Why AI Use Surveys Differ: Workers vs Businesses

Why can the headline percentages all be accurate?

A worker reads that about four in ten people use generative AI for work, then sees that fewer than one in five businesses report AI use. The gap can look like a contradiction. It is not, unless both surveys asked the same question of the same kind of unit over the same period and used comparable weights.

A Federal Reserve Board research note comparing U.S. survey series reports that the Real-Time Population Survey (RPS) estimated work-related generative AI use at about 41% of the workforce in November 2025. The note describes RPS as a quarterly, nationally representative survey whose GenAI module typically gets 5,000–6,000 responses. Census’s Business Trends and Outlook Survey (BTOS) showed about 18% of firms reporting AI use at year-end 2025. The Fed note calculates its BTOS comparison using a four-period moving average to smooth biweekly volatility and account for rotating cohorts; that is the note’s calculation, not a claim that Census publishes the figure as a four-period average. Census broadened BTOS’s question in November from AI in producing goods or services to AI in any business function, so the year-end estimate is not directly comparable with a one-month estimate under the earlier wording. The Atlanta Fed’s Survey of Business Uncertainty (SBU), which surveys senior business leaders, estimated in its one-time November AI module that firms reporting adoption employed about 78% of workers. These U.S. estimates use different dates, denominators, samples, and measures; they are not a controlled comparison of one common quantity.

The 78% figure does not mean that 78% of employees personally use AI. It means the estimate weights firms by employment, so a large adopting company counts more than a small one. A business can report that someone uses AI somewhere in the organization even when many employees do not use it in their own tasks. Conversely, an employee might use a permitted public tool without a formal company-wide deployment. Those are plausible situations implied by the different units, not separate rates established by the surveys.

The Fed says differences in sampling distributions and units of analysis likely drive the largest variation: the firm-focused BTOS, executive-focused SBU, and household RPS represent firms and workers differently, and large firms employ many more people than their share of businesses. The note also says question framing, whether respondents count incidental use as material adoption, information gaps between respondents, and social-desirability effects may contribute. These are plausible sources, not a percentage-point breakdown of the gap. BTOS averaged about 20,000 responses across its four surveys before year-end 2025 despite sampling about 1.2 million businesses; the SBU AI questions received 1,032 responses. The headline estimates therefore remain survey estimates with differing designs and potential response error, not a census of every worker or firm.

Sources: Monitoring AI Adoption in the US Economy

What exactly counts as AI use in each survey?

Start with the technology label and the activity the question counts. RPS asks individuals about generative AI for work. BTOS has asked businesses about AI more broadly, giving examples such as machine learning, natural-language processing, virtual agents, and voice recognition. One person may count a text-generating tool used to draft a note; a business respondent may think the question means AI embedded in a production process. Both can answer accurately according to the question they heard.

Census documentation records a concrete wording change. Before November 17, 2025, BTOS asked whether a business had used AI in producing goods or services during the previous two weeks. The revised question asks whether it used AI in any business function. Census says expert feedback and cognitive testing found some respondents thought the old question did not apply to their business; others initially said no, although they used AI for hiring, project management, accounting, research and development, or software workflows. Census created a new time series after observing a level shift alongside the wording change. The revised series should therefore not be treated as a seamless continuation of the old one.

Pew Research Center’s American Trends Panel surveyed U.S. workers employed for pay in October 2024 about how much of their job was done with AI. Wave 157 received 5,395 responses from 6,490 sampled panelists. Pew reports that 16% said at least some work was done with AI and 63% said little or none; among non-AI users—workers who said they used AI little or not at all, or had not heard of workplace AI use—31% said at least some of their work could be done with AI. These are self-reported views from a survey sample, not a census or direct observation of tasks; sampling and nonresponse error remain possible. Perceived applicability is not current use, and these findings are not interchangeable with a yes-or-no adoption measure from another survey.

Frequency adds another distinction. “Used at least once,” “used in the last week,” “used daily,” and “used enough to change a process” describe different levels of activity. In the Federal Reserve note, RPS’s November 2025 estimate was 40.7% for any work-related GenAI use, 35.2% for use at least once in the previous week, and 12% for daily use in the previous week. These answers come from a nationally representative quarterly survey module, not workplace logs; self-report and question framing can affect what respondents count. A chart that shortens all three to “workers use AI” hides meaningful differences.

Sources: Monitoring AI Adoption in the US Economy; Workers’ exposure to AI; BTOS AI Core Question Updates

Why do the denominator and weighting change the answer?

A firm-weighted question asks how widespread adoption is across businesses. An employment-weighted question asks how much of the workforce is attached to firms that report adoption. A worker-level question asks what sampled individuals say they do. Those are different arithmetic questions before any sampling error is considered.

The Federal Reserve note compares reported firm-size distributions to explain why weighting can matter. It says the BTOS sample distribution mirrors the firm population, but its table shows that the shares are not identical: for firms with 1–49 employees, for example, the note reports 89.6% in the BTOS sample distribution versus 95% of firms in the population. The same note reports that firms with 1–49 employees account for 26.6% of employment, while firms with 250 or more employees are about 0.9% of firms and account for 56.2% of employment. Its analysis says large firms show the highest adoption in both firm-level series, and that sample distributions and adoption differences likely explain a considerable share of the BTOS–SBU gap. This is the Fed’s interpretation, not a complete numerical decomposition or evidence that the sample precisely reproduces every population share.

An illustrative comparison makes the denominator visible. Imagine a city with many small firms and a few very large employers. If a handful of large employers adopt an AI system, the share of firms adopting may remain modest, while the share of workers employed by an adopter could be much higher. This example is arithmetic, not a description of a measured city or a claim about how many workers actually use the system.

The Census survey’s own description also matters: BTOS is a biweekly, nationally representative view of U.S. businesses, and its current-use question refers to the past two weeks. The worker samples in the comparison use people as their unit, and the cited RPS figures refer to generative AI. A headline comparison should therefore name the unit, technology, field date, reference period, and weighting method. If those details are missing, treat the comparison as a rough contrast, not evidence that one survey disproves another.

Sources: Monitoring AI Adoption in the US Economy; Large Firms With at Least 20 Employees Biggest AI Users

What should a worker conclude about personal exposure?

A national use rate is context, not a personal job-risk estimate. It does not reveal whether your employer formally adopted a tool, whether your team uses it, how often your tasks are assisted, whether output is checked, or whether work has been reorganized. It also does not measure labor demand or displacement. Technical capability, exposure of a task, observed use, employer adoption, and a job outcome are separate signals.

Pew’s October 2024 findings, described above, distinguish present use from perceived applicability. That perception is not an assessment of technical feasibility, likely automation, or displacement. The survey describes worker responses, not every role or workplace.

When you meet a headline, use this short comparability test: (1) Is it generative AI or AI of any kind? (2) Is the respondent a worker, business, or executive? (3) Is the denominator people, firms, or employment at firms? (4) What period and frequency count? (5) Did question wording or the survey series change? If several answers differ, do not subtract the percentages or read the gap as hidden adoption or job loss.

For your own decision, list three to five recurring tasks and record whether AI is actually used, what part it supports, and what judgment, verification, relationships, or accountability remain with you. Watch for observable changes in task allocation and skill expectations. If you want a structured prompt for that inventory, the free task-level checker can organize change-pressure signals; it does not calculate a validated probability of displacement. Your next move should follow evidence from your actual work, not a mismatch between survey headlines.

Sources: Monitoring AI Adoption in the US Economy; Workers’ exposure to AI; BTOS AI Core Question Updates

Questions readers ask