In June, the US unemployment rate stood at 4.2%, close to full employment. That includes the occupations most exposed to artificial intelligence. I read this as a significant departure from what many executives expected after three years of hearing that AI would empty entire offices.
The most comfortable explanation would be that adoption is still limited. The data presented by Peter McCrory, Anthropic's head of economics, says otherwise. About 20% of companies already use AI in at least one function. In the information sector, the figure is 40%. Quality-adjusted AI output grew by more than 2,000% per year in 2024 and 2025. AI already operates at meaningful scale, yet the expected effect on unemployment has not appeared so far.
This matters because many organizational decisions made over the past two years followed a simple line of reasoning: AI performs tasks, tasks make up jobs, so automated tasks mean eliminated jobs. The logic is clean and easy to present in a boardroom, but it keeps producing the wrong forecast. Understanding why it fails can determine whether a company reorganizes its work or dismantles capabilities it still needs.
Jobs are not fixed bundles of tasks
When part of the work moves to a machine, the job often reshapes itself around what remains.
A credit analyst who used to spend half the week building spreadsheets, checking documents, and writing standardized assessments does not become half an analyst when that work is automated. The analyst reviews more cases, examines exceptions that once remained in the queue, and discusses risk policy with the people making the final decision. The person remains busy, but the content of the day changes.
The data supports this 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 frequently cited gain was broader scope, with people doing more things and doing them more proficiently. A reduction in work was not the dominant pattern.
For anyone making budget decisions, this distinction is decisive. Automating tasks creates additional capacity. Turning that capacity into cost reduction requires the company to decide actively that it will not use the broader scope it has just gained.
The return on experience rose
The clearest example comes from a seven-month study of Claude Code usage. A reasonable expectation was that programming with agents would level the field. If the agent writes the code, the advantage held by someone with twenty years of experience should shrink.
The opposite happened. People with deeper domain knowledge delegate more effectively to the agent, get the first attempt right more often, and recover better when the agent makes a mistake. I read this as a shift in where experience creates value: routine coding matters less, while judgment matters more.
This makes sense when you consider what the agent requires from the person. It needs a well-framed problem, context about the surrounding system, and someone who can quickly recognize an output that is plausible but wrong. These abilities come from experience accumulated in a specific domain, not simply from knowing how to use the tool.
The same mechanism appears outside engineering. In legal work, procurement, pricing, and complex customer service, the agent produces volume. The person who can distinguish a correct answer from a convincing one determines whether that volume helps the company or creates liability.
Pressure is showing up in entry-level hiring
The absence of higher unemployment does not mean work is unchanged. The pressure is appearing elsewhere.
It shows up as fewer entry-level openings, slower replacement when someone leaves, and higher expectations for each person who remains. Hiring rates for young workers in AI-exposed occupations have weakened. McCrory is careful on this point: part of the decline can be explained by the current US labor market, which has low hiring and low firing regardless of AI.
Even so, the pattern deserves attention because its effects accumulate over time. A company that freezes junior hiring for three years will not feel the impact next quarter. It will feel it five years from now, when it needs people with enough experience to exercise sound judgment and discovers that it trained no one. Today's junior employee is the professional who, in 2031, will know when the agent is wrong.
McCrory also defines the limits of the argument. He is talking about the next year, not the next decade. As agents become more capable and AI begins to contribute to its own innovation, this pattern may change. Treating today's stability as permanent would be as misguided as assuming immediate mass unemployment.
The question that changes the budget decision
Many boardrooms are asking how much headcount AI allows the company to cut. That question produces a quick number and a fragile decision. A more useful question is whether the company is using AI mainly to reduce its workforce or to expand what experienced people can take on.
The evidence so far supports the second reading. There is also a simple indicator companies can track internally. Instead of measuring only how many hours were saved, measure what now fits into the workweek of someone who previously lacked the time. If the answer is "nothing, there is just time left over," the company captured efficiency without changing anything fundamental.
I would be cautious about treating AI primarily as a cost-cutting program. A company may discover too late that it eliminated precisely the kind of experience agents made more valuable. The ability to recognize that an answer is wrong cannot be hired on short notice. It accumulates over years inside the work itself.