In June, the unemployment rate in the United States stood at 4.2%, close to full employment. That includes the roles most exposed to artificial intelligence. It is not the number most executives expected after three years of hearing that AI would empty out entire offices.
The comfortable explanation would be that adoption is still small. The data presented by Peter McCrory, who leads economics at Anthropic, says the opposite. Around 20% of companies already use AI in at least one function. In the information sector, it is 40%. Quality-adjusted AI output grew more than 2,000% per year in 2024 and 2025. The scale is already there, and the expected effect on unemployment, so far, has not shown up.
This matters because most of the organizational structure decisions made in the last two years started from a simple line of reasoning: AI does tasks, tasks make up jobs, therefore automated tasks mean eliminated jobs. The reasoning is clean, easy to present to a board, and it keeps getting the prediction wrong. Understanding why it gets it wrong is the difference between reorganizing a company and dismantling it.
Jobs are not fixed bundles of tasks
When part of the work moves to the machine, the job does not disappear. It recomposes around what is left.
A credit analyst who used to spend half the week building spreadsheets, checking documents and writing standardized opinions does not become half an analyst when that half is automated. That person starts reviewing more cases, looking at the exceptions that used to sit in the queue, and discussing risk policy with whoever makes the final call. The person is still busy, but the content of the day changed.
The data supports that reading. No occupation in the O*NET database has all of its tasks systematically absorbed by Claude. Among the 81,000 users analyzed, the most cited gain was scope: people doing more things, with more proficiency. Reduction of work was not the dominant pattern.
For anyone deciding a budget, that distinction is decisive. Automating tasks creates additional capacity. Turning that capacity into cost reduction requires the company to actively decide not to use the scope it just gained.
The return on experience went up, not down
The most concrete example comes from a seven-month study on the use of Claude Code. The reasonable expectation was that programming with agents would level the field: if the agent writes the code, the advantage of someone with twenty years of experience shrinks.
The opposite happened. People with deeper domain knowledge delegate better to the agent, get it right more often on the first attempt, and recover better when the agent is wrong. The return on routine coding work fell, and the return on judgment rose.
That makes sense once you look at what the agent requires from the person. It needs a well-formulated problem, context about the surrounding system, and someone able to notice quickly that the output is plausible but wrong. That is not tool skill. It is experience built up in a specific domain.
The same mechanism shows up outside engineering. In legal, in procurement, in pricing, in complex customer service, the agent produces volume. Whoever can tell a correct answer from a convincing answer determines whether that volume helps the company or creates liability.
The pressure is showing up in entry-level hiring
If unemployment did not rise, that does not mean nothing is happening to work. It means the pressure is somewhere else.
It shows up as fewer entry-level openings, slower replacement when someone leaves, and higher expectations on each person who stays. Hiring rates for young workers in AI-exposed roles have weakened. McCrory is careful on this point: part of it is explained by the current American labor market, which hires little and fires little, regardless of AI.
Even so, the pattern deserves attention because it accumulates over time. A company that freezes junior hiring for three years does not feel the impact next quarter. It feels it five years from now, when it needs people with enough experience to judge well and finds out it trained no one. Today's junior is the professional who, in 2031, will know when the agent is wrong.
McCrory himself limits the reach of the argument. He is talking about the next year, not the next decade. As agents become more capable, and as AI starts contributing to innovation itself, this pattern may change. Treating the current stability as a permanent state would be the mirror image of the panic.
The question that changes the budget decision
The question circulating in boardrooms today is how much headcount AI allows a company to cut. It produces a fast number and a fragile decision. The more useful question is different: is AI being managed as a headcount reduction program, or as a way to expand what experienced people can take on?
The data so far favors the second reading. And there is a simple indicator to track internally. Instead of measuring only how many hours were saved, measure what now fits into the work week of people who previously had no time. If the answer is "nothing, there is just more free time", the company captured efficiency and changed nothing structural.
Anyone treating AI as a cost cut may find out, late, that they eliminated exactly the kind of experience the agents made more valuable. The ability to notice that an answer is wrong cannot be hired under pressure. It accumulates over years, inside the operation itself.