AI is coming for the paycheck before the job
Everyone has been watching for AI layoffs, and they haven't come. A new study finds AI showing up somewhere quieter, in the pay of the jobs it touches most, and the lowest-paid feel it most.
What the evidence says 5 min read
For three years the question about AI and work has been about layoffs. When does the wave come? Which jobs go first?
This month Goldman Sachs said the wave never arrived in 2026. The layoff rate stayed historically low. Its economists expect a “low-hire, low-fire” job market for another year.
So is AI doing nothing to work yet? A study from Apollo, the investment firm, suggests it’s doing something, just not where we’ve been looking. It isn’t showing up in who gets let go. It’s showing up in what people get paid.
What the study did
Apollo’s economists, Sania Edlich and Torsten Slok, took 321 US occupations and compared pay before and after ChatGPT arrived at the end of 2022.
Most studies of AI and jobs start from what AI could do in a job. This one starts from what people are actually doing with it. It uses Anthropic’s Economic Index, which measures how many of a job’s tasks turn up in real conversations with Anthropic’s AI, Claude. Jobs where at least half the tasks turn up count as highly exposed.
Only eleven jobs cleared that bar. Between them they employ about 5.8 million people, roughly one American worker in 27.
| Job | Share of tasks seen in use |
|---|---|
| Computer programmers | 75% |
| Customer service representatives | 70% |
| Data entry keyers | 67% |
| Medical records specialists | 67% |
| Market research analysts | 65% |
| Medical transcriptionists | 64% |
| Wholesale sales representatives | 63% |
| Database architects | 58% |
| Financial and investment analysts | 57% |
| Software testers | 52% |
| Statistical assistants | 51% |
What it found
After 2023, real pay (pay after inflation) in those eleven jobs fell behind pay in other jobs by about 6.7%. The number of people working in them didn’t change in any way the study could detect.
Then the authors split the workers by how much they earn, and the picture tilted.
The less you earn, the bigger the gap
How far real pay in the most AI-exposed jobs fell behind other jobs after 2023
The best-paid workers in exposed jobs showed no clear effect at all. The cost landed on the people with the least room to absorb it.
It isn’t a pay cut. It’s a smaller rise.
This is the part the headlines skipped, and it matters if you’re in one of these jobs.
Take customer service representatives. Between 2022 and 2024, their real pay actually went up 2.6%. Nobody’s payslip got smaller. What the study measures is a gap: how much less their pay rose than it would have if it had kept pace with jobs AI barely touches.
That makes the effect easy to miss. You wouldn’t notice a raise you never got. You’d notice that the rent went up and your pay didn’t quite keep up, and you’d blame inflation, or your employer, or the economy. You probably wouldn’t blame the chatbot your team started using last year.
Why would AI hold pay down?
The study’s authors think employers are keeping the gains from AI rather than passing them on in wages. There’s another way to read it, and it fits evidence we’ve written about before.
AI tools help beginners far more than experts. In customer service, a new starter with a good assistant can handle calls that once took months to learn. That’s good news for the new starter. But it shrinks the gap between a new worker and an experienced one, and that gap is a big part of how people in these jobs earn a raise.
Our own data points the same way. Of the Claude conversations linked to customer service work, about two in three are people getting help with a task rather than handing it over. The tool mostly makes the worker faster. It doesn’t replace them. Which is exactly what you’d expect to show up in pay rather than headcount.
What the study can’t tell you
It’s early, and the authors say so.
- One AI company’s view. The exposure measure comes only from Anthropic, the one AI company that publishes this kind of data. People use other tools too, so the real exposure is probably higher than the study can see. (Anthropic’s data appears on this site too. We always say so, and it never feeds into any score.)
- Only part of the workforce. The study could match only 321 of roughly 800 US occupations to the AI data.
- A strange few years. Pay after 2023 was still settling after the pandemic, and the study can’t fully separate that from AI.
- Other studies see it differently. Stanford’s team, tracking young workers, finds the opposite pattern: fewer jobs, not lower pay. Both could be true at once. Employers may be hiring fewer young people and paying the ones already there a little less than they would have.
So read the 6.7% as a first sighting, not a measurement you can bank on.
What it means if you work in one of these jobs
The scores on this site describe how a job’s tasks hold up, not what it pays. But this study adds something worth knowing if your work is on that list, or close to it.
- Check your pay against prices, not against last year. A small raise can still be a pay cut once you allow for rent and groceries.
- Get good at the cases the tool can’t handle. The angry customer, the mistake nobody spotted, the request that doesn’t fit the script. That’s where experience still counts, and where a raise is easiest to argue for. Our guide to auditing your working day can help you find them.
- Ask what happens to the time saved. If a new tool makes your team faster, it’s fair to ask whether that shows up in pay, in training, or only in a bigger queue.
Goldman’s economists called today’s wage growth “a subdued pace”. The layoff wave may still come. But if this study is right, AI has already arrived in a lot of working lives. It came in quietly, in the size of a raise.
What this article is built on
- The Impact of AI on the U.S. Labor Market: Early Evidence from Observed AdoptionWhite paper, July 2026
Sania Edlich and Torsten Slok, Apollo Global Management · Quoted with attribution
- Anthropic Economic Indexrelease 2026-06-26, period beginning 2026-05-01
Anthropic · CC-BY
- AI Economic Indicators: Canaries DashboardRelease of 21 October 2026, data to September 2026 (vintage 16 October 2026); with the paper Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (Brynjolfsson, Chandar and Chen, August 2026)
Stanford Digital Economy Lab and ADP Research · No data licence stated; downloads published openly for tracking, used with attribution
Every source on the site, with its licence, is on the methodology page · data 0.8-observed-use
The jobs in this article
-
Exposed: Gradient 3Written profile Customer service representative AI could take a fair chunk of all six tasks: logging requests, sending documents, taking payments and passing work along. With nothing safely out of reach, it scores second lowest of the twenty. 42 holds up 73 AI helps 75 AI does it 33.6 Hands-on -
Exposed: Gradient 3Written profile Software developer AI could take a fair chunk of six of the eight tasks, led by testing, working out requirements and documenting fixes. Only talking things through with engineers and customers is mostly out of its reach. 47 holds up 77 AI helps 63 AI does it 14.2 Hands-on -
Exposed: Gradient 2Written profile Data analyst The ILO files this under a catch-all group for database and network professionals, whose tasks are mostly about data security. AI could take a fair chunk of watching for threats and planning data protection; the rest is mostly out of its reach. 51 holds up 77 AI helps 58 AI does it 4.1 Hands-on -
Exposed: Gradient 4ISCO 4132 Data entry clerks AI could take a fair chunk of all five tasks: entering data, registering documents, checking and correcting entries and running calculating machines. It has the lowest score of all 427 jobs. 30 holds up 76 AI helps 81 AI does it 30.6 Hands-on -
Exposed: Gradient 4ISCO 4110 General office clerks AI could take a fair chunk of seven of the eight tasks, including typing up and proofreading information and preparing invoices. Sorting and sending the post is mostly out of its reach. 41 holds up 63.5 AI helps 54 AI does it 27.6 Hands-on -
Exposed: Gradient 3ISCO 4226 Receptionists (general) Booking appointments and handing out forms and brochures are where AI could take a fair chunk. Welcoming visitors and pointing them the right way are mostly out of its reach. 44 holds up 67 AI helps 66 AI does it 23.9 Hands-on
Read next
What the evidence says 11 min read
Graduates in AI-exposed fields are being hired less. So where did they go?
Three new studies seem to disagree about AI and new graduates. They don't. Fewer start in the jobs they trained for, but barely more are out of work. Here is where they went.
What the evidence says 14 min read
The new jobs are coming. Nobody has built the road to them yet.
AI is now the reason US employers give most for job cuts, yet cuts are falling. McKinsey says the real test is moving 11 million people into new work, and most of the roads aren't built.
What the evidence says 10 min read
AI is writing the doctor's notes. The time it saves is the least of it.
AI scribes now listen in on GP appointments and write up the notes. The biggest study found they save about two minutes an hour. The bigger change is where the doctor is looking.