Generative artificial intelligence is rapidly entering the workplace. Economists and business leaders want to know how this technology is altering daily routines and shifting labor demands. To find out, a team of researchers surveyed thousands of workers about their specific daily activities.
The findings, published as an NBER working paper, show that generative AI use is widespread across many occupations but shallow within them. Even in jobs where the technology is common, fewer than half of the workers actually use it to assist with their tasks.
Previous attempts to measure the labor market impact of generative AI often relied on exposure scores. These scores are predictions based on whether a software model seems capable of performing a specific job function. Other studies analyzed anonymous chat logs from platforms like OpenAI and Anthropic to see what users were typing.
Alexander Bick of the Federal Reserve Bank of St. Louis, along with colleagues from Vanderbilt University and Harvard University, identified a gap in these approaches. Predictions do not capture real-world use, and chat logs lack context about the user’s actual profession.
Mapping tasks to workers
The research team used the Real-Time Population Survey, a recurring national online survey of the US labor market. They collected data across four waves between August 2025 and May 2026. Each survey wave collected responses from roughly 5,000 individuals.
The researchers matched each employed respondent to a detailed occupation profile using a government database of job characteristics. The survey then presented respondents with the ten most important tasks associated with their specific job. Respondents indicated which of those tasks they regularly perform and whether they use generative AI to help complete them.
The analysis revealed that generative AI has reached a wide variety of workplaces. More than 80 percent of occupations and 40 percent of job tasks have at least some workers using the technology.
Adoption is highly concentrated in certain fields. Management, finance, and computer-related occupations report the highest usage. Tasks involving data analysis, report writing, and interpreting information see adoption rates near 60 percent.
Conversely, physical and personal service jobs see very little use. The lowest adoption rates appear in jobs requiring manual labor, equipment maintenance, or direct patient care.
Predictions versus reality
The researchers compared their survey data to existing exposure scores. They found that predictions only account for a fraction of actual usage patterns.
Clerical and administrative roles are heavily over-predicted by exposure scores. Tasks like processing insurance claims or preparing medical documents seem suited for automation, but actual adoption remains low. The authors suggest this gap may stem from strict compliance rules, data privacy barriers, or a lack of employer-provided training.
On the other hand, roles with high autonomy see more adoption than predicted. Chief executives, research scientists, and information security analysts use the tools at high rates. The researchers interpret this as evidence that workers with fewer bureaucratic barriers are more likely to experiment and find uses for the technology.
The learning process
The study highlighted massive variation among people doing the exact same job. For roughly 40 percent of tasks, at least a fifth of workers use generative AI, while more than half do not.
The researchers found that personal experience is a strong predictor of this behavior. Workers who use generative AI for one task are significantly more likely to use it for their other tasks. They are also more likely to use it at home.
The researchers argue this pattern points to a learning process. Once a worker pays the initial cost of figuring out how to prompt the software and navigate its limitations, they easily apply that knowledge to other areas of their work.
Context is missing in chat logs
The authors also compared their occupation-based data to measurements drawn directly from AI platform chat logs. The two methods produced vastly different pictures of how the technology is used.
Chat logs tend to classify a huge portion of user interactions as basic, generic actions. For example, chat data might classify 15 percent of all prompts as editing written materials. In the survey data, basic editing makes up a much smaller share of work.
The authors explain that chat classifiers analyze isolated text without knowing the user’s job. A scientist asking for help formatting data might have their prompt logged as basic editing. The survey method instead maps that activity to a higher-order goal, like preparing a research report.
Workplace barriers to adoption
This research indicates that predicting the economic impact of generative AI requires looking beyond the technical capabilities of the software. Adoption is heavily influenced by the worker’s individual willingness to experiment and the surrounding workplace environment.
For tasks hampered by compliance or privacy concerns, software improvements alone will not drive adoption. Usage in these areas will likely remain stalled until organizations adjust their internal policies or deploy secure, specialized versions of the tools.
