Where the Work Goes When AI Makes It Cheap

When a tool gets cheaper, people use far more of it. That pattern is now visible as artificial intelligence lowers the cost of cognitive work. Through the first half of 2026, employment and wages in the occupations most exposed to AI have been climbing, across the United States, Europe, and parts of Latin America. The total amount of paid work is growing. What is shifting is who gets to do it, and on what terms. The same force that is expanding work is also redistributing it.

Economists have a name for the first half of that sentence. In 1865 a British economist, William Stanley Jevons, noticed that as steam engines burned coal more efficiently, Britain consumed more coal overall, because efficiency made it worth using everywhere. Cheaper meant more. The same logic now reaches analysis, code, copy, legal review, and research: as the cost of producing them falls toward zero, the appetite for them rises. That is the reassuring half, and it is real. The harder question is the second half. When the work expands, where does it actually go?

Start with what the data shows

The jobs apocalypse that filled headlines in 2025 has not arrived in the numbers. Global unemployment sat at roughly 4.9% through 2026, according to the International Labour Organization, close to a historic low. In the United States, total employment held near record levels with unemployment around 4.3%. The European Union recorded its highest employment rate on record in early 2026. A fixed-pie collapse, the kind where machines simply absorb the work and leave people idle, would already show up in these series. It does not.

Look closer at the people most exposed to AI and the picture gets more encouraging. Vanguard's research, published in December 2025, found that the United States occupations most exposed to AI grew employment by 1.7% over the two years to mid-2025, faster than the 1% pace before the pandemic, while real wages in those same occupations rose 3.8%, against barely 0.1% before. Exposure to AI, so far, looks more like a wage premium than a layoff.

The clearest single rebuttal to the fear comes from the field that was supposed to fall first. Software development is where AI writes the most capable output, and the prediction was that programming jobs would be the first to fall. Instead, Microsoft's global diffusion report in May 2026 found United States software-developer employment about 4% higher than a year earlier, a record, even as code-generation tools spread through the profession. When the cost of writing software falls, companies build more of it, and they need more people to direct the building.

This is the mechanism Torsten Slok, chief economist at Apollo, has been pointing to: make legal work, accounting, consulting, or analysis cheaper and you expand the demand for all of it. History backs the shape of the claim. The economist David Autor has shown that around 60% of the jobs Americans do today did not exist in 1940. After every major technology shock, new kinds of work have appeared. That is the rule. The fear that there is a fixed amount of work to go around, and that each machine permanently removes a slice, has been wrong at every major turn for two centuries.

So the optimists have the aggregate, and they have it honestly. If the story ended here, the advice would be simple: adopt, relax, wait for the new jobs. The story does not end here.

The gains have an address

Growth in total work says nothing about who receives it. And when you sort the 2026 data by age, the mood changes.

Researchers at Stanford's Digital Economy Lab, working from the payroll records of the largest United States payroll processor rather than from surveys, found that workers aged 22 to 25 in the most AI-exposed jobs saw their employment fall about 16% relative to older colleagues over roughly three years, a figure they revised upward from an earlier estimate. Workers over 30 in those very same jobs kept growing, by 6 to 13%. The decline is concentrated at the entry rung, and it is sharpest where the tools are strongest: junior software developers in that age band fell close to 20% from their late-2022 peak.

One study can mislead. Two studies using different methods, reaching the same conclusion, are harder to dismiss. A separate Harvard analysis, built from company hiring and résumé data covering some 285,000 firms, found that junior employment at companies adopting generative AI fell 7.7% within six quarters compared with firms that did not adopt. The authors gave it a precise name: seniority-biased technological change. The effect is concentrated, and it lands at the bottom of the ladder.

The wage data points the same way. PwC's analysis of more than a billion job advertisements across 27 countries found that workers with AI skills now command a 62% pay premium over peers in the same role, up from 56% a year earlier. The benefit is real, and it is collecting in the hands of people who already have experience and scarce skills. Meanwhile, underemployment among recent United States graduates reached 42.5% in late 2025, a twelve-year high. The exposure is not evenly shared either. The ILO found that 29% of female-dominated occupations are highly exposed to generative AI, against 16% of male-dominated ones, because clerical and administrative work, long a route into the workforce for women, sits squarely in the technology's path.

Put plainly: the expansion is real, and so is a squeeze, and they are landing on different people. The gains flow to experience, scarce skills, and whoever already holds a seat. The pressure falls on the people trying to get their first one.

There is a trap inside this for any business reading the entry-level decline as good news. Senior people are made, slowly, out of juniors who were given real work, allowed to make real mistakes, and coached through them by someone who had done it before. If the cheapest way to handle junior-level tasks in 2026 is to hand them to AI, the entry pipeline thins. And the senior generalists every company will be competing for in 2030 are the juniors nobody is training in 2026. Cutting the entry rung is a saving today and a bill later.

The honest doubt: a productivity engine no one can see yet

There is a hole in the optimist case, and intellectual honesty means naming it plainly.

The whole argument that new work will replace displaced work rests on a productivity boom: AI makes people far more productive, that surplus funds new products, new demand, and new jobs. The trouble is that the boom is loud at the level of the individual task and almost silent at the level of the whole economy. Controlled studies are striking. The economist Erik Brynjolfsson found that AI assistance raised worker output by about 15% on average, and by 34% for novices and the lowest-skilled, while barely helping seasoned experts. Yet aggregate productivity has barely moved. Daron Acemoglu, who won the 2024 Nobel in economics, estimates AI will add only about 1.1 to 1.6% to GDP over a decade, with roughly 5% of tasks automated. The engine that is supposed to pay for the new jobs has not yet shown up in the macro numbers.

This gap matters because of what companies are doing in the meantime. A Harvard Business Review analysis in January 2026 found that most layoffs attributed to AI are being made on the basis of what executives expect AI to do, well ahead of what it has measurably done. Firms are restructuring around a promise. And the promise is not reliably paying off: Gartner's 2026 work found that cutting staff did not, on average, produce better returns on AI at all. Around 80% of organisations that piloted AI reduced headcount regardless of whether the technology was working, and what actually separated the high-return adopters was investing in their people rather than shedding them. Acemoglu has a phrase for the danger here: "so-so automation," where firms cut labour without the productivity transformation that would have justified it. The worst of both worlds, in which workers absorb the cost and the gains never materialise.

Keep one asymmetry in view whenever you read a confident forecast. The displacement is observed. The net job creation is forecast. The decline in junior hiring, the occupations contracting at the United States Bureau of Labor Statistics, the AI-attributed cuts tracked by Challenger, Gray and Christmas at about 49,000 so far in 2026: those are measurements of things that have already happened. The reassuring counterweights, the World Economic Forum's projection of 78 million net new jobs by 2030 and similar figures, are models of things that have not. Both can be cited in good faith. Only one of them has happened yet.

We have been here, partly

The optimists reach for history, and history mostly supports them, but with a footnote large enough to change the advice.

The favourite example is the cash machine. Automated tellers spread across the United States from the 1980s, and bank-teller employment rose, because cheaper branches meant banks opened more of them. It is a real complementarity, though it owed as much to 1990s banking deregulation as to the machines. The cleaner case is the spreadsheet. After it arrived, the number of bookkeeping and accounting clerks fell by roughly 400,000 in the United States over four decades, while the number of accountants and auditors rose by around 600,000. The work moved upmarket, from manual calculation to judgment, advice, and analysis. This is the heart of why the fixed-pie fear keeps failing: technology that kills a task often grows the profession around it.

Now the footnote. Aggregate recovery and individual recovery are not the same claim, and the cases where they came apart are the ones that should give an optimist pause. When trade with China hollowed out American manufacturing towns from 2000, the economy as a whole adjusted, but the displaced workers themselves carried earnings losses that persisted for more than twenty years, with little of the reallocation textbooks promise. Agricultural mechanisation eventually enriched everyone and permanently ended a way of life for millions of farm workers, in the United States, in Britain, in Brazil, where agriculture fell from 60% of employment in 1950 to about 8% today. "The economy adjusts" and "the people who were displaced are fine" are different sentences. The first has almost always been true. The second often was not.

And AI may strain the old adjustment mechanism in three specific ways. It moves faster than past technologies, in months and years rather than the decades it took electricity or the assembly line to diffuse. It hits many sectors at once rather than one at a time, so a displaced worker cannot simply step sideways into an untouched industry. And it reaches cognitive work, the very ground that displaced manual workers were always told to retreat to. The historical pattern is genuine cause for hope. It is not a guarantee, and it never came free.

What is left when execution is nearly free

Step back from the data for a moment, because there is a deeper shift underneath it.

The cost of producing a unit of cognitive output has fallen by something like 99% in three years. A workload that cost thousands of dollars a month to run through these models in 2023 now costs a small fraction of that. When execution becomes nearly free, it stops being the scarce thing, and value moves to whatever is still scarce: judgment about what is worth doing, taste in how it is done, accountability for the result, and the relationships through which work actually gets won and trusted. A job is a bundle of tasks someone pays you to complete. Work, in the older and larger sense, is the act of deciding what should exist and taking responsibility for bringing it about. AI is very good at the first and, so far, cannot do the second at all.

This is why the investor Naval Ravikant has floated the idea of a billion-dollar company run by four or five people. Whether or not the exact number holds, the bet underneath it is serious: that AI collapses the value of headcount as a measure of capability, and rewards small groups with strong judgment and outsized output. There is an optimistic edge here that the gloom misses. The same Brynjolfsson finding that unsettles the entry-level picture, that AI helps novices far more than experts, also means the technology can narrow the skill gap, handing a beginner a large part of what used to take years to acquire. Used well, that widens access to good work rather than closing it.

The people building these systems no longer agree on where this goes, and their disagreement is the most honest signal available. Sam Altman, who runs OpenAI, said in May 2026 that the company had been "pretty wrong" about an imminent jobs apocalypse. Yann LeCun, Meta's chief AI scientist, frames the near future as "AI as the employee, humans as the boss." Others inside the labs still warn of steep white-collar losses. When the people closest to the technology split this sharply, the appropriate posture for everyone else is to prepare for both.

What this means for you

Forecasts are not a plan. Here is what the evidence supports doing now, whatever your seat.

If you run a company, stop letting your AI strategy be a headcount number. The market has started to treat AI-justified layoffs as a confession rather than a credential: through 2026, several large companies that framed job cuts as an AI strategy were marked down hard by investors, while the ones rewarded were using the same technology to produce more. Point the saving at output, not just at cost. And keep hiring juniors while almost no one else is, at today's wages and against little competition, because you are buying the experienced talent the whole market will be bidding for in a few years. Above all, build the business on what compounds and cannot be rented: judgment, trust, proprietary data, relationships. We argued in defensibility in the AI era that models commoditise while those assets appreciate, and a falling price of AI only sharpens the point. If your only advantage is access to a model your competitor can also rent by the month, you do not have an advantage.

If you work for yourself, this is the expansion to ride, with one condition. AI lifts beginners most, which means a solo operator can now do work that recently required a team. But access to the tool is not a moat, because everyone has the same access. The differentiation is in the parts a buyer pays a person for: the judgment to frame the right problem, the taste to do it well, the accountability to stand behind it. That advantage is genuine. The four-or-five-person company with real reach is no longer a fantasy. The people who capture it are the ones who treat cheap AI as an input to something it cannot manufacture on its own.

If you are early in your career, the honest news is that the first rung got narrower, and the honest response is to get onto it anyway and then climb fast. The roles most protected are the ones where you direct AI rather than compete with it. Aim to be the person who directs the tool rather than the task it absorbs, and get inside a place where real work and real feedback are still on offer.

And if you are reading this from outside the United States, the picture has its own shape. An operator in Madrid, São Paulo, or Mexico City who rents AI by the month carries none of the risk sitting on the balance sheets of the companies building it. The capability is the same; the exposure is far lower. Real demand is opening: specialist wages across Latin America have been rising at double-digit rates, and remote hiring from the United States has surged. But those gains route to a narrow, often English-speaking, highly skilled tier, while only about 1% of global AI investment reaches Latin America at all. Outside the wealthy economies, the skill investment matters even more, because the gap between those who can direct this technology and those who cannot is widening fastest where the support to cross it is thinnest.

The decision the data leaves you

The pie is growing. It is also splitting. Both halves are true at once, and refusing either one is how people end up surprised.

The reassurance is earned: there is no sign in the 2026 data of work disappearing in aggregate, and good reason, from the shape of the technology and the weight of history, to expect the appetite for cognitive work to keep expanding. The warning is earned too: the expansion is reaching experience and scarce skill far more than the entry rung, the productivity boom that is meant to fund the new work has not yet appeared, and the people displaced along the way have not always been made whole, even when the economy was.

What this leaves you with is not a forecast to wait on but a decision to make. The force is the same for everyone. Where the work goes depends on what you build on top of it. The operators who do well from here will use cheap, abundant AI as fuel for something a falling price cannot commoditise: the judgment to know what is worth doing, and the trust to be the one who does it.