Will AI take your job? What the payroll data shows so far

Not the whole workforce, and not yet. But one group is already measurably worse off, and it is not the group most of the headlines are about.

Analysis by Inzonex · published 2026-09-25 · counted from published open data · how we work

Key figures

Across the economy there is no evidence of widespread displacement.

Workers aged 22 to 25 in the most AI-exposed occupations are about 19% below where they would be had they kept pace with less-exposed peers.

The gap is produced by hiring that did not happen, not by people being let go.

What was measured

The Stanford Digital Economy Lab used payroll records from ADP covering roughly 4.6 million United States workers across more than 730 occupations, from November 2022 to June 2026. Payroll records are the strongest evidence available on this question because they count people who were actually paid, rather than asking employers what they intend to do.

Occupations were sorted into five bands by how exposed they are to what current AI systems can do, and the bands were then compared with each other over time.

The one group that moved

Ages 22 to 25, most-exposed jobs11.0Ages 22 to 25, least-exposed jobs10.0
Employment change, November 2022 to June 2026 (first bar is a fall, second is a rise) (percent).

Employment of 22 to 25 year olds in the two most exposed bands fell about 11% over the period. In the three least exposed bands the same age group grew about 10%. Experienced workers show no comparable gap in either direction.

July 202515.0June 202619.0
The gap for 22 to 25 year olds in AI-exposed jobs, against their peers (percent).

The gap was 15% when it was first reported in July 2025 and 19% a year later. It is widening, steadily, rather than spiking.

Fewer hires, not more firings

This is the part that gets lost. The adjustment runs through reduced hiring rather than increased separations. Nobody is being marched out. Positions that would have been opened for someone at the start of their career are not being opened.

That distinction matters for anyone trying to act on this. It means the visible signal is not a redundancy announcement. It is a job advert that never appears, which is the hardest kind of change to notice from inside a company and the easiest to dismiss from outside it.

The declines also concentrate where AI is more likely to automate a task rather than augment the person doing it. Same technology, opposite effect on headcount, depending on which of the two a given job mostly involves.

Which jobs people are asking about

The four occupations people search for most alongside this question are accountants, lawyers, software engineers and doctors. The study does not publish a verdict for any single one of them, and neither will this page. What it does establish is the shape of the risk: not whether a profession disappears, but whether its entry-level rung is still being built.

A profession can look completely stable in headline employment while its junior intake quietly halves. Those two things stay compatible for years, and then stop being compatible all at once.

How to read this about your own job

Worth holding alongside this: in the only randomised trial of AI on real work, experienced developers took 19% longer with AI tools while believing they had been 20% faster. Hiring decisions are made on expectations, and the expectations have not been tracking the measurements.

Three honest limits. This is United States payroll data, so it does not automatically transfer to other labour markets. It covers to June 2026, so it cannot speak to anything that has happened since. And exposure bands are an estimate of what AI can do to a job description, not a measurement of what any particular employer decided.

Within those limits the finding is narrow and solid: across the economy, no widespread displacement; among people at the very start of a career in the jobs AI is best at, a gap of about a fifth, made of hiring that did not happen.

Sources and method