Who Owns the Productivity?

Imagine that an employee has a task that usually takes eight hours. Research, drafting, checking, revisions, the whole day. Then the employee learns to use AI well and finishes the same work to the same standard in four hours. Nothing magical has happened. The person has simply become more productive: the same useful result now requires less time.

What happens to the other four hours? That question sounds almost childish. It is actually one of the largest economic questions of the AI era. The employee might do more work. The company might serve more customers. The customer might get a lower price or a faster answer. The worker might earn more. The worker might go home earlier. The company might need fewer new employees. Profits might rise. Some combination of those outcomes might occur.

AI does not answer that question. AI creates capacity. People and institutions decide what happens to the capacity.

Chapter Seven asked what happens when one person can produce what once required a team. This chapter asks who receives the value of that extra productivity. Economists call the extra value an economic surplus. It might become more pay, a lower price, a larger profit, or saved time. How it is shared depends on agreements, competition, ownership, public policy, and each party’s ability to negotiate.

When Delegation Adds Work

Before deciding who receives a productivity gain, we should establish whether the gain exists. An assistant can complete part of a task quickly while increasing the work required to finish the whole job.

I encountered that problem while trying to arrange a ride to my morning swim. I already knew how to compare Uber and Lyft and request the cheaper ride. I had been doing it for a year. The fares could change from minute to minute, so the useful service was a comparison close to departure followed by a booking within my budget and arrival deadline. I wanted to hand over that recurring job.

The assistant asked for my pickup location, spending limit and timing. Those were reasonable details for an assistant able to complete the assignment. But its ability to complete it had not been established. It found that Uber offered browser booking, could not verify a working Lyft browser-booking route, and encountered a failure while testing Uber sign-in. No ride was booked. Those were the limits of that attempt, not proof that either service could never be accessed by an assistant.

The proposed fallback was for me to open the apps, capture their live fares and send screenshots so the assistant could help me choose. That returned much of the original task to me and added another handoff. A separate sign-in option also asked me to resolve access before the assistant could establish whether it would work. I explained that I already knew how to do the job. I had hoped to delegate it.

This was a small, personal example rather than a controlled productivity study. I did not measure the extra time or a financial loss, and no more expensive ride was purchased. The experience nevertheless exposed a cost that a count of useful replies would miss: I was now managing the assistant's attempts to help while the original job remained unfinished.

Delegation fails when managing the delegate costs more than doing the task. Explanation, setup, checking, corrections and recovery all belong in the calculation. So does the human work that remains after the assistant has produced its answer. A business that counts only the time saved on a first draft can miss the time spent verifying it. A customer offered an automated service may end up supplying information repeatedly and completing the transaction elsewhere. These are possibilities to test, not savings to assume.

Some setup work can be worthwhile when it enables reliable help across many future tasks. The question is whether there is evidence for that return. Before drawing the user into a long sequence of instructions, an assistant should establish which parts of the job it can perform, which permissions it needs and where human action will still be required. It should disclose uncertain access and stop when further attempts are unlikely to improve the outcome. Required safeguards should remain in place; their cost belongs in the assessment of whether this is a useful task to delegate.

The efficient answer may be that the user's existing method is faster. An assistant should be willing to recommend that path promptly, even when it leaves the assistant with less to do. Continuing to offer workarounds can look helpful while shifting more work to the person who asked for relief. Nothing about that outcome requires a machine to have a motive to stay involved.

The test is straightforward: does delegation reduce the human's total burden, or merely change the work the human has to do? Compare the complete result with the existing method, including time, money, attention, quality and the consequences of failure. Only after that comparison shows a gain does the next question arise: who gets to keep it?

The Afternoon Test

A useful way to make the problem concrete is what I will call the afternoon test. If AI turns a full day of work into a morning of work, watch what happens to the afternoon. The destination of those hours tells you something important about who is capturing the productivity gain.

If the employee is given another assignment, the company captures most of the gain as additional output. If the employee keeps the same pay and leaves earlier, part of the gain becomes leisure. If the employee earns a bonus for producing more, part becomes compensation. If the company lowers prices, customers receive part of the gain. If the company uses the savings to invest, hire in a different area, or improve service, the gain is recycled into the business. If the company cuts positions or avoids planned hiring, part of the gain appears as lower labor cost.

None of these outcomes is built into the technology. That is the first principle of this chapter. Efficiency creates an opportunity. Distribution requires a decision.

The afternoon test also exposes why arguments about AI and work often pass each other without meeting. One person says AI is wonderful because a report now takes two hours instead of six. Another says workers are being squeezed because expectations rose as soon as the tool arrived. Both can be describing the same productivity gain. They are arguing about its destination.

Productivity Is Not Pay

Economists use productivity in several technical ways, but the basic idea is simple enough for ordinary life. Productivity asks how much useful output we get from a given amount of input. If a carpenter can build two cabinets in the time it once took to build one, productivity rose. If a call-center employee can solve more customer problems in an hour, productivity rose. If an analyst can produce a better report in half the time, productivity rose.

A productivity increase can make higher wages possible because each hour of work can support more economic value. Possible is the important word. Higher productivity and higher pay are different events. A company may share a gain through wages, bonuses, equity, benefits, shorter hours, better working conditions, or greater job security. It may also retain the gain through higher margins, use it to cut prices, spend it on new technology, pay suppliers, reduce debt, or reward shareholders.

Competition complicates the picture. If every firm in an industry gets the same AI advantage, prices may fall and customers can capture much of the benefit. If one firm gains the advantage first, that firm may keep more of the gain for a while. If workers with scarce AI-related judgment are hard to hire, they may capture more through wages. If the relevant skills become common, employers may have more bargaining power. The technology enlarges or rearranges the pie. Markets and institutions still decide the slices.

This is why the phrase “AI will increase productivity” is only the beginning of an economic argument. The next sentence has to ask: productivity for whom, measured how, and converted into what?

Five Places the Gain Can Go

Most AI productivity gains inside a business have five obvious destinations. They can go to workers, owners, customers, reinvestment, or time. The categories overlap, but they give us a useful map.

Workers can receive higher wages, bonuses, promotion opportunities, better tools, or more secure jobs. Owners can receive higher profits, dividends, or a more valuable company. Customers can receive lower prices, faster service, better quality, or products that did not exist before. Reinvestment can create new products, new markets, new equipment, and new jobs. Time can become shorter work, less overtime, fewer late-night emails, or simply more room to think.

There is a sixth destination that is less pleasant to name: displacement. A company can use productivity gains to produce the same output with fewer people or to avoid hiring people it once expected to need. Chapter Five examined that margin directly, and Chapter Six examined what it can do to the first rung of a career.

The point here is narrower. A company can become more productive while different people experience entirely different results. An investor may see higher earnings. A customer may see a faster answer. A senior employee may receive a raise. A junior applicant may encounter a job that no longer exists. One productivity number can hide several economic stories.

The Klarna Question

Klarna, the company this book has followed since Chapter Five, provides a useful running case because it has made productivity per employee part of its public story. Chapter Five used Klarna to examine labor without employment and the changing relationship between headcount and capacity. Here the same company illustrates a different question: once a firm produces much more revenue with each employee, where does the value go?

Klarna has reported rapid growth in revenue per employee while reducing headcount sharply from its 2022 level. The company attributes much of this change to AI-enabled productivity, alongside changes in its business and costs. It calls this operating leverage: revenue can grow without costs rising at the same pace. The changing figures appear in Current Evidence. The broader point is that business growth may require fewer additional employees than before.

That does not tell us the final distribution. Revenue per employee is not the same thing as profit per employee, and it is not a clean scientific measure of AI productivity. Revenue can rise because of pricing, customer growth, product mix, acquisitions, or economic conditions. Expenses can move for reasons unrelated to AI. A serious reader should resist the temptation to turn one corporate ratio into proof of a technological revolution.

Still, the ratio forces the right question. If a company can serve far more business with a smaller workforce than before, the old link between company growth and payroll growth weakens. The value that once automatically required another salary can now be divided in more ways.

The Worker Can Capture Time

Money is not the only way workers can benefit. A technology that saves time can return part of life to the person using it. That sounds sentimental until you remember that time is what employment actually consumes.

A field experiment across 66 firms found that employees who actively used a generative AI tool spent less time on email and reduced work outside regular hours. The researchers did not find a broad change in the overall quantity or composition of tasks during the experiment. All 66 firms were very large and mostly multinational, so the result is best read as evidence about large-company knowledge work. In plain English, some workers got part of the productivity gain back as a less intrusive workday rather than as a bigger paycheck.

That possibility matters because the eight-hour workday can hide productivity improvements. If a salaried employee finishes faster and the employer immediately fills every saved minute with another task, output rises while the employee’s day feels unchanged or more intense. If some of the saved time becomes breathing room, training, family time, or fewer evening messages, the worker receives a different kind of dividend.

There is no universal formula saying how much of a gain should become pay and how much should become time. Different jobs, industries, and competitive conditions make that impossible. But organizations should at least recognize that saved time is an economic asset. Pretending it has no owner simply guarantees that the strongest party in the arrangement will usually decide.

What Stays Human

Imagine finishing a task before lunch that used to consume most of an afternoon. The first draft is ready. The spreadsheet has been organized. You have checked the important figures and corrected the mistakes. For once, the next hour has no assignment attached to it.

What will you do with that hour?

The answer matters because time saved and time reclaimed are different things. An employer may fill the space with more work. You may spend it correcting another AI result. A task completed faster can leave the day just as crowded. But when some time does come back to you, there is a choice worth making before habit makes it for you.

You could attempt something you have postponed. You could learn a skill, talk with a customer, or give a difficult decision the attention it deserves. You could also close the laptop and spend the hour with someone you love. A useful technology should leave room for a life beyond producing more useful technology.

This is the personal side of the productivity question. Once we have asked who receives the gain, we can ask what the person receiving it wants that gain to make possible.

Experience gets another opportunity

My experience with AI has made that question concrete. In writing and publishing The Delegation Age, I can ask for research, compare arguments, develop drafts, and prepare material for readers. Ideas that would once have required more technical help or a larger organization are easier for me to pursue. I still have to decide which ideas deserve the reader's time and which claims I am willing to put my name behind.

That last decision draws on a lifetime outside the software. My years advising people on financial and business decisions matter because a recommendation has to fit a person and a situation. A calculation can be correct while the recommendation built around it is unsuitable. The missing fact may concern an obligation, a competing goal, or a risk the person cannot afford to take.

AI can help uncover those facts too. It can suggest a question I overlooked or challenge an assumption I brought into the conversation. Experience earns its value through being tested, rather than through being older than the tool. A confident human can overlook something important just as a confident machine can.

What encourages me is the chance to use more of what I have learned. I can spend less effort getting an idea into a form I can examine and more effort deciding whether it deserves to go further. That opportunity is available in different ways to a teacher preparing a lesson, a business owner considering a new service, or a writer trying to make a difficult argument clear.

Years of experience are useful material to bring into that work. So are the questions a beginner is willing to ask. Nobody needs a lifetime of expertise before taking a first step. The task is to choose a step whose result can be checked and whose consequences can be managed.

More options need a standard

Consider a small business owner deciding how to price a new service. This is an illustration, rather than an account of a particular business. AI produces three plans with revenue estimates and polished explanations. Preparing those alternatives may have taken minutes. Learning whether customers will accept the prices has barely begun.

The owner needs to check the assumptions. How many customers would actually buy? How much time does each sale require? Does the price cover the work that happens after the invoice is paid? A forecast can look attractive because it assumes repeat business that has never been demonstrated.

The owner can use AI to examine these questions, then try a limited offer with a few customers and compare the result with the forecast. The useful gain comes from reaching a test sooner. A generated plan becomes more informative when it meets something outside the conversation.

Writers face a similar choice. Ten versions of an opening paragraph may help a writer see possibilities. The writer still needs a standard for choosing. Does the paragraph say something true? Does it sound like the person writing the book? Will a reader understand why the subject matters? A version can be smoother and still weaken the argument.

Taste is the ability to recognize what suits the purpose. It develops through close attention to examples and through learning why one choice works better than another. Judgment extends that attention to evidence and consequences. A pleasing sentence can be inaccurate. An efficient business plan can damage a relationship the owner hopes to preserve.

AI can participate in both kinds of evaluation. It can compare drafts, expose inconsistencies, and propose tests. The person directing the work must still establish which standards matter, obtain help where needed, and answer for the decision. The boundary concerns authority and accountability as well as ability. We do not need to claim that machines can never exercise useful judgment to insist that a decision affecting people needs someone who can explain it and respond to its consequences.

Build judgment while using the tool

Advice to exercise judgment is incomplete unless people have a way to develop it. Someone entering a profession cannot simply borrow the experience of someone leaving it. Nor should a beginner be expected to repeat every inefficient task that an earlier generation performed. Chapter Six's discussion of training debt raises the practical question: how do we preserve learning while making the work easier?

One place to begin is before the AI answers. Write down your own view of the problem in a few sentences. What are you trying to accomplish? What do you expect the difficult part to be? Which assumption could change the decision? This takes little time on a modest task and gives you something independent to compare with the result.

Without that first view, it is easy to let the answer quietly define the question. A recommendation to reduce a price may draw your attention toward discounts when the actual problem is unreliable service. A request for a more persuasive message may produce better wording for a promise you should reconsider making.

Next, check the claim that carries the most weight. A recommendation may contain twenty details, while depending on one uncertain number. Find that dependency. For a research summary, read the original passage supporting the main conclusion. For a spreadsheet, test the formula that determines the outcome. For a plan, ask what happens if the central assumption fails.

The depth of checking should match the stakes. An invitation and a decision affecting someone's livelihood deserve different arrangements. If the decisive claim is outside your competence, involve a qualified person or use a reliable test. Your ability to delegate should include knowing when the work requires someone else's expertise.

Then explain the choice in your own words. Why this version? What did you reject? What would make you change your mind? If the explanation merely repeats the AI's confident language, return to the evidence. You need enough understanding to defend the decision to the person who will live with it.

Finally, revisit the outcome. Keep a short record of important decisions, including what you expected and what actually happened. A sales message can sound convincing and receive no replies. A lesson can look clear while leaving students unable to explain the idea. Feedback gives you an opportunity to improve the next assignment and the standard used to judge it.

These habits can fit inside ordinary work. A junior employee can explain an AI draft to a mentor. A business owner can review one recent decision each Friday. A writer can compare a revised passage with the original and identify what became clearer or less honest. The finished work matters, and so does the understanding gained while producing it.

Give some of the time a purpose

When you do recover time, it helps to decide what at least part of it is for. Otherwise, the easiest new task is likely to claim it. Generating another draft takes less effort than calling a customer whose answer may be uncomfortable. Producing another plan is easier than finding out why the first one failed.

Start with one recurring task and a modest experiment. If AI helps you finish a weekly report sooner, include the time spent checking and correcting it when estimating the saving. Choose a use for part of the time that remains. It might be a conversation with someone affected by the report, a period of uninterrupted learning, or an earlier finish to the day.

Employees may need to negotiate that choice with a manager. A team could agree that some saved time goes toward additional output and some toward training or improving the service. Owners have more direct control, but can still allow every gain to disappear into expansion. Having the authority to choose does not ensure that you will pause long enough to exercise it.

There is a particular opportunity in spending some of that time with people. A customer can explain why an apparently sensible offer feels wrong. A colleague can tell you where a process repeatedly breaks. A student can show the difference between completing an assignment and understanding it. AI may help prepare those conversations and interpret what you hear. Making room for the conversation remains a decision about how to use your day.

There is also room for work done slowly by choice. Read something without immediately asking for a summary. Try solving a manageable problem before requesting the answer. Write a page in your own voice and then ask what needs attention. Independent practice gives you material against which to judge the help you receive.

And some hours can remain free of professional purpose. Time with your children, a walk, or an evening without unfinished administration can be a worthwhile return. A person does not owe every saved minute to the next increase in output. Where circumstances allow, reclaiming time can be part of the goal itself.

A useful question before the next handoff

Before your next AI assignment, ask how you will know that the work helped. Name the result you want and the evidence you will use to assess it. Decide who can check the parts you cannot. Consider the cost of an error before giving permission for an action that will be difficult to reverse.

Add one more question: if this saves time, what would I like some of that time to make possible?

The answer can be small. Understanding one subject better. Having a conversation you have delayed. Testing an idea with a real customer. Finishing early enough to enjoy the evening. You do not need a grand plan to begin using a useful tool well.

What stays human is not a fixed inventory of tasks that technology will never reach. It is our obligation to decide what the work is for, to take the people affected seriously, and to remain answerable for what we authorize. AI can help us think through those obligations. Greater capability gives us more opportunities to act on them.

If a tool helps you complete the work sooner, pause before filling the space. There may be something you have wanted to learn, someone you have wanted to see, or an idea you have wanted to pursue. Give some of the recovered time to that. The gain becomes yours when it helps you do something you have chosen.

The Bargain Changes Before the Paycheck

The traditional employment bargain is easy to describe. The employer buys a person’s time and capability for an agreed form of compensation. The employee accepts direction, standards, and a degree of organizational control. Productivity improvements have always changed that bargain. Better machines, software, logistics, and management have repeatedly altered how much one worker can produce.

AI makes the adjustment unusually visible because the gain can arrive at the level of the individual task. A worker can know that a job that took three hours now takes thirty minutes. A manager can see the same thing. The measurement may still be rough, but the change is experienced directly.

That can create a trust problem. If workers believe every visible AI gain will become a higher quota, more surveillance, or a smaller team, they have a rational reason to hide how much time the technology saves. If managers believe employees will keep all the gains as private slack, they have a rational reason to measure more aggressively. Both sides can end up spending part of the productivity dividend policing each other.

A better bargain makes the destination explicit. Some organizations may share gains through bonuses. Others may reduce repetitive work and increase training. Others may promise that a portion of time savings will be protected for higher-value work rather than simply added volume. Some may use gains to lower prices or expand. The specific arrangement can vary. The important thing is that the arrangement exists in the open.

Current Evidence: Where the AI Dividend Is Landing

Reviewed September 10, 2026. Anthropic public benefit fund announcement added October 19, 2026.

The evidence in 2026 does not point to one universal destination for AI productivity gains. It points to several at once.

Klarna reported that revenue per employee reached nearly $1.4 million in the first quarter of 2026, about four times its 2022 level. In its prior quarter, the company said revenue had grown 104% since late 2022 while operating expenses fell 8% and headcount declined 49%. Klarna describes this as AI-led operating leverage. Those are company-reported figures, and revenue per employee should be treated as an operating signal rather than a pure productivity measure.

PwC’s 2026 Global AI Jobs Barometer offers a broader, though still advisory, signal. It reported that companies in the most AI-exposed sectors had stronger productivity growth than the least exposed group, while also showing faster growth in staff cost per employee and faster headcount growth. PwC labels staff cost per employee as wage growth. That measure can rise because employees are paid more, but it can also rise when a company employs fewer lower-paid junior workers. Given the Chapter Six evidence, some of the difference may reflect that workforce mix. The figure still says something useful about where value is accumulating; by itself, it does not prove that the same workers received raises. The broader finding also rejects the simple story that company productivity gains must come only from fewer workers. Some firms appear to be using greater productivity to grow.

A Danish study linking adoption surveys with administrative labor-market records found widespread chatbot adoption and worker-reported productivity benefits but no meaningful average effect on earnings or recorded hours during the first two years. The researchers could rule out effects larger than about 2%. Employers were nevertheless reorganizing work, including creating new tasks in AI oversight and AI integration, and adopters were moving toward higher-paying occupations where chatbots were more relevant, although not in numbers large enough to move average earnings. Productivity was changing before average pay or hours clearly moved. That is almost a laboratory version of this chapter’s argument: the gain can appear before its distribution is settled.

Anthropic’s survey of roughly 81,000 Claude users found that, among respondents who actually named a beneficiary of their productivity gains, most described benefits flowing to themselves through speed, wider capability, or freed time. About 10% said employers or clients were asking for and receiving more work. The benefit was not evenly reported across career stages: 60% of early-career respondents said they personally benefited, compared with 80% of senior professionals. The survey has obvious limits: it covered personal Claude users who chose to respond, and Anthropic used AI classifiers to interpret open-ended answers. It is useful as a signal about experience, not a population estimate.

An IMF working paper published in July 2026 looked at AI usage across countries and estimated the labor-cost equivalent of time saved at roughly $2.7 trillion a year across its sample, about 3.4% of combined GDP. The authors are careful: this is an indicative value of saved labor time, not a forecast that global GDP has already risen by that amount. They also found that the gains generated by current AI use tilt toward higher-paid occupations in nearly every country studied. In developing economies, nearly all of the usage-based value was concentrated inside a relatively small professional enclave. The gain is large enough to matter, and its distribution is uneven enough to matter separately.

On the day this chapter was drafted, Reuters reported a striking corporate example from Wipro. Chief technology officer Sandhya Arun said the Indian technology-services company’s AI initiatives had freed capacity equivalent to the output of about 20,000 employees, roughly 8% of a workforce of about 243,000. Wipro said those employees had been redeployed rather than removed, and more than 100,000 employees had received advanced AI training or certifications. Arun argued that the goal should move from raw productivity toward business outcomes. Reuters also quoted analyst Manoj Chandra Jha of Nord-IQ Research, who described Wipro as still in a cost-absorbing phase of AI monetization. That counterweight matters. The company’s account shows one possible destination for productivity gains, redeployment and reskilling, while the economics of the investment are still developing.

RocketAir offers a worker-side example with a different destination. The New York design and strategy firm has operated a four-day workweek since 2022. Founder and CEO Taylor Rosenbauer has described an internally developed AI tool as helping the team sustain the shortened schedule through faster data synthesis and other routine work. RocketAir’s four-day week predates its current AI use, so AI did not cause the policy. The shorter week is also demanding rather than leisurely: Rosenbauer told The Washington Post that the schedule creates constant pressure to deliver value. Even with that caveat, the company has publicly chosen to use AI-supported efficiency in a system that preserves a shorter workweek rather than simply converting every saved hour into more working hours. Its current careers page still lists the four-day workweek as a benefit. That is the afternoon test with a name attached.

On October 18, 2026, Anthropic announced a plan, subject to stockholder approval, to commit 5% of its shares to a public benefit fund. The fund would be a wholly owned company. Anthropic named disruption to workers as one possible use and said more details about governance would follow. Further share commitments would depend on company growth and board discretion.

Such a fund could spread some of AI’s financial gains beyond the people who own or use the technology. The promise still leaves a decision to be governed: who qualifies for help, who decides, and how can affected people question that decision? Setting money aside and giving people a voice are separate choices. Readers should look for published decision rules and reported results before treating a pledge as an effective answer to displacement.

On November 2, 2026, OpenAI published LifeVal, an evaluation of everyday tasks such as arranging childcare, managing household responsibilities and accessing government services. The company says it used time-use research and a survey of 3,995 people to identify where assistance would be most valuable. This is a provider-designed evaluation, not an independent demonstration that using an assistant reduces a person's total burden in daily life.

The useful next question is what happens outside the test. Can the assistant reach the required service, complete the authorized job and leave the person with less work overall? Setup, checking and recovery belong beside the result. A strong task score can justify trying an assistant; permission to act and evidence of a net benefit still have to be established in the actual workflow.

The Distribution Choice

The evidence should make us suspicious of easy slogans. AI productivity does not automatically enrich workers. It also does not automatically impoverish them. The same technology can support higher wages in one company, fewer hours in another, lower prices in a third, and layoffs in a fourth.

What determines the difference? Partly economics. A company with strong competition may have to pass gains to customers. A company fighting for scarce talent may have to pass gains to workers. A company with pricing power may retain more as profit. A young company may reinvest almost everything. An independent worker may capture the gain directly because a faster project creates room for another client. A salaried employee may see the same saved hour become another assignment.

Management choices matter too. Leaders decide whether a productivity target becomes a hiring freeze, a growth plan, a bonus pool, a shorter process, a lower price, or another layer of output expectations. Those decisions are usually presented as ordinary management. In the AI era, they are also decisions about who participates in the productivity dividend.

There is no requirement that every gain be divided equally, and equality is not the only measure of a good outcome. A company that reinvests heavily may create more jobs later. Lower prices can benefit millions of customers. Higher profits can finance risk and reward capital that made the investment possible. Higher wages can reward scarce skills. More free time can improve a worker’s life without appearing in GDP. The serious question is not which destination is always correct. It is whether the destination is visible enough to be debated honestly.

Accountability Analysis

The Delegation Doctrine does not contain a wage formula. It does contain a useful way to think about the decision.

Principle 1 says the owner remains the owner. If management authorizes AI to change how work is performed, management still owns the consequences of that operating choice. Saying “AI made us more efficient” does not explain why the gain became layoffs instead of growth, bonuses instead of price cuts, or more work instead of less overtime. Those are human choices layered on top of technological capability.

Principle 5 moves accountability upstream. Once AI can perform work faster than managers can inspect every step, accountability shifts toward the rules set before the work begins: productivity targets, staffing plans, quality thresholds, compensation systems, training budgets, and the conditions under which saved capacity becomes reduced headcount.

Principle 6 asks whether consequences reach the owner. A manager may receive credit for saving money while workers lose income or customers spend more time correcting mistakes. An organization pursuing efficiency should identify who benefits and who bears the costs of the change. Workers, applicants, customers, communities, and shareholders can experience the same decision very differently.

The practical test is simple. When an organization celebrates an AI productivity gain, ask where the gain went. Did output rise? Did price fall? Did wages rise? Did hours fall? Did quality improve? Was capacity redeployed? Did headcount shrink? Was the gain invested? Did the AI supplier capture part of it through fees? A productivity claim without a destination is only half an economic statement.

Follow one saved hour

Choose a task that AI helps you finish faster. Count the time needed to check and correct the result before deciding how much you saved. Then follow that saved time: did it become another assignment, training, a shorter day, or a benefit for the customer? Ask who decided. Use the afternoon test to make the destination visible.

Accountability Verdict

AI can make an hour of human work worth more. It cannot decide who receives the difference.

The productivity dividend can become wages, profit, lower prices, investment, better service, shorter hours, new jobs, fewer jobs, or some mixture of all of them. The destination will vary, and reasonable people can disagree about the best division. What should disappear is the fiction that the division happened by itself.

For employers, the obligation is to measure more than output. Track where the saved capacity goes and what happens to quality, pay, hours, training, staffing, and customers. For workers, the useful question is broader than “Will AI take my job?” Ask whether AI is increasing the value of your judgment, who can capture that value, and whether your employment arrangement lets you participate in the gain.

When AI gives back four hours, watch the afternoon. That is where the economic settlement is being written.

Chapter Nine turns from the division of gains inside work to a larger problem. What happens when access to the tools, skills, capital, and bargaining power needed to create those gains is unequal from the start?

Sources and Notes

1. Klarna Group plc, “Klarna Delivers Strong Start to 2026 With $1Bn Revenue and $68M Adj. Operating Profit,” May 14, 2026, and “Klarna Accelerates U.S. Growth and Delivers $1bn Revenue Driven by Rapid Banking Service Adoption,” February 19, 2026. https://investors.klarna.com/News--Events/news/news-details/2026/Klarna-Delivers-Strong-Start-to-2026-With-1Bn-Revenue-and-68M-Adj--Operating-Profit/ and https://investors.klarna.com/News--Events/news/news-details/2026/Klarna-Accelerates-U-S--Growth-and-Delivers-1bn-Revenue-Driven-by-Rapid-Banking-Service-Adoption/default.aspx. Accessed September 10, 2026. Klarna reported nearly $1.4 million in revenue per employee in Q1 2026, about four times its 2022 level. In Q4 2025 it reported revenue up 104%, operating expenses down 8%, and headcount down 49% since Q4 2022. Company source used for company-reported operating figures and management attribution. Source record: DA-CH8-S1. Archive reference: pending.

2. Anders Humlum and Emilie Vestergaard, “Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI,” NBER Working Paper 33777, May 2025, revised March 2026. https://www.nber.org/papers/w33777. Accessed September 10, 2026. Using adoption surveys linked to Danish administrative records, the authors report widespread chatbot adoption and worker-reported productivity benefits but precise null effects on average earnings and recorded hours, ruling out effects larger than 2% during the first two years after ChatGPT. Employers also reorganized tasks, including new work in AI oversight and integration, and adopters moved toward higher-paying occupations where chatbots were more relevant, though too few to shift average earnings. Source record: DA-CH8-S2. Archive reference: pending.

3. Eleanor W. Dillon, Sonia Jaffe, Nicole Immorlica, and Christopher T. Stanton, “Shifting Work Patterns with Generative AI,” American Economic Review: Insights, online article, DOI 10.1257/aeri.20250275; open working-paper copy: NBER Working Paper 33795, revised November 2025. https://www.aeaweb.org/articles?id=10.1257/aeri.20250275 and https://www.nber.org/papers/w33795. Accessed September 10, 2026. In a field experiment across 66 firms and 7,137 knowledge workers, 80% of treated workers who used the tool spent about two fewer hours on email each week in the second half of the experiment and reduced work outside regular hours. The study did not detect broad shifts in task quantity or composition from individual-level access. All 66 firms were very large and mostly multinational. The paper discloses that some authors were employed by Microsoft, maker of the tested M365 Copilot, while retaining discretion over results. Source record: DA-CH8-S3. Archive reference: pending.

4. PwC, “2026 Global AI Jobs Barometer,” June 15, 2026. https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html. Accessed September 10, 2026. PwC reports that companies in the most AI-exposed sectors showed higher productivity growth than the least exposed group and also faster growth in staff cost per employee and headcount. The report cites 24% growth in staff cost per employee versus 17%, headcount growth of 52% versus 36%, and productivity growth of 34% versus 24%, all from a 2018 baseline. PwC defines the wage measure as total staff costs divided by headcount, so it can rise because employees are paid more or because a workforce contains fewer lower-paid junior workers. This is employer-advisory research and is treated as a current signal, not causal proof. Source record: DA-CH8-S4. Archive reference: pending.

5. Maxim Massenkoff and Saffron Huang, “What 81,000 people told us about the economics of AI,” Anthropic Economic Index, April 22, 2026. https://www.anthropic.com/research/81k-economics. Accessed September 10, 2026. Among respondents who named a beneficiary of their productivity gains, most described benefits flowing to themselves through speed, expanded scope, or freed time; 10% said employers or clients were asking for and receiving more work. Sixty percent of early-career respondents said they personally benefited, compared with 80% of senior professionals. Only about a quarter of interviews named a recipient. The sample consists of personal Claude users who opted into the survey, and the analysis relies partly on Claude-powered classifiers. Vendor research is used only as a signal about user-reported experience. The author uses Anthropic products as part of the Living Edition’s external critique process; see the project transparency statement. Source record: DA-CH8-S5. Archive reference: pending.

6. Rachel Yuting Fan and Ha Minh Nguyen, “Aggregate Gains from AI and Their Distribution: Global Evidence from Usage Data,” IMF Working Paper 2026/147, July 10, 2026. https://www.imf.org/en/publications/wp/issues/2026/07/10/aggregate-gains-from-ai-and-their-distribution-global-evidence-from-usage-data-577586. Accessed September 10, 2026. Using five waves of Anthropic Economic Index usage data, the paper estimates an annual labor-cost equivalent of roughly $2.7 trillion, about 3.4% of combined GDP in the 86-country sample. The authors stress that this is an indicative value of saved labor time, not a direct GDP estimate. They find that current AI usage-based gains tilt toward higher-paid occupations in nearly every country studied and that, in developing economies, nearly all usage-based value is concentrated in a small professional enclave. IMF Working Papers represent research in progress and do not necessarily represent IMF views. Source record: DA-CH8-S6. Archive reference: pending.

7. Sai Ishwarbharath B and Abhirami G, “Wipro’s AI push frees capacity equivalent to 20,000 workers, CTO says,” Reuters, September 10, 2026. https://www.reuters.com/world/india/wipros-ai-push-frees-capacity-equivalent-20000-workers-cto-says-2026-09-10/. Accessed September 10, 2026. Wipro chief technology officer Sandhya Arun said AI initiatives had increased productivity by capacity equivalent to the output of 20,000 employees, about 8% of a workforce of roughly 243,000, and that the employees had been redeployed. More than 100,000 employees had received advanced AI-related training and certifications. Reuters also quoted Nord-IQ Research analyst Manoj Chandra Jha, who described Wipro as still in a cost-absorbing phase of AI monetization relative to peers. Source record: DA-CH8-S7. Archive reference: pending.

8. RocketAir, “RocketAir moves to a 4-day workweek,” March 27, 2023; Danielle Abril, “These companies say AI is key to their four-day workweeks,” The Washington Post, December 31, 2025; RocketAir Careers. https://www.rocketair.com/news/rocketair-moves-to-4-day-workweek, https://www.washingtonpost.com/business/2025/12/31/ai-four-day-workweek/, and https://www.rocketair.com/careers. Accessed September 10, 2026. RocketAir states that its four-day policy began in August 2022. The Washington Post reports that Rosenbauer said an internally developed AI tool helps the team deliver on the shortened schedule, while also describing the constant pressure to deliver value. Because the four-day week predates the current AI use, this is evidence of AI helping sustain a worker-time allocation, not proof that AI caused the policy. Source record: DA-CH8-S8. Archive reference: pending.

9. Anthropic, "Investing in the public benefit," October 18, 2026. https://www.anthropic.com/news/public-benefit-fund. Accessed October 19, 2026. Company announcement: initial 5% commitment is subject to stockholder approval; the fund is a wholly owned LLC; further equity allocations depend on valuation growth and board discretion; additional governance details remain forthcoming. Used for the narrow announced plan and structure, not as independent proof of completed transfers, effective worker relief or public control. Independent corroboration of the announcement: Reuters, "Anthropic to commit 5% stake to new 'public benefit fund'," October 18, 2026, https://www.reuters.com/business/anthropic-commit-5-stake-new-public-benefit-fund-2026-10-18/. The second paragraph is the author's analysis. Anthropic products are used in this Living Edition's external critique workflow, as disclosed in the existing chapter. Source record: DA-CH8-S9. Update 104. Archive reference: pending.

10. Author’s conversation with an AI assistant, November 1, 2026, Eastern time (November 2 UTC), summarized in DA_Delegation_Cost_v1.pdf. First-person experience and author analysis; not a measured productivity study. Exact pickup location and account details are excluded. Public-platform limitations are confined to the particular attempted workflow. Source record: DA-CH8-S10. Update 127.

11. OpenAI, “LifeVal,” November 2, 2026. https://openai.com/index/lifeval/ Accessed November 3, 2026. Provider-designed evaluation and account of task selection; not independent evidence of net time savings, representative field effectiveness or permission to act. The complete-workflow assessment is the author's analysis. OpenAI supplies the assistant used in this book's research, drafting and production workflow. Source record: DA-CH8-S11. Update 129. Archive reference: pending.