Does AI replace workers? What the evidence shows about execution, strategy and entry-level jobs
No economy-wide job loss yet, but US employment of 22-25 year-olds in AI-exposed jobs is 19% below trend. Which work AI absorbs, and where judgment still pays.
Does AI replace workers, or only the execution part of their jobs?
By Andrej Ruckij · June 7, 2026 · updated November 4, 2026
TL;DR: Not across the economy, at least not yet. Stanford’s study of US payroll data finds no widespread displacement, but employment of 22-to-25-year-olds in the most AI-exposed occupations is about 19% below where it would otherwise be, and the gap comes mostly from fewer of them being hired. The experiments point the same way: AI takes over routine, checkable execution work and leaves the decisions about what that work is for. My read: if your job is mostly execution, the risk is real, and it shows up first as a hire that never happens rather than as a layoff.
Does AI actually replace workers?
The best employment-level evidence I know of is “Canaries in the Coal Mine?” by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab. They first published in August 2025 and revised the paper on August 12, 2026, using ADP payroll records covering millions of US workers through June 2026.
Their first finding is the one that gets left out of headlines: “We find no evidence of widespread, economy-wide job displacement.” The second is the one that gets put in. Workers aged 22 to 25 in highly AI-exposed occupations now have employment about 19% below their counterfactual, compared with people the same age in less-exposed occupations. That gap was 15% at the July 2025 data vintage, so it has widened over the year. Experienced workers in the same occupations show no comparable gap. In plain levels, the exposed group’s young-worker employment fell about 11% between November 2022 and June 2026, while the least-exposed group’s grew about 10%.
Three details matter more than the headline number:
- It’s hiring, not firing. In the authors’ words, the adjustment works “primarily through reduced hiring of young workers rather than increased separations.” People already in these jobs aren’t leaving at higher rates; fewer young people are getting hired into them.
- It depends on how AI is used. The declines sit in occupations where AI mostly substitutes for human tasks. Where it mostly complements workers, employment is flat or rising, especially for experienced people.
- Pay hasn’t moved yet. Adjustment is happening through employment rather than base pay.
The authors call these “early, descriptive indicators” and say plainly they are not causal estimates. The gaps shrink when you control for education, and some of the trends began before generative AI was widely used. A February 2026 follow-up from the same team looks at how interest rates and timing fit into the picture. So the honest answer to the question is: there’s a measurable effect on entry-level hiring in exposed jobs, and nothing yet that looks like mass replacement.
One source you can’t use to answer this either way is vendor case studies. They report hours saved and output per person. In the ones I’ve read for this wiki, none reports how many people were let go or not hired, so they can’t tell you whether jobs went away.
Which part of the job AI takes
The controlled experiments explain why the hit lands where it does. They keep finding that AI is strongest on work with a known shape and a checkable answer.
| Study | Who | What happened |
|---|---|---|
| Brynjolfsson, Li & Raymond, Generative AI at Work (NBER) | 5,179 customer-support agents | Issues resolved per hour up 14% on average, 34% for novice and low-skilled agents, “minimal impact” on experienced and highly skilled ones |
| Noy & Zhang (Science, 2023) | 444 college-educated professionals on writing tasks | Time down 40%, quality up 18%; the work shifted toward idea generation and editing and away from rough drafting |
| Dell’Acqua et al. (Harvard Business School and BCG) | 758 consultants | On tasks inside AI’s capability, 12.2% more tasks done and 25.1% faster; on a task outside it, AI users were 19 percentage points less likely to get the right answer |
My reading of these studies together: AI is absorbing the drafting, the first response, the variant, the routine resolution. What it doesn’t absorb is deciding which problem to solve and noticing when the output is wrong, and Dell’Acqua’s outside-the-frontier result shows what happens when people skip that second part.
There’s an apparent contradiction here. If AI helps novices most, why are novices the ones being hired less? My read, which none of these papers tests directly: if a tool gives a new hire much of what a year of experience used to give, an employer also needs fewer new hires to get the same output. The tool helps the junior people already in the job and reduces the number of openings for the next ones. I go through who gains more in does AI help beginners or experts more?
Where the market pays for judgment
Pay data inside one function shows the same split. Modash surveyed 499 influencer marketers for its 2026 salary report. According to the figures recorded in our wiki from the full report, owning campaign strategy came with about $14,830 more in pay, and execution-heavy tasks went with some of the lowest salaries. I couldn’t reopen the gated report to recheck that number for this update, so treat it as one opt-in, self-reported survey in one field.
Some hiring is starting to move the same way. Felix Rieseberg, an engineering manager at Anthropic, wrote on October 27, 2026 that he has stopped giving traditional coding interviews because AI now beats people at the small challenges they use. He looks at portfolios for small prototypes that show judgment and empathy for users instead, and he says he hasn’t found a full replacement for the interview yet. His line: “Execution isn’t expensive anymore.” That’s one manager’s practice, not company policy.
The warning that goes with it comes from a Microsoft Research and Carnegie Mellon survey at CHI 2025 (319 knowledge workers, 936 examples): people who trusted generative AI more reported less critical thinking. If judgment is the part of the job that keeps its value, leaning on the tool can wear down the very thing you’re being paid for. I cover what that means for interviews in how to hire in the age of AI.
Where I don’t think the pattern applies
This part is my judgment, and I don’t have data that tests it.
Brand-building doesn’t seem to split the same way. AI can make a single brand asset faster, but the mechanism is years of consistent exposure building memory in buyers’ heads, and there’s no high-volume task inside that to automate.
Enterprise sales looks similar for a different reason. The core of a large B2B deal is a specific person trusting another specific person over months. Prospecting research can be automated; carrying the relationship can’t, as far as I can see.
So the rule I use is narrower than “AI eats execution.” It eats execution that is high-volume and has a checkable answer. Execution that builds up slowly over years, or that runs through one relationship, holds up better.
What I’d do with this
If you run a team: list the weekly tasks in each role and mark the ones where there’s a known right answer someone could check. Those are where AI will substitute, and where the Stanford data says entry-level hiring is already thinner. Then ask where your next senior people will come from if you stop hiring the junior ones who used to learn on that work. Nobody in the research has an answer to that yet.
If you’re early in your career: the evidence says the door is narrower in exposed jobs, not closed. Rieseberg’s advice is the practical version. Show a small thing you built and the choices you made, because polish alone no longer tells anyone much.
If you own a function: move your own week toward the decisions the work feeds: what to test, what to measure, what to stop. Let the tool take the volume, and keep checking its output yourself, because that checking is the part of the job the market is still paying for.
Sources
- Brynjolfsson, Chandar & Chen: Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (Stanford Digital Economy Lab, revised August 2026) — ADP payroll data through June 2026; the 19% gap, hiring vs separations, substitution vs complement
- Stanford Digital Economy Lab: No widespread displacement, but the AI employment gap for young workers has widened to 19% (August 2026) — the 15% to 19% widening, the 11% and 10% level changes, the education caveat
- Stanford Digital Economy Lab: Canaries, interest rates, and timing (February 2026) — the authors’ follow-up on other drivers
- Brynjolfsson, Li & Raymond: Generative AI at Work (NBER Working Paper 31161) — 5,179 support agents, +14% average, +34% novices
- Noy & Zhang: Experimental evidence on the productivity effects of generative AI (Science, 2023) — 444 professionals; time and quality effects, task restructuring (working-paper version)
- The Harvard Crimson: the HBS and BCG “jagged frontier” study — Dell’Acqua et al., 758 consultants, inside and outside the frontier
- Modash: How much do influencer marketers make? (March 2026) — the 499-marketer salary survey; the strategy premium is from the full report as recorded in the Primores wiki
- Felix Rieseberg: Hunger beats tenure, advice for junior engineers (October 27, 2026) — dropped coding interviews, hires on prototypes
- Lee et al.: The Impact of Generative AI on Critical Thinking (CHI 2025, Microsoft Research and Carnegie Mellon) — 319 knowledge workers, 936 examples, self-reported
- Primores wiki: automation eats execution, AI skill leveling, strategy vs execution