What AI Is Actually Used For at Work: Everywhere, and Barely Anywhere
Two sentences from the same study, and you need both of them.
Generative AI is now used somewhere in nearly every occupation in the United States. Within almost none of those occupations is it used for most of the work.
That pair explains a lot of disagreements people are currently having with each other, and it changes what your own adoption numbers are telling you.
What was measured, and why it is better than the usual figure
Alexander Bick, Adam Blandin, David Deming and Tyler Schumacher published this through the St. Louis Fed at the start of September, with the full working paper at the NBER. They ran a A survey whose respondents are selected so the result can be projected onto the whole population, rather than describing whoever happened to answer. It is the expensive way to do this and the reason matters here: most AI usage figures come from a vendor’s own users, who are by definition people who already signed up. A representative sample includes the people who never did. of roughly 14,000 US workers across four quarterly waves between August 2025 and May 2026.
The part that makes it unusually useful is what they linked it to. Occupations came from Bureau of Labor Statistics codes and tasks came from O*NET, the government database that breaks jobs into their component activities. So this is not “do you use AI.” It is “do you use AI for this specific task that the government already says your job contains.”
That produces the first task-level adoption index anyone has published.
For scale, adoption is still climbing: among US adults, use rose from 45% in August 2024 to 62% by May 2026, and use specifically for job tasks went from 33% to 45%.
Wide
The breadth is genuine and it is the half most people already believe.
At least 20% of workers use AI in more than 80% of occupations. It has reached almost everywhere. At the top end the numbers are high: computer and information research scientists at 87.3%, information security analysts at 85.4%, network and systems administrators at 82.4%.
At the bottom end are jobs where you would expect it, and the zeroes are instructive: driving trucks, presenting menus, assisting with medical procedures, collecting biological specimens. Physical work with people and objects in it.
And shallow
Now the half that gets left out.
Fewer than 3% of tasks show adoption above 50%. Not one task exceeds 70%.
Only 40% of occupations pass 50% adoption, and just 16% pass 70%. The researchers’ own summary is that “current adoption is widespread but shallow: genAI is used across many occupations and tasks, yet within most of them, fewer than half of workers adopt.”
Where it does go deep tells you what this technology is currently for. The three highest-adoption tasks are reading documents to gather technical information (61.3%), preparing research reports (60.7%) and analyzing data to identify trends (57.5%).
Notice what those three have in common. They are all “read a lot, then summarize or synthesize.” That is the job AI is genuinely doing at work in 2026, and it is narrower than the conversation suggests.
This corrects a number I published here
I need to deal with something, because I wrote a post on this and the new study undercuts part of its foundation.
How Much of Your Job Can AI Actually Do? About a Fifth was built on Google’s ATLAS research, which mapped 14.7 million conversations to occupations and tasks and found AI touching 68% of occupations but only 21% of tasks in the median job. Wide and shallow, from a different angle.
The Bick paper says that kind of measurement runs high. Their words: platform chat logs “tend to over-classify chats into generic activities spanning many occupations.” An An estimate of how much of a given job AI could plausibly do, usually built by matching what models can do against lists of job tasks. It is a capability estimate rather than a behavior measurement, and the two get quoted interchangeably. Vendors publish a lot of them, for the understandable reason that a large number is good for business. derived from a vendor’s own traffic will tend to spread one generic chat across many occupations it only loosely belongs to.
So what remains from the earlier post and what does not?
The shape remains. Both studies find AI present in most jobs and dominant in almost none. That was the argument, and it holds.
The specific figure should be treated as an upper bound. The 21% came from platform data of exactly the type this paper says over-classifies. I would not now write that number without this caveat attached, and I am flagging it here rather than quietly leaving it.
The direction of the error is worth remembering generally. When a capability estimate and a behavior measurement disagree, the capability estimate is almost always the larger one, and it is almost always the one in the press release.
The most interesting finding is where the records are sensitive
Buried in the occupation-level results is the thing I keep thinking about.
The researchers compared actual adoption against what exposure scores predict. Most jobs land somewhere near the prediction. Medical secretaries use AI at 16.8%, against a predicted 61%.
That is a forty-four point gap, and the explanation given is privacy rules and the cost of an error. The capability is there. The permission is not.
If you have read Shadow AI or the argument for a one-page AI policy, this is that problem showing up in national statistics. A rule that permits AI in general and forbids it on the records the job actually runs on does not produce cautious adoption. It produces a forty-four point hole, and nobody inside the organization experiences it as a policy decision. They experience it as the tool not being much use.
Some occupations run the other way and adopt at roughly double what exposure predicts, which the authors read as workers finding uses nobody anticipated. Both directions point the same way: what people are permitted and inclined to do explains more than what the technology can do.
What to take from it
- Your adoption rate is measuring breadth. If your number is “what percentage of the team has used AI,” you are measuring the easy half. The question that matters is which tasks, and how often.
- Expect the top three shapes. Read-and-summarize, draft-a-report, find-the-trend. If your team is not using it for at least one of those, the obstacle is local rather than technological.
- Treat vendor exposure numbers as a ceiling, not a forecast, and especially when they are quoted at you in a business case.
- Check whether your data rules are the constraint. If a task scores high on capability and near zero in practice, look at the policy before you look at the people. That is the medical secretary problem and it is more common than resistance.
The takeaway
AI at work in 2026 is a mile wide and a few inches deep, and the depth is the number nobody reports.
Stop asking how many people use it and start asking which tasks it has actually taken over. Fewer than three in a hundred tasks have it past half the workforce, the ones that do are all versions of reading and summarizing, and where a job looks capable on paper and empty in practice, the usual reason is a rule rather than a reluctance.
Sources: the figures are from Alexander Bick, Adam Blandin, David Deming and Tyler Schumacher, “What Work Does Generative AI Do?” (opens in a new tab) (Federal Reserve Bank of St. Louis, 1 September 2026), with the full paper as NBER Working Paper 35677 (opens in a new tab) (August 2026). The survey is the Real-Time Population Survey, roughly 14,000 US workers aged 18 to 64 across four quarterly waves from August 2025 to May 2026, with occupations coded to BLS categories and tasks drawn from ONET. I have read the St. Louis Fed write-up and the NBER abstract, not the full working paper, and it is a working paper rather than peer-reviewed work. The “widespread but shallow” and “over-classify chats” phrasings are the authors’. The earlier figure I am revising is from Google’s AI & Economy ATLAS v1.0 (opens in a new tab). The reading of the medical secretary gap as a policy problem rather than a capability problem is mine, and while the write-up names privacy rules and error cost as the explanation, no part of this study tested whether changing a data rule would change adoption.*
Related: How Much of Your Job Can AI Actually Do? About a Fifth is the post this one partially corrects, and still the better read on what a low task share feels like day to day. AI Productivity Gains: Everyone’s Using It, Almost No One Is Saving Time is the same wide-and-shallow gap measured in hours instead of tasks.