What AI does to work

Stand at a bank window in 1985 and list what the person behind it does. Counts notes into a customer's hand. Takes a deposit and keys it in. Checks a signature against a card. Explains, for the third time that morning, why the account is overdrawn. Talks a furious man out of closing it.

Five things, and a machine standing in the same lobby could already do the first two.

Nobody automates a job. A job is a bundle of tasks held together by a single employment contract, and a machine arrives for some of the tasks in the bundle. What happens to the bundle afterwards is a different question, with its own answer, and the answer is not supplied by the machine. Almost every confusing thing written about AI and employment comes from collapsing those two questions into one.

The bank's answer, and why it expired

What happened to tellers is documented, and it is not the story either side of the argument expects. The economist James Bessen set the numbers out in 2015. In the United States, between 1988 and 2004, the number of tellers needed to run a branch in the average urban market fell from twenty to thirteen. Over the same years the number of urban branches rose 43 per cent. The machine did exactly what it was bought to do, and the count of teller jobs did not fall.

The reason is the whole lesson, and it is the part that gets dropped. Cheaper branches were worth opening. Banks were fighting each other for market share, so when the cash machine cut what a branch cost to staff, they spent the saving on more branches instead of pocketing it. Meanwhile the tasks the machine could not touch became the valuable part of the work. In Bessen's words, cash handling became less important and human interaction more important. The teller who kept the job was now expected to sell.

That story gets repeated as proof that automation creates work. It proves nothing of the kind, because every step of it was contingent on something other than the technology. It ran on a bank's appetite for branches, and that appetite is gone. The US Bureau of Labor Statistics counted about 339,200 tellers in 2025, and says branch numbers have been falling because of technological change, as customers moved to banking on their phones. The cash machine never decided whether tellers kept their jobs. What decided was whether cheaper branches made banks want more branches, and that was never a question about the machine.

Bessen points at the nineteenth century for the same mechanism running the other way. Power looms took over 98 per cent of the labour needed to weave a yard of cloth, and the number of factory weaving jobs went up. Cloth got cheap, cheap cloth sold in volumes nobody had planned for, and volume needed weavers even at a fraction of the labour per yard. The surviving skill was no longer the weaving. It was coordinating work across several looms at once, and over the late nineteenth century weavers' wages rose sharply against other workers'.

Demand can also simply fail to arrive. Where people already have as much of something as they want, making it cheaper sells no more of it, and the labour saved is just saved. Nothing in the technology tells you which case you are standing in. That is the first reason predictions miss: the forecaster is studying the machine, and the answer is in the market.

The number everyone quotes

In September 2013 two Oxford researchers, Carl Benedikt Frey and Michael Osborne, published an estimate for 702 detailed occupations of how susceptible each was to being done by computer. About 47 per cent of total US employment landed in their high-risk category. That figure has been in circulation ever since, almost always as the claim that half of all jobs are about to disappear.

The paper says otherwise, in a sentence rather easier to find than the number. We make no attempt to estimate how many jobs will actually be automated. What they measured was the share of employment that could potentially be substituted by computers, from a technological capabilities point of view, over what they called some unspecified number of years, perhaps a decade or two. That is an engineering judgement about tasks. It contains no economics and no clock. It was never a forecast, and it was quoted as one for a decade.

Then there is the distance between what a thing can do and how much of it anyone has actually installed. The US Census Bureau started asking firms directly, and between September 2023 and February 2024, months when this was the loudest subject in the world, the share of American businesses using AI in producing goods or services rose from 3.7 to 5.4 per cent. The same survey found that firms using it often did so to substitute for worker tasks, and that few of them reported cutting employment because of it.

That gap between what exists and what has spread has a name, and now you have somewhere to put it: economists call it diffusion. Capability and diffusion are different curves and they can run years apart. Most of what a technology will eventually do to work has not happened yet at the moment everyone begins arguing about it, which is also the moment the argument is least informative.

What has actually been measured

The measurements that exist are real, and every one of them is narrower than the way it travels.

One much-quoted productivity result comes from a controlled experiment reported in 2023: developers given an AI assistant finished 55.8 per cent faster than developers without one. Look at the task before spending the number. The developers were recruited for the study, and what they were asked to do was implement an HTTP server in JavaScript as quickly as possible. That is a self-contained exercise with a known shape and a finish line the experimenter drew. No existing codebase to fit into, and nobody else's opinion to satisfy. An earlier page put a launch video and a deployment on either side of the same gap. A productivity experiment sits on the launch-video side. It measures something honestly; that something is not a working week.

A second result travels far less well, because its headline is an average that describes nobody. Researchers followed 5,172 customer support agents at a company introducing an AI assistant and found issues resolved per hour up about 15 per cent. Underneath that, the least experienced and lowest-skilled agents got both faster and better, while the most experienced and highest-skilled got slightly faster and slightly worse. One number, two opposite experiences. If you were the best agent on that floor, the tool was worth roughly nothing to you, and the average reported that you had gained.

Evidence about employment itself is thinner and much newer. In November 2025 three Stanford researchers reported six findings drawn from payroll records held by the firm ADP. In the occupations they rank as most exposed, workers aged 22 to 25 showed a 16 per cent employment decline measured against comparable workers elsewhere, while employment for experienced people in those same occupations stayed stable. The split inside that result is the useful part. Employment fell in the occupations where the software does the task, and the authors report only muted effects where it assists the person doing it — young workers' employment changes, they say, are not ordered by how exposed an occupation is to being assisted. One direction shows a decline and the other shows no clear pattern at all, which is a weaker and more honest finding than a mirror image would have been.

Notice what kind of claim that is. It is a comparison between groups inside one payroll provider's records rather than a count of jobs destroyed, the exposure ranking is not a record of AI being used at all but a model's rating of how exposed each occupation's listed tasks are, checked afterwards against what people were observed asking an assistant to do, and it is a working paper rather than a reviewed and settled finding. The authors say as much, writing that the facts they document may in part be influenced by factors other than generative AI. It is still one of the few direct measurements of employment that exist at all, which tells you how early this is.

The average is not anybody

A total that holds steady is compatible with an enormous amount of rearrangement underneath it. The count of tellers held through the cash machine, and the job stopped being the job. Somebody who was quick and exact with money, and who had taken the work partly because it did not involve talking strangers into purchases, still appeared in the statistics as employed. What they had lost is in no series at all.

That is the ordinary shape of the thing. The gains from automation are thin and very wide, spread across everyone who buys cloth or visits a bank, in amounts too small for any one person to notice. The costs are narrow and deep. They land on the particular people whose particular skill was the part that got taken, in the years before anything has arrived to replace it. Bessen makes the timing point without drama: building the training institutions and labour markets for a major new skill has sometimes taken decades. Nobody has decades of working life to spend waiting for one.

The work that has no code yet

Automation also makes work, and some of the work this round has made sits inside AI itself. Models are shaped by people who write example answers, choose between two candidate outputs, and rate what comes back. Where training data comes from sets out what one project disclosed about that workforce and then deliberately stops, because the present-day picture of the work is not in the published record. Treat that hole as real. The confident general claims about those jobs are unsourced in both the admiring and the damning direction.

There is a second reason new work is hard to see, and it belongs to the instruments rather than to the work. Occupational statistics count people into categories, and the categories are revised on a schedule. The American classification in use today is the 2018 edition; the federal review now under way asks, among other things, whether to add new detailed occupations, and it is aimed at 2028. Until it lands, a genuinely new kind of work is counted inside whichever old category sits nearest to it.

So the instruments are lopsided. A job that shrinks does so inside a category that already exists, and it shows up on schedule. A job that appears has to wait for somebody to write a code for it, and until then it is filed under something else or not seen at all. Every reading of the totals is tilted a little toward loss, for a reason that has nothing to do with what is happening to work.

Three questions that need not agree

A headline asks whether AI can do somebody's job. There is no answer, because a job is not the kind of thing that gets done. Take it apart into what the person actually spends the week on, and then ask three questions, expecting them to point in different directions.

Which of those tasks does the machine take? What does taking them do to the value of the ones it leaves? And does the work getting cheaper make anyone want more of it? The teller's answers were two tasks, a job that turned into selling, and yes for as long as banks wanted more branches. The power loom's were nearly all of them, coordination as the surviving skill, and yes on a scale nobody had planned for.

That leaves one habit worth keeping, because you will need it long after the current examples are stale. A projection is an argument about which tasks look automatable, and it can be built at a desk before anything has happened. A signal is a count of something that already occurred, in a named group of people, over a stated period, which would have come out differently had the claim been false. The 47 per cent was a projection and told you nothing about 2023. Five thousand support agents and a payroll file are signals, and they are small, recent and awkwardly shaped, which is what signals look like before they grow large enough to be obvious.

When someone tells you what AI is going to do to work, the useful question is not whether they sound excited or alarmed. It is which of those two things they are holding.