GenAI reaches 80% of jobs, but adoption stays shallow

Workers use generative AI across 80 percent of US occupations and more than 40 percent of tasks, but that reach says little about how deeply it has been adopted. In most occupations where workers use GenAI, fewer than half of the people doing the job report using it.

The findings come from a nationally representative survey that links workers' reported GenAI use to detailed occupations and tasks. The research, published by the National Bureau of Economic Research, is based on Real-Time Population Survey data and measures actual workplace use rather than estimating exposure from the capabilities of an AI system.

A job can contain tasks that a generative model could assist with without workers actually using one. The study’s authors say adoption is widespread but shallow, with substantial differences among people doing similar work.

“GenAI is used in many occupations and tasks, but few of these exhibit very high adoption rates.”

The study reports that four out of five detailed occupations have adoption rates above 20 percent. But only about one out of six occupations exceeds 70 percent adoption. The occupations with the highest use are concentrated in management and professional work, particularly finance, business and computer-related jobs. Adoption is lowest in personal-service work and jobs involving manual activity or interpersonal interaction.

The task-level results are even less concentrated than the occupation-level results. Only 2.8 percent of tasks have adoption above 50 percent, and no task exceeds 70 percent adoption. GenAI is often applied selectively within a job rather than becoming a standard tool for everyone who performs it.

Why chat-log estimates can look larger

The researchers also challenge a common way of measuring AI’s effect on work: classifying the tasks represented in chatbot conversations. The paper says those measures differ conceptually from worker surveys and tend to assign chats to broad, generic activities that occur across many occupations.

One example is editing written material or documents. OpenAI chat data classified 15 percent of chats as involving that kind of activity, while the US Department of Labor’s O*NET database indicates that only 2.4 percent of workers are in jobs that include the task. The mismatch does not mean either data set is necessarily invalid; it means chat volume is not a direct measure of the share of workers whose jobs include a particular activity.

The Department of Labor’s O*NET system provides the occupational framework used for this comparison. O*NET describes jobs through knowledge, skills, abilities, tasks, work activities and other descriptors. Its content model includes nearly 277 descriptors, and its database is updated through ongoing worker surveys and, in some cases, occupation-expert input, with new versions released on an annual schedule.

That structure gives the survey a more grounded denominator than a list of activities inferred from platform traffic. A chatbot may be used to edit a document by someone whose formal occupation does not center on editing, while O*NET asks how the work itself is defined across the US economy. The result is a narrower estimate of how many workers are actually using GenAI as part of their jobs.

Work use trails personal use

As of May 2026, 45 percent of US adults ages 18 to 64 reported using GenAI for work. Personal use was higher: 55 percent used GenAI for non-work reasons. Counting either category, overall adoption reached 62 percent.

The overlap between those groups is not specified in the supplied findings, so the 62 percent overall figure should not be treated as the sum of work and personal use. It indicates the share reporting use in at least one of the two settings.

The researchers identify prior experience as one factor associated with broader adoption. Workers who begin using GenAI in one domain tend to adopt it in other domains as well. That points to a diffusion problem as much as a task-automation problem: access, familiarity and individual willingness may determine use even when two workers perform substantially similar jobs.

The results also put a limit on claims based solely on occupational exposure. Exposure scores explain some, but “far from all,” of the variation in adoption across occupations and tasks, according to the paper. A capability map can show where AI might help; it cannot show how many workers are willing or able to incorporate it into daily work.

For companies measuring returns on enterprise AI, the gap is between a tool being relevant to an occupation and being used consistently by its workers. The data show broad relevance, but the high-adoption pockets are relatively small: about 15 percent of occupations exceed 70 percent adoption, and no individual task reaches that level. The question is less whether GenAI can touch a job than why most workers in the same job still do not use it.