MW Which workers are using AI surprisingly often? Dry cleaners and construction workers.
Andrew Keshner
The adoption of artificial intelligence at work has been widespread, but more 'shallow' than expected - for now
A new study looks at who's using AI at work, and how they're using it.
Artificial intelligence has been a sea change for the workplace, with its capacity to research and analyze reams of information at lightning-fast speeds.
Yet when it comes to the ways AI is filtering into Americans' jobs, many workers are looking at a trickle so far, rather than a flood.
Worker use of AI has been "widespread but shallow," with a broad range of people using the technology for just a small portion of their work, according to a new study released by the National Bureau of Economics on Monday.
What's more, researchers found the jobs where people use the technology are not always the ones that experts would expect.
For example, larger shares of dry cleaners and construction workers said they use AI on the job when compared to administrative assistants and data-entry clerks.
The study from researchers at Vanderbilt University, Harvard University and the Federal Reserve Bank of St. Louis comes at a time of growing anxiety that AI's spread is happening too fast.
A rising number of people are wary that AI will take their jobs, even though such displacement largely hasn't shown up yet in jobs data. Last week, Microsoft (MSFT) co-founder Bill Gates said too many people are underestimating the upheaval coming for the job market and elsewhere because of AI. The future will need "human-reserved" jobs, Gates argued.
By contrast, Meta Platforms (META) CEO Mark Zuckerberg said AI's advance will establish all sorts of new jobs, even if companies do become smaller.
The new research paper shows there's more time before that debate shakes out, said Adam Blandin, an economist at Vanderbilt University and one of the co-authors. "Even though a lot of people are using [AI], there's still a lot of room left to run," he said.
The study's findings show that 4 out of 5 occupations have adoption rates where at least 20% of workers use AI for job tasks. But there's just 1 in 6 jobs where at least 70% of workers are using it.
Across professions, Blandin noted, there are many tasks where at least 20% of workers are using AI, "but there are no tasks where almost everyone uses it."
For now, Blandin views AI's reach like a lake that's large enough to cover plenty of ground, but where people can still stand in most places.
Most under-predicted occupations
Occupation Actual % Predicted % % Gap
Construction-equipment operators 60.7 19.5 41.2
Computer and office machine repair 75.7 36 39.7
Industrial electrical and electronics repair 73.2 34.2 39
Special education teachers 70.9 38.6 32.3
Computer and info. research scientist 87.3 55.1 32.2
Laundry and dry-cleaning workers 49 20.6 28.5
Information security analysts 85.4 57.3 28.1
Chief executives 79.7 52.4 27.3
Network and computer systems adminis. 82.4 56.6 25.8
Shuttle drivers and chauffeurs 55.9 30.1 25.8
Source: NBER
Most overpredicted occupations
Occupation Actual % Predicted % % Gap
Receptionists and information clerk 7.6 54 -46.4
Medical secretaries and admin. assistants 16.8 61 -44.2
Tellers 18.1 51.8 -33.6
Customer service reps 30.1 61 -30.9
Data entry keyers 26.7 56.1 -29.4
Operations research analysts 33 61 -28
Ins. claims and policy processing clerks 34.4 61 -26.6
Logisticians 30.7 57 -26.2
Food preparation and serving supervisors 26.2 52.4 -26.2
Animal caretakers 5.3 30.6 -25.3
Source: NBER
AI-powered jobs are not always what you think
There has been plenty of polling on how AI is taking root at work. A Gallup poll last month showed that when workers said they used AI, around half said they used it for tasks like writing, editing and research.
Monday's research queried people about their jobs, how often they use AI and how they use it. The data came from an ongoing survey, collected between August 2025 and May 2026. It analyzed people's roles against academic gauges on a job's "exposure" to AI. Here, "exposure" counted as AI's capacity to double the productivity of a job task, either with the human or by itself. Jobs with high exposure tended to hinge on information analysis, while jobs with low scores tended to focus on manual work and direct interactions.
Sometimes, however, there was a wide gap between actual AI use on a job and the predicted use from exposure scores. Six in 10 construction-equipment workers and nearly half of laundry and dry-cleaning workers said they used AI for at least one task on the job. Three-quarters of computer and machine repairers said AI was part of the job.
What's going on? Blandin's read is that "there's some aspect of a task that requires some information retrieval - and AI is really good at info retrieval."
For example, he said, a construction-equipment operator might check AI about a safety rule or machine assembly before going ahead and doing the job.
It's possible many people are using AI for details beyond the immediate job duty. One hairdresser told researchers about using AI to put together promotion deals, Blandin noted. Dry cleaners could also be owners or managers who may be using AI for business decisions, he added.
Meanwhile, it's also possible people in some jobs use AI much less than it would seem on paper. For example, around 8% of receptionists and information clerks said they used AI for at least one of their tasks, the findings showed, while around 17% of medical secretaries and administrative assistants said they used AI for at least one job duty. Those jobs had much higher AI-exposure scores.
Regulations and legal liability concerns might be standing in the way of greater adoption, the study said. For example, medical offices may be slow to incorporate AI because of medical privacy laws.
"It could just be a miss on that exposure prediction," Blandin said. "It could be those people have a lot of information in their head that AI does not know." On the other hand, "another possible story is those jobs might be slow in allowing AI to do that work," he said - adding that there isn't "strong evidence which of those two stories is happening," though it could potentially be both.
-Andrew Keshner