Researchers Publish U.S. AI Work Map Showing Broad but Shallow Use

In more than four in five U.S. occupations, at least one in five workers use generative AI on the job. A new working paper based on nearly 14,000 workers offers a map of reported workplace use that is broad across jobs yet shallow within occupations and tasks.

That distinction matters. Measures of AI exposure estimate which work a technology may be able to affect. The researchers instead measure the share of workers in an occupation who use AI for their jobs, and the share of people performing a task who say AI regularly helps them complete it.

The difference arrives as overall use continues to rise. Between August 2024 and May 2026, the share of U.S. adults reporting generative-AI use climbed from 45% to 62%, while workplace use among workers rose from 33% to 45%. The indexes ask: which parts of jobs receive AI assistance?

A survey built around the work people report

• The occupation index measures what share of workers in a given occupation use AI on the job.
• The task index measures what share of people who perform a particular task use AI to help complete it.
• Respondents first identified their occupation, then selected from the 10 tasks rated most important for that occupation in O*NET, a Labor Department-sponsored occupational database.

More than 80%Occupations with at least 20% workplace AI use

Reported workplace AI adoption reached at least one in five workers in more than 80% of occupations.

Fewer than 3%Tasks with more than 50% AI use

Fewer than 3% of detailed work tasks had AI adoption above 50%, and no task exceeded 70%.

Different workers use AI for different slices of a job

More than 40% of detailed tasks cleared 20% adoption, but fewer than 3% cleared 50%. In the occupations with the highest overall use, the researchers say the pattern is not one universally AI-assisted task. Rather, many workers use AI, but for different parts of their jobs.

Reported adoption was highest among computer and information research scientists, at 87.3%, followed by information security analysts at 85.4% and network and computer systems administrators at 82.4%. Reading technical documents, preparing research reports and analyzing data for trends were the most-assisted tasks. At the other end were animal caretakers, receptionists and information clerks, and licensed practical or vocational nurses.

Exposure is useful, but it does not settle adoption

Existing exposure measures do a reasonable job ranking which occupations and tasks have higher adoption, explaining roughly half of the variation in some comparisons. But actual use can diverge sharply from those predictions. Medical secretaries and administrative assistants reported 16.8% adoption, well below a 61% exposure prediction; the researchers point to sensitive records, privacy rules and the cost of errors as factors in the gap.

Occupation also says less about whether a particular worker will adopt AI. The paper finds that age, education and sex explain little of those individual differences. Workers with at least six months of AI experience use it for more work tasks, evidence the authors say is consistent with learning and experimentation carrying from one setting into another.

That is a useful constraint on sweeping claims about which jobs AI will transform. A job’s technical exposure can describe opportunity, while privacy, error risk and a worker’s familiarity may shape whether the tool becomes part of everyday work. Future survey waves will show whether today’s shallow pattern deepens as experience accumulates.