Worker and business surveys report different AI-use rates because they measure different things: individual use, firm adoption, or the share of workers employed at adopting firms. The Federal Reserve’s November 2025 comparison illustrates the distinction with roughly 41% of individuals reporting work-related GenAI use, about 18% of firms reporting AI use in December, and about 78% of the labor force employed at firms reporting adoption. These dated estimates differ in respondent, technology scope, timing, and weighting; none is a job-loss estimate. Census records also show that broadening its Business Trends and Outlook Survey question from production to any business function coincided with a level shift. Its cognitive testing identified uses the narrower wording could miss, but the available record does not quantify how much of the shift reflects previously unreported versus newly adopted use. Compare each rate with its definition attached, and use changes in your recurring tasks—not a national percentage—to guide a career decision.
Why can all three headline rates be true?
The rates can all be correct because they count different units and kinds of activity. In “The Fed - Monitoring AI Adoption in the US Economy,” the December 2025 Business Trends and Outlook Survey (BTOS) figure is about 18% of firms reporting AI adoption; the November 2025 Real-Time Population Survey (RPS) figure is about 41% of individual workers reporting work-related generative AI use; and the November 2025 Survey of Business Uncertainty (SBU) figure is about 78% of workers employed at firms reporting AI adoption. The last is an employment-weighted firm measure: it describes where workers are employed, not how many personally use AI. The figures also differ in date and technology scope. They are therefore not three estimates of one shared quantity, and none measures job displacement. Their spread illustrates why survey rates need their respondent, scope, and weighting attached; it does not quantify how much each difference contributes. A firm can count once after reporting adoption even if use is concentrated in one part of its operation; a person-based measure instead depends on the individual worker’s own reported activity. The employment-weighted rate assigns the adopting-firm status across that firm’s workforce in the calculation. That difference in what enters each numerator is enough to make the percentages diverge without an arithmetic error.
What do workers report that a business survey cannot see?
A worker survey answers from the person’s vantage point: did this respondent use AI for work, or believe it could perform some of their work? That can capture a personal interaction that has not become a formal, organization-wide rollout. It can also miss activity a respondent does not recognize as AI or does not recall within the question’s time frame. These answers describe reported experience and perception; by themselves, they do not show whether a tool can perform a task reliably, whether an employer has approved it, or whether the task or job has changed.
Pew Research Center’s “Workers’ exposure to AI” reports two distinct October 2024 responses from U.S. workers. Sixteen percent said at least some of their work was done with AI. Separately, among workers who did not report using AI, 31% said AI could do at least some of their work. The first is a report of use; the second is a view about possible applicability among non-users. The latter is not evidence that AI was already doing that work, that a capable system could do it to an acceptable standard, or that the employer intended to adopt one. Nor does the difference establish why a person was not using AI: the survey result alone cannot distinguish lack of access, preference, policy, opportunity, or other circumstances.
This split matters because actual use and perceived potential answer different questions even when the same people are asked about their work. A respondent may describe a narrow trial or occasional assistance as use, while another may judge that AI could help with a task without having tried a tool. Those are reasonable forms of self-report, but a percentage based on either response inherits the question’s wording and the respondent’s interpretation. Neither response supplies a task inventory, frequency measure beyond what the survey explicitly asks, or independent check of output quality.
The personal vantage point can reveal activity organizational questions may not capture: for example, a worker might voluntarily try a publicly available tool for a small work task, or use a feature without knowing whether it is part of a formal deployment. Conversely, a company can report adoption even when an individual respondent has not encountered the tool in their own role. These are plausible routes to different answers, not prevalence findings from Pew. Organizational adoption and personal use can coexist at different levels because adoption does not imply that every worker has access to, needs, or uses the system.
Self-reports are useful precisely because they ask about first-person experience that an employer may not observe. Their limit is the same first-person vantage: memory, interpretation of “AI,” and the boundary a respondent draws around “work” can vary. Pew’s 16% and 31% therefore illuminate reported use and perceived potential in October 2024, not the share of tasks technically automatable or the share of workers whose jobs were redesigned. To understand organizational reach, a firm-level question is a complementary measure; it answers from a different reporting position rather than correcting the worker’s account.
Sources: Workers’ exposure to AI
What does a business adoption question count?
A business adoption rate usually counts responding organizations that report at least one use within the question’s scope. The unit is the firm or establishment named by the survey design, not each employee, task, or hour of work. In “The Fed - Monitoring AI Adoption in the US Economy,” the December 2025 BTOS figure is a share of firms reporting AI adoption. Read that as organizational incidence: it does not mean that the same share of employees personally used AI, or that the typical adopting firm has changed a large part of its production process.
This distinction matters because a single in-scope use can determine a yes response. Imagine, as an illustration rather than a survey result, a company where a small finance team uses an AI feature to classify invoices while production, sales, and most other staff do not use AI. If the business respondent knows about and reports that activity, the firm can count as an adopter even though most workers have no direct interaction with it. The binary firm measure records presence within the organization; it does not encode breadth, frequency, task importance, or the fraction of staff reached. Those would require separate questions and denominators.
The organization’s answer also depends on what falls inside the survey’s definition of AI use. A broad AI label can include systems or embedded features that respondents do not think of as generative AI; a GenAI question narrows attention to systems that generate text, images, code, or other content. These are scope labels, not rankings of importance. A narrow GenAI measure may better match a question about generative tools, while a broader AI item may include other applications. Neither label alone tells a reader how consequential the use is for a particular worker or process.
The respondent’s knowledge and reporting position shape the firm answer too. A manager or other business representative may know about approved procurement and formal workflows but not see every employee’s informal experiment. Another respondent may report a feature incorporated into ordinary software without identifying it as AI. The result can therefore omit activity that workers recognize, or include organizational use that individual workers never encounter. The measure is a report from an organizational vantage point, not an audit of every tool or a census of employee behavior.
This makes the firm rate useful for a bounded question: how many surveyed organizations say they have at least one use under this instrument’s definition? It is not a direct measure of how much work has changed. Conversely, worker self-reports can register a person’s isolated use even where no organization-wide deployment exists. The Federal Reserve’s comparison is informative precisely because it identifies distinct respondent units and scopes; comparing its rates as though each estimated the same quantity would erase those design choices.
What did Census change, and what can the break establish?
The Census Bureau changed the BTOS core question’s operational boundary. “BTOS AI Core Question Updates” records that, beginning November 17, 2025, the earlier item asked whether a business used AI “in the production of goods or services”; the revised item asks whether it used AI “in any business function.” This is a change from a producer-centered question to an organization-wide scope. It can affect a yes/no response even if the business’s underlying activity did not change, because uses outside producing goods or services now qualify explicitly.
The timing matters for interpreting the series. Census says a level shift appeared in conjunction with the wording change and that it began a new time series with data released December 4, 2025. That establishes a comparability break around the revision: readers should not treat values on either side as a seamless measure under identical wording. Census reports the coincident shift and explains the decision to create a new series. Its documentation does not establish what fraction of the shift resulted from the broader wording, how much came from new adoption, or whether other factors contributed. The sequence is evidence of a measurement discontinuity, not a quantified causal decomposition.
Census’s separate “BTOS - Help” page gives a qualitative reason for broadening the item. In cognitive testing, some respondents who initially answered no then described AI uses outside production, including hiring, project management, accounting, and research and development. The page also notes software workflows with embedded AI features and respondents who felt a production-oriented question did not apply to their business. These examples show how a narrow frame could fail to elicit recognized organizational activity: respondents may interpret “use” through their central product or service rather than support functions, or may not recognize AI embedded in a familiar tool.
Those observations explain a plausible omission mechanism, but cognitive interviews do not estimate its prevalence across all BTOS respondents. They do not tell us how many businesses changed from no to yes solely because the words changed, nor how many actually began using AI during the same period. The examples are evidence about interpretation and question comprehension, not a numerical bridge from the old series to the new one. The careful conclusion is therefore that broader coverage plausibly surfaced previously unreported uses, while the size of that contribution remains unknown from the official documentation.
“Measuring AI Adoption among Firms: How You Ask Matters,” from the St. Louis Fed, offers a related illustration using European business surveys. It contrasts the U.S. production-focused BTOS wording with the EU-ICT-Firm survey’s broader any-purpose framing and examines how the latter can capture uses in areas such as marketing, sales, organization, and accounting. The comparison is consistent with the Census account: functions beyond production can matter to firm incidence. But it compares survey settings, so differences in wording are entangled with other differences in instruments and context; it cannot isolate a pure wording effect for U.S. businesses.
The St. Louis Fed analysis presents a U.S. figure of roughly 34% only as a conditional extrapolation from the European comparison. It is not an observed U.S. adoption rate, and it should not be substituted for BTOS results or described as what Census found. The extrapolation asks what a broader measure might imply under stated assumptions; it does not repair the series break with direct U.S. observations.
Taken together, the Census record and the St. Louis Fed illustration support a narrow interpretation. The question revision changed which business functions could qualify; cognitive testing identified examples that the old production wording could miss; and the reported BTOS level shift coincided with the revision. That combination makes a coverage effect plausible. It does not show that coverage explains the entire shift. Newly adopted uses may also have contributed, and the available official material does not assign shares to either mechanism. When comparing rates across the break, keep the wording regime attached to the number and leave the unexplained portion unresolved.
Sources: BTOS AI Core Question Updates; BTOS - Help; Measuring AI Adoption among Firms: How You Ask Matters
How do denominators and weights change what a percentage means?
A denominator is the total counted in the comparison; the weight determines how much each observation contributes to that total. The percentage is a ratio, so changing either its numerator or denominator changes what the result describes. In a firm-weighted rate, each responding business contributes one firm to the denominator, regardless of whether it employs 8 people or 8,000. In an employment-weighted adoption rate, firms contribute according to the number of workers they employ. The question changes from “what share of businesses reports adoption?” to “what share of workers is employed by a business reporting adoption?” Those are different quantities, even if they draw on the same broad business landscape.
Consider a deliberately small hypothetical economy with 100 firms: 90 small firms employ 10 people each, and 10 large firms employ 1,000 people each. Suppose 9 small firms and 5 large firms report AI adoption. The firm-weighted rate is 14 adopters out of 100 firms, or 14%. Yet the adopting firms employ 90 small-firm workers plus 5,000 large-firm workers: 5,090 out of 10,900 workers, or about 46.7%. This is an arithmetic illustration, not survey evidence. The employment-weighted result is much higher because adoption is concentrated among firms with many employees; the firm rate gives each organization equal weight.
Now suppose that within each adopting firm only one in ten employees personally uses an AI tool. In this illustration, that would mean about 509 personal users, or 4.7% of all workers, even though 46.7% work at an adopting business. The gap does not mean that one calculation is defective. The employer-based numerator assigns the adopting status to the firm’s full headcount; a worker-use numerator counts only individuals who say they used AI. It can therefore show organizational reach through large employers without showing how deeply tools have entered workers’ daily tasks.
The Federal Reserve’s “The Fed - Monitoring AI Adoption in the US Economy” distinguishes its business and worker estimates along these lines: its SBU figure is employment-weighted, while RPS asks individual workers about their own work-related GenAI use. Its earlier review, “Measuring AI Uptake in the Workplace,” also describes employment weighting as one reason a business survey can yield a higher worker-associated estimate than a firm-weighted rate. The mechanism is useful context for reading a percentage, but it does not make the estimates identical in date, scope, or survey design.
Employment weighting can answer a practical scale question: how much of the workforce sits inside organizations that report at least some adoption? It can signal that adoption has reached firms with substantial headcount, even where the overall share of firms remains modest. But a large employer may have deployed a tool only in a small department, while many of its other employees have no access or relevant use. If adoption clusters in functions unlike the reader’s own, the headline still says little about that person’s tasks. The weight captures firm size, not the distribution of use within the firm.
So read an employment-weighted percentage as a measure of where people work, not as the percentage personally exposed to a tool, the intensity of its use, or the fraction of jobs at risk. For personal exposure, the relevant evidence is whether a worker’s recurring tasks are changing, who uses the system, and what review or accountability remains. The weighting scheme clarifies the population represented by the number; it cannot substitute for task-level evidence or establish displacement.
Sources: The Fed - Monitoring AI Adoption in the US Economy; Measuring AI Uptake in the Workplace
What changes when a survey counts any use, recent use, or frequent use?
A reference window sets how far back a respondent is asked to remember, while a frequency threshold determines what counts within that window. A worker who tried a generative AI tool once at work, used it at least once in the previous week, and uses it every day could qualify under three different definitions. Each measure can be valid for its own question: whether use has ever occurred in a period, whether it is recent, or whether it is routine. The percentages will differ because the qualifying activity differs, even if the respondents and denominator stay constant.
In “The Fed - Monitoring AI Adoption in the US Economy,” the November 2025 Real-Time Population Survey estimates 40.7% of individuals used GenAI for work in some form, 35.2% used it at least once in the prior week, and 12% used it daily in that week. The 40.7% figure includes a broader set of reported work-related use than the narrower weekly and daily thresholds. The fall from any use to weekly use is consistent with some reported activity being less recent; the smaller daily figure identifies a more frequent subset. These are prevalence and frequency measures, not estimates of how much time AI saved, how important the affected task was, or how much the worker’s job changed.
A longer lookback can catch episodic activity that a short window misses: a monthly reporting task, an occasional client request, or a tool trial that happened earlier in the period. It can also retain experimentation that has since stopped. A short window gives a more current snapshot of use, but an unusually quiet week may omit seasonal or infrequent work. Neither automatically answers whether AI use has become embedded in a recurring workflow. Persistence requires either repeated measurement over time or a question specifically designed to establish continuing use.
The time-window effect also appears in business surveys. The Federal Reserve’s historical review, “Measuring AI Uptake in the Workplace,” compared instruments that asked about use over different periods, including BTOS’s longer six-month reference with surveys asking about a much shorter period. A firm that used AI once during a longer period may count as an adopter even if it has no current use; a short-period measure is more sensitive to recent activity. The review treats this alongside employment weighting as a reason survey estimates can diverge, not as a correction factor that can be applied mechanically to one headline rate.
For Census’s Business Trends and Outlook Survey, the two-week reference period is useful context for interpreting the business report as recent activity. The period alone does not say whether use is daily, experimental, or sustained across successive weeks. Likewise, a worker’s report of use within a week does not establish that the employer has standardized the practice. To compare persistence, the survey would need comparable questions asked across multiple periods or a direct measure of continued use; a single recall window cannot provide that history.
The right window depends on the decision the statistic is meant to support. Recent-use measures help describe what respondents report doing now; longer windows can be more informative where work is intermittent. When comparing a headline with another survey, keep the window and threshold attached to the rate rather than translating “any use” into “regular adoption.” When judging a personal work change, ask whether a tool appears repeatedly in the same task, whether it has displaced or altered steps, and whether the output still receives human review. Frequency tells you how often a reported behavior occurred under the survey definition; it does not, by itself, tell you its workplace consequence.
Sources: The Fed - Monitoring AI Adoption in the US Economy; Measuring AI Uptake in the Workplace
Can paired worker and employer surveys settle the disagreement?
Parallel employer and worker surveys can clarify why two rates differ, but they settle only questions their design and measures can answer. The OECD project page, “OECD AI surveys of employers and workers,” describes an initial program fielded in early 2022 across manufacturing and finance in Austria, Canada, France, Germany, Ireland, the United Kingdom, and the United States. Its employer survey covered 2,053 firms; its worker survey covered 5,334 workers. These are substantial samples across seven countries, but their stated scope is those sectors and settings at that time. The page also says the surveys were repeated in 2026 with revised questionnaires and broader country and sector coverage. That is a description of the repeat, not a result: the page does not report what the 2026 respondents said, so no trend or current prevalence conclusion follows from the announcement alone.
The value of collecting both perspectives is that they observe different parts of workplace change. An employer can report whether a firm has implemented AI, where it is being used, and what organizational reasons or barriers shape that decision. A worker can report firsthand whether a tool is available or used in their own work and how the experience affects their tasks. The OECD project describes these as complementary perspectives: implementation context from firms and experience from workers. A gap between them could therefore be informative rather than evidence that one side is mistaken. For example, organizational rollout may be limited to a particular function, while a worker’s personal use may be informal or outside a formal deployment. Conversely, a firm may have adopted a system that many employees in the sampled roles do not use.
To call the surveys paired in the stronger statistical sense would require evidence that particular worker responses can be linked to particular employer responses, or that the same units are followed under a defined panel design. The OECD page reports separate sample counts and a repeated survey program; it does not say that individual workers are matched to their employers, that the same respondents answered in both waves, or that a worker and firm answer the same operational question. Revised questionnaires and broader coverage may improve relevance and comparison, but they can also alter the populations or wording being compared. Without the published instruments and linkage design, the careful description is complementary employer and worker surveys, not matched observations proving whether each rollout reached its employees.
Even an ideally matched design would answer only the adoption and use part unless it also measured workplace outcomes. If a firm says AI is in use and its employees report using it, that agreement would show alignment on reported use within the measurement frame. It would not establish which recurring tasks were reassigned, whether output quality or productivity changed, how much review remained, whether staffing levels shifted, or whether anyone lost a job. Those require separate measures: task-level before-and-after allocation, quality and review data, staffing or hours records, and employment transitions over a suitable period. A survey can ask about some of these, but the conclusion must track what was actually asked and observed.
The strongest alternative interpretation is that agreement across linked firm and worker reports could be a meaningful early signal that implementation has reached people’s work, especially if repeated over time and broken down by role or function. That would improve on treating an employer’s adoption report as universal employee use. It still would not make use equivalent to displacement: tools may assist existing tasks, shift work toward review, or change demand without reducing headcount. The OECD survey design can help locate the difference between organizational rollout and first-person experience. To infer employment effects, however, the evidence must measure changed work and employment directly; agreement on adoption is a starting observation, not the outcome.
What should you check next?
Use survey rates to understand a population and a measurement choice, not to forecast your own job. When you read a headline, note its date, which AI it includes, who answered, whether the unit is a person or a business, how observations were weighted, the recall window, and the exact wording regime. Those details tell you what the rate can describe. For your own decision, list three to five recurring tasks and record whether AI is actually used, how often, who reviews the output, and whether expectations or handoffs have changed. Separate a trial from a repeated workflow, and a faster first draft from a task that no longer needs human judgment. Then choose a proportionate response: learn the tool if use is recurring and reviewable, strengthen verification or domain judgment where those remain central, or investigate an adjacent role if important tasks are being redesigned. Revisit the inventory when processes or standards change; it is a snapshot, not a validated displacement estimate. The free task checker at /ai-job-risk-checker can help organize task-level change-pressure signals, but it does not calculate a validated probability of job loss. If you need to compare paths against constraints, the paid career roadmap can structure scenarios and a 30/60/90-day plan; it does not guarantee employment or income.
Questions readers ask
Why can worker and business surveys report different AI-use rates?
A worker survey counts individuals reporting their own use; a business survey counts firms reporting organizational use. The surveys may also differ in AI scope, question wording, time window, denominator, and weighting.
Does the share of workers at AI-adopting firms mean those workers use AI?
No. An employment-weighted firm rate estimates the share of workers employed at businesses reporting adoption. It does not show how many employees personally use AI, how often they use it, or which tasks have changed.
Why did Census change its AI-use question?
Census broadened the Business Trends and Outlook Survey question from AI used in producing goods or services to AI used in any business function. Census cognitive testing identified examples of uses the earlier wording could miss. The level shift alongside the change marks a comparability break, but the available documentation does not quantify how much reflects newly captured existing use versus newly adopted use.
Can AI-use survey rates predict whether my job will be lost?
No. They describe reported use under particular survey definitions; they do not estimate an individual's displacement risk. For a personal decision, examine recurring tasks, actual use, review responsibilities, and observable workflow changes.
Sources and notes
- The Fed - Monitoring AI Adoption in the US Economy
The dated November–December 2025 U.S. estimates measure different things: firm-weighted BTOS broad-AI adoption, individual RPS work-related GenAI use, and SBU employment-weighted firm adoption. RPS frequency estimates further distinguish any use, last-week use, and daily use.
- Workers’ exposure to AI
Pew reports U.S. workers' October 2024 self-reported workplace AI use separately from non-users' perceptions that AI could do some of their work.
- BTOS AI Core Question Updates
Census documents the change effective November 17, 2025, from asking about AI in producing goods or services to AI in any business function, and says a level shift appeared in conjunction with the wording change, leading to a new time series from December 4, 2025.
- BTOS - Help
Census says cognitive testing found respondents who initially answered no but described uses outside production, including hiring, project management, accounting, R&D, or software workflows; some respondents also felt the old producer framing did not apply to them.
- Measuring AI Adoption among Firms: How You Ask Matters
The St. Louis Fed compares narrower production-focused and broader any-purpose business questions using European survey evidence and presents a conditional U.S. extrapolation; this illustrates how question scope can alter reported firm incidence.
- Measuring AI Uptake in the Workplace
The Federal Reserve review of 16 workplace AI surveys explains that firm versus worker units, broad AI versus GenAI scope, reference periods, employment weighting, and sampling design can produce different reported uptake; its BTOS comparison describes a longer six-month window and employment weighting as factors in a higher estimate.
- OECD AI surveys of employers and workers
OECD describes early-2022 employer and worker samples across seven countries and manufacturing and finance, explains their complementary first-person and implementation perspectives, and states that surveys were repeated in 2026 with revised questionnaires and wider country and sector coverage.
Apply this to your own work
See the whole job market at once.
Explore which occupations AI may reshape, then turn the signal into a practical response.
Explore the job map