Who Gains Most From AI at Work? What 20 Studies Found

Every article about AI and work argues about who benefits, and almost none of them cite the studies that measured it. This page is the table those articles are missing: twenty studies that measured how the productivity effect of AI differs across workers or across tasks, each with its design, sample, average effect, and the exact figures for who gained more or less.

The short answer is that gains are uneven and usually largest for the least experienced. Twelve of the twenty rows find larger gains for lower-skilled, less experienced or lower-performing workers. Three go the other way or show experts slowed. The direction flips by task, and no row links a productivity gain to that worker's pay.

The scope is deliberately narrow. Usage logs and self-report surveys are not causal estimates, so they sit in a separate table. Surveys on whether workers admit to using AI sit in a third. One widely cited study is excluded because it was withdrawn, and it is named so that nobody adds it back.

Rows are ordered roughly by how often they are cited. Peer-reviewed journal versions are marked; the rest are working papers or research-organisation reports. "Who gained" quotes the study's own split. Every figure was read at a primary source, which for several paywalled journals means the author's working-paper text or the journal abstract; the methodology section lists which.

Where a study splits workers by skill, experience or prior performance, the larger gain usually goes to the weaker group. The chart shows the three rows with the cleanest paired figures.

The mechanism most authors propose is diffusion of best practice. The support-agent paper describes the assistant as spreading what the best agents already did. The creativity study finds the tool "effectively equalizes" weaker and stronger writers. The P&G experiment finds employees outside the core job, working alone with AI, matching teams that included an expert.

Part of the senior gap is adoption rather than usefulness. The Ant Group study attributes the smaller effect on senior programmers to lower engagement, with acceptance rates the same across experience levels. The three-firm Copilot trials found short-tenure developers more likely to keep using the tool.

Speed gains are more uniform than quality gains. In the first legal study, time saved was roughly the same regardless of baseline speed while quality gains went to the weakest. The most skilled support agents got small speed gains and small quality declines. And task type matters as much as the person: consultants outside the frontier got worse, agentic tasks gained less than analytical ones, and maintenance coding rose more than new features.

• Not "AI makes experts worse". Most rows show smaller gains for experts, not losses. Losses appear in four settings: small quality declines for top support agents, the outside-the-frontier consulting task, the Kenyan trial, and METR's experienced developers.
• Not settled for agentic tools. Most rows used 2022 to 2024 chat or autocomplete tools. The one agentic-era RCT is flagged as biased by its own authors, and the one study that compared agentic with analytical tasks found agentic tool-use tasks gained least.
• Not evidence of higher pay for those who gain. No row links a worker's productivity gain to that worker's earnings. The Danish administrative study finds no pay effect at all.
• Not a universal levelling law. The effect narrows gaps in assisted output. The education-gap study found the gap partly re-emerged once AI was removed, and two rows run the other way.
• Not one comparable number. Outcomes range from resolutions per hour to lines of code, graded quality and revenue. Averages from −19% to +55% are not on a common scale, and we do not average across studies.
• Not representative of all work. Eight of the twenty rows are coding. Two have vendor-affiliated authors.

These sources are cited as often as the studies above and measure something different: what people say they saved, or what a vendor's logs show. They are here so they are not confused with causal estimates.

The gap between the two tables is the point of our post on why AI adoption statistics disagree. A self-reported 3x speed-up and a measured 19% slowdown can describe the same developers, as METR's two exercises show.

The census measures gains the researchers could see. A large share of workers say they keep their AI use from their employer, which means employer-visible gains may understate who is benefiting, or attribute them to the wrong person. The figures are not comparable with each other, because each survey asked a different question.

What an employer should do with a hire who is quietly twice as productive is the subject of our companion post on two hires and one time tracker. The broader statistics on agent productivity and return are in our agent productivity data points.