1 SEP 2026 — A study using the Real-Time Population Survey finds generative AI has reached 80 per cent of occupations, and that in most of them fewer than half of workers use it. The number nobody is quoting is the task one: 2.8 per cent of tasks exceed 50 per cent adoption, and none exceeds 70.
What was measured
The work is by Alexander Bick of the Federal Reserve Bank of St Louis, Adam Blandin and Tyler R. Schumacher of Vanderbilt, and David J. Deming of Harvard, published on 1 September and drawing on survey data as of May 2026. It measures what workers say they do, cross-referenced against the United States Department of Labor's O*NET occupational database, rather than inferring usage from vendor chat logs.
As of May 2026, 45 per cent of American adults aged 18 to 64 used generative AI for work, 55 per cent used it outside work, and 62 per cent used it at all.
Occupation adoption and task adoption are different measurements
Four out of five detailed occupations show adoption above 20 per cent. Only one in six exceeds 70 per cent, and those are mostly computer-oriented roles.
At the task level the picture changes completely. Of the tasks that make up those occupations, 2.8 per cent have adoption above 50 per cent and not one is above 70. Even in jobs where most people use the tools, they use them for a narrow slice of the work.
The gap between occupation-level and task-level adoption is the core finding, and the one most likely to be dropped in summaries. An occupation-level number answers whether people in this line of work touch the tool. A task-level number answers how much of the work it has actually taken over. Only the second bears on productivity.
Why this is the study the earlier numbers should have been
Last week we reported that a survey finding 80.8 per cent of engineers use AI agents daily had been administered to a panel screened for people already using AI agents. That figure measured intensity among users and was read as adoption across a profession.
This is the same subject approached properly. A population survey asks a representative sample of workers, including those who use nothing, so the denominator is the workforce rather than the customer base. Cross-referencing against O*NET means the occupational categories are standard ones rather than whatever the respondents call themselves.
The contrast runs in both directions. A vendor panel found 80.8 per cent daily agent use among engineers; this population survey finds 45 per cent for any work use across all adults. Neither is wrong about what it measured. Only one of them describes the labour market.
What broad but shallow implies for the productivity argument
The case for large economy-wide productivity gains rests on the tools absorbing a substantial share of the work, not on people having tried them.
A task-level ceiling below 70 per cent, with only 2.8 per cent of tasks past halfway, is consistent with bounded gains: time saved on drafting, summarising and code completion, and very little change to most of the work in most jobs. It is also consistent with an early stage of a longer diffusion, and this data cannot distinguish those two readings.
It establishes that anyone forecasting from adoption breadth is using the wrong number. Eighty per cent of occupations sounds like saturation and describes contact, not penetration.
The occupations at 70 per cent are the ones you would guess
Only 15 per cent of occupations show adoption above 70 per cent, and they are primarily computer-oriented.
That concentration matters because it is where almost all of the public discussion originates. Software engineers, data analysts and technical writers are both the heaviest users and the people most likely to be writing about it, which produces a systematic overestimate of how the rest of the economy is behaving.
The corrective is in the same dataset. In most occupations, fewer than half of workers use these tools at all, and that majority is not writing about their experience.
How this reads from Southeast Asia
The study is American, and the direction of the bias for this region is not obvious.
Two forces pull in opposite directions. Regional white-collar work is concentrated in high-adoption services and technology sectors, which would push a local figure up. Paid tiers, however, cost the same in Manila as in San Francisco while wages do not, pushing it down, and the less-capable, rate-limited free tiers that dominate here limit use further.
No equivalent population instrument exists for any ASEAN country, which is the actual gap. The vendor surveys that circulate regionally have the sampling problem described above. At task level, nobody knows the regional adoption rate. A government or statistics office wanting to answer it would need to add questions to an existing labour force survey, which is cheap and has not been done.
View author profile → · Editorial policy
