop there are not being optimistic. They are declining to measure the part that determines whether the investment produced anything worth having.
Skills intelligence, career pathways, and workforce planning matter more in this environment, not less — but their role is precise. They make the changing structure of work visible early enough to decide well. They do not determine how gains become wages, ownership, prices, or public benefit. That is set by compensation design, governance, market structure, contracts, and institutions.
Which means it is set by decisions. Someone's decisions. If leaders want AI to produce broad human benefit, they have to design the path from productive capacity to economic participation and then measure it, because nothing in the technology does it on its own.
Fifteen percent went somewhere. Find out where.
Sources
Acemoglu, D., & Restrepo, P. (2019). Automation and new tasks: How technology displaces and reinstates labor. Journal of Economic Perspectives, 33(2), 3-30. https://doi.org/10.1257/jep.33.2.3
Acemoglu, D., & Restrepo, P. (2020). Robots and jobs: Evidence from US labor markets. Journal of Political Economy, 128(6), 2188-2244. https://doi.org/10.1086/705716
Arrieta-Ibarra, I., Goff, L., Jiménez-Hernández, D., Lanier, J., & Weyl, E. G. (2018). Should we treat data as labor? Moving beyond "free." AEA Papers and Proceedings, 108, 38-42. https://doi.org/10.1257/pandp.20181003
Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889-942. https://doi.org/10.1093/qje/qjae044
Buchele, R., Kruse, D., Rodgers, L., & Scharf, A. (2009). Show me the money: Does shared capitalism share the wealth? (NBER Working Paper No. 14830). National Bureau of Economic Research. https://doi.org/10.3386/w14830
Grigoryeva, A. (2025). The shift to stock-based compensation and gender inequality in wealth in the United States. Socio-Economic Review, 23(2), 541-566. https://doi.org/10.1093/ser/mwaf009
Hui, X., Reshef, O., & Zhou, L. (2024). The short-term effects of generative artificial intelligence on employment: Evidence from an online labor market. Organization Science, 35(6), 1977-1989. https://doi.org/10.1287/orsc.2023.18441
Humlum, A., & Vestergaard, E. (2025). Still waters, rapid currents: Early labor market transformation under generative AI (NBER Working Paper No. 33777). National Bureau of Economic Research. https://doi.org/10.3386/w33777
Jones, D., & Marinescu, I. (2022). The labor market impacts of universal and permanent cash transfers: Evidence from the Alaska Permanent Fund. American Economic Journal: Economic Policy, 14(2), 315-340. https://doi.org/10.1257/pol.20190299
Moll, B., Rachel, L., & Restrepo, P. (2021). Uneven growth: Automation's impact on income and wealth inequality (NBER Working Paper No. 28440). National Bureau of Economic Research. https://doi.org/10.3386/w28440
Nimier-David, E., Sraer, D., & Thesmar, D. (2026). The effects of mandatory profit-sharing on workers and firms: Evidence from France. The Quarterly Journal of Economics, 141(3), 2205-2267. https://doi.org/10.1093/qje/qjag022