Will AI replace accountants, lawyers and doctors?
The professions people search about most are not the ones the evidence says are moving first, and the movement is happening at one specific point in a career.
Key figures
Accountants are among the occupations identified as most AI-exposed, alongside customer service representatives, software developers and receptionists.
Early-career customer service employment fell about 11% from its November 2022 peak.
Health aide roles have been growing faster for young workers than for older ones over the same period.
Which occupations the data actually names
Stanford's payroll analysis is the most concrete evidence available, because it counts people who were paid rather than asking employers what they plan to do. The occupations it identifies as most exposed are customer service representatives, software developers, accountants and receptionists.
Note what is not on that list. Lawyers and doctors are the two professions people search about most, and neither appears among the occupations where employment has actually moved.
Automate or augment
The finding that explains the pattern is not about how clever a profession is. Declines concentrate in occupations where AI is more likely to automate a task than to augment the person doing it.
That is a distinction about the task, not the job title. Reconciling a ledger, answering a routine enquiry and writing boilerplate are tasks a system can complete end to end. Deciding what a client should do, examining a patient, or standing behind a judgement are not, and the same technology in those roles makes the person faster rather than unnecessary.
Over the same period, health aide work has been growing faster for young workers than for older ones. Physical presence turns out to be a strong moat.
Accountants, lawyers, doctors
| Profession | What the evidence supports | What it does not support |
|---|---|---|
| Accountants | Named among the most AI-exposed occupations in payroll data | Any measured disappearance of the profession |
| Lawyers | Nothing measured in this dataset either way | Confident claims in either direction |
| Doctors | Nothing measured in this dataset either way | Exam performance as evidence of clinical replacement |
For accountancy the honest statement is that exposure is high and the entry-level effect visible elsewhere is likely to apply. For law and medicine the honest statement is that nobody has the data yet, and anyone giving you a date is guessing.
The prediction that already ran its course
There is one profession where this argument has already had a decade to resolve, and the result is instructive.
In 2016 Geoffrey Hinton, later a Nobel laureate and widely called the godfather of AI, said that radiologists would be replaced by computers within five years.
Between 2014 and 2023 the number of radiologists in United States Medicare-affiliated practices rose from 30,723 to 36,024, an increase of about 17%, and the specialty is projected to keep growing. Hinton has since said he was wrong about the timing and that he had meant image analysis specifically rather than the job as a whole.
That is the cautionary case for every confident prediction on this page. The most qualified person in the field named a profession and a horizon, and employment then moved in the opposite direction, because reading the image was a smaller part of the job than it looked from outside.
Passing the exam is not doing the job
Most confident predictions about these three professions rest on a model scoring well on a professional exam. That is a weak basis, for a reason the agent benchmarks make concrete: systems that perform strongly on isolated questions perform far worse on multi-step real work, where the task is ambiguous, the context is unwritten and somebody has to be accountable for the outcome.
An exam is a closed question with a known answer and no consequences. A client is an open question with no known answer and consequences that land on a named individual. Professional work is mostly the second kind.
What to watch instead of headlines
Three signals are worth more than any forecast. Whether graduate intake is still being recruited at the same rate. Whether the work given to juniors is still the work that teaches judgement. And whether the profession's liability sits with a person or with a tool, because that has not moved at all, and it is what keeps a human in the chain.
Sources and method
- Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Brynjolfsson, Chandar and Chen, Stanford Digital Economy Lab. Source of the named exposed occupations, the customer service figure and the health aide comparison.
- AI won’t replace radiologists, but will dramatically change their jobs, Knowable Magazine, 2026, on Geoffrey Hinton’s 2016 prediction and what happened instead. Source of the 2016 prediction, the growth from 30,723 to 36,024 radiologists between 2014 and 2023, and the later qualification.
- Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, Becker, Rush, Barnes and Rein, METR, submitted 12 July 2025. The randomised trial showing that measured productivity effects can run opposite to expert expectation, which is why this page declines to forecast individual professions.
- This is United States payroll data. It covers occupations as recorded by employers and does not separate specialisms within a profession.
- No forecast for any individual profession is offered here, because the underlying study does not publish one.
- Every source here was opened and checked on 25 September 2026. Where a question has no measurement behind it, this page says so rather than dressing an opinion as data.