Do Firm Surveys Undercount Workers’ Generative AI Use?

What is the short answer?

Yes, some firm surveys can miss AI use workers report. A question asking whether a business uses AI “in producing goods or services” may miss internal support work, or use the respondent does not recognize. The U.S. Census Bureau’s “BTOS AI Core Question Updates” documents a shift to “any of its business functions” after expert feedback. This demonstrates a measurement gap, not a correction factor for every firm survey. A gap between firm and worker percentages is not a count of hidden use. The Federal Reserve’s “Monitoring AI Adoption in the US Economy” compares three U.S. surveys with different respondents, units, scopes, and weighting; its late-2025 figures differ. A worker reporting generative AI use does not mean the employer integrated it into a workflow. A firm reporting adoption does not show how many employees use AI or how often. For a career decision, treat national rates as context, not a forecast about your job. Start with recurring tasks: which tools are available and permitted, what outputs they change, how much verification remains, and whether duties or expectations are shifting locally. That evidence can support a measured next step; a headline alone cannot establish your personal displacement risk.

Sources: BTOS AI Core Question Updates; Monitoring AI Adoption in the US Economy

How can a firm survey miss work that an employee can see?

A firm survey can miss work that an employee sees for two different reasons: the question may leave that work outside its definition, or the person answering may not know it is happening. The U.S. Census Bureau’s Business Trends and Outlook Survey (BTOS) provides a documented example of the first mechanism. Before November 17, 2025, its core question asked whether a business had used AI “in producing goods or services” during the prior two weeks. The revised version asks about use “in any of its business functions.” The wording change alters what counts; it is not a correction factor for all firm surveys. The Census Bureau says expert feedback and cognitive testing led to the broader wording. In testing, some respondents thought the original question did not apply because they did not see their business as a producer of goods or services. In other interviews, respondents initially answered no, but discussion showed that their businesses used AI in hiring, project management, accounting, research and development, or software workflows. These are examples documented by Census, not estimates of how commonly those activities occur. They show how a production-focused frame can overlook support functions that matter to an organization even when they do not make its product directly. For knowledge workers, possible internal uses include drafting, research, meeting summaries, and routine analysis. If a firm survey asks only about production, an employee may reasonably report work-related tool use that the business-level measure does not capture. The distinction is about coverage; it does not establish that the employee’s use is frequent, authorized, relied upon, or part of a redesigned workflow. Nor does it show that the survey respondent deliberately concealed use. A second issue is visibility. A business-level survey asks someone to characterize the organization, while a worker can answer from their own tasks. In its review of 16 workplace surveys, generally fielded from late 2023 to mid-2024, the Federal Reserve’s “Measuring AI Uptake in the Workplace” notes that worker use may not be known to management; it cites one survey in which 29 percent of respondents said management was unaware of their AI use. That result belongs to that survey and period. It is evidence that respondent knowledge can differ, not a universal rate of unobserved use. An individual representative may know formal pilots and policies yet lack a complete view of every employee’s day-to-day experimentation. Check question coverage and organizational visibility separately. A broader question may capture support-function use, while a worker can report activity an organization-level respondent cannot observe. Neither mechanism measures the total amount missed. If a firm rate looks low, inspect its wording and respondent before concluding that workers are not using generative tools or that an employer is hiding adoption.

Sources: BTOS AI Core Question Updates; BTOS Help; Measuring AI Uptake in the Workplace

Why do the percentages look like they disagree?

The percentages look inconsistent when readers treat them as answers to one question. They are not. A firm-weighted share asks what fraction of businesses report adoption; an individual survey asks what fraction of people say they use generative AI for work; an employment-weighted firm estimate asks what fraction of workers are employed by businesses that report adopting. These quantities differ. A worker may use a tool at an adopting firm, at a firm that does not report adoption, or not at all. The firm measure cannot identify each employee’s case. The Federal Reserve note “Monitoring AI Adoption in the US Economy” puts these contrasts side by side for late 2025. It reports about 18 percent of firms adopting AI in the Census Bureau’s Business Trends and Outlook Survey (BTOS), about 41 percent of individuals reporting work-related generative AI use in the Real-Time Population Survey (RPS), and employment-weighted Survey of Business Uncertainty (SBU) estimates of roughly 78 percent of workers at firms adopting AI and 54 percent at firms using large language models. These are historical snapshots, not current rates. Their distance is not a count of hidden workers: BTOS asks businesses about AI broadly, RPS asks individuals about generative AI, and SBU asks senior business respondents about firm use. Weighting explains one difference. Imagine many small businesses and a smaller number of large employers. If each business contributes according to its place in the firm population, the many small firms influence the firm-weighted result. If firms instead receive weight based on employee headcount, a large employer counts more because more workers work there. Thus “78 percent of workers at adopting firms” does not mean that 78 percent personally use AI; it describes employment at firms whose respondent says the firm adopts it. The Fed note says BTOS generally mirrors the distribution of firms, while SBU gives more representation to firms with greater employment weight. It identifies unit and sample distribution as major sources of variation, alongside question framing and other possible differences. Before comparing two rates, line up six fields: **unit** (firm, worker, or worker at an adopting firm); **respondent** (business representative, executive, or individual); **technology** (AI broadly or generative AI/large language models); **reference period** (dates or recall window); **frequency threshold** (any, recent, or regular use); and **weighting** (each firm, sampled person, or employment). “Measuring AI Uptake in the Workplace,” the Federal Reserve’s review of 16 mostly U.S. surveys fielded from late 2023 to mid-2024, found estimates ranging from about 5 to 40 percent and described differences in wording, respondents, samples, and definitions. That range reflects varied measurements in that period, not a current adoption interval. Its comparison reinforces the practical rule: do not subtract or rank percentages until their units and definitions match. A gap may reflect different questions, with no survey error required.

Sources: Monitoring AI Adoption in the US Economy; Measuring AI Uptake in the Workplace

What does a worker report reveal that a firm rate cannot?

Worker surveys can reveal generative AI use that a firm rate may miss: an employee may try a tool without a formal program or a firm respondent knowing. But a report of use does not show frequency, permission, reliance on the output, or changed responsibilities. It signals contact with workers’ tasks, not the career consequence.

Statistics Canada’s analysis of the Canadian Survey on Working Conditions covers employed people aged 15 to 69 in Canada’s provinces, with collection waves from September 2024 through July 2025. Respondents reported whether they had used generative AI in their main job or business during the previous 12 months. The share rose from 17% in September 2024 to 30% in July 2025, averaging 22% across the period. The authors explicitly say this measure captures whether use occurred during the year, not its frequency or intensity. A brief trial and daily workflow use can both qualify. The analysis compares this worker figure with a separate business estimate of 12% for the 12 months leading up to the second quarter of 2025. It suggests worker use may occur alongside or independently of formal organizational strategy. Since the surveys have different measures and periods, their difference is not a count of hidden use or proof that firms failed to report known programs. These Canadian findings describe Canada, not an individual workplace elsewhere. [Statistics Canada, “Workplace artificial intelligence use: A profile of sociodemographic and job characteristics.”]

A U.S. study shows how a shorter recall window changes the question. “The Rapid Adoption of Generative AI,” a peer-reviewed study using nationally representative surveys, reports that among employed U.S. respondents in late 2024, 27% had used generative AI for work at least once in the previous week. Ten percent used it every workday and 17% on some, but not all, workdays. The weekly threshold captures recent contact; categories separate daily from less regular use. They still do not establish permission, quality, reliance, or changed duties. These U.S. results cannot be merged directly with Statistics Canada’s Canadian past-year series. [“The Rapid Adoption of Generative AI,” Management Science.]

Self-reports depend on what respondents understand by “use,” remember, and count as work. They may capture experimentation as well as settled practice, without distinguishing the two. Worker surveys reveal individual task contact that firm answers may miss, but do not establish that management was unaware. An occasional AI outline for a memo and a team’s routine drafting, checking, and approval process are different evidence, even if an “any use” question counts both. This is an illustration, not a reported case. Before treating use as career pressure, verify task frequency, permitted tools, who checks results, whether outputs affect decisions, and whether expectations changed. Worker reports reveal contact; frequency, authorization, workflow reliance, and changed responsibility need separate evidence.

Sources: Workplace artificial intelligence use: A profile of sociodemographic and job characteristics; The Rapid Adoption of Generative AI

How much use is enough to count as workplace adoption?

There is no single threshold for workplace adoption because “use” means different things. A worker trying a tool once, a team using it repeatedly with approval, and a firm building it into a workflow are distinct observations. A survey counting any contact measures worker trials; a firm adoption survey measures whether the organization reports use under its definition. Neither alone tells you how much work has changed. It helps to keep the stages separate. Capability means a system can perform part of a task under some conditions. Exposure means a task could be affected by that capability. Individual use means a worker reports using it. Organizational adoption means a firm deploys or permits a tool or workflow. Redesign means responsibilities, task allocation, accountability, or staffing actually change. This ladder is a practical way to read evidence, not an official classification. Movement between rungs should be observed, not assumed. The surveys illustrate why frequency matters. In “The Rapid Adoption of Generative AI,” nationally representative U.S. surveys found that in late 2024, 27% of employed respondents had used generative AI for work at least once in the previous week. The study found 10% reported using it every workday, while 17% used it on some but not all workdays. These are frequency reports, not evidence of employer authorization, reliance, or job redesign. “At least once last week” and “every workday” therefore answer different questions even within one worker survey. A firm measure has a different vantage point. “Monitoring AI Adoption in the US Economy” compares a biweekly Census Bureau business survey, an individual worker survey, and an executive survey. It notes that the firm survey broadened its question in November 2025 from AI use in “producing goods or services” to use in “any of its business functions.” The Census Bureau’s “AI Question Wording Updates” documents that change. Broader wording may capture supporting functions; it does not establish use frequency or workflow change. Consider an illustrative task pair. A generative tool may help prepare a first draft of a meeting summary; that is assistance on one component. A client recommendation may still require checking source material, understanding context, resolving uncertainty, and taking responsibility for advice. A tool producing plausible text does not show that the accountable task has transferred. Occasional worker use does not show employer incorporation. For a career decision, look for recurring task changes and changed expectations. Note whether use recurs, is permitted, improves or weakens quality, adds verification time, and shifts who owns the final result. Persistent changes support an upgrade or adjacent-path experiment more than a capability demo or one-time trial.

Sources: The Rapid Adoption of Generative AI; Monitoring AI Adoption in the US Economy; BTOS AI Core Question Updates

Does more reported use mean jobs are already disappearing?

No. A survey showing that more firms or workers report AI use tells you the technology has reached work; it does not show that jobs have disappeared. “AI Adoption and Firms’ Job-Posting Behavior” compared online postings with AI adoption across U.S. industries and firms from September 2023 through November 2025. The authors report no evidence that higher measured adoption was associated with fewer postings. Statistically significant results pointed to a small positive relationship, which the authors caution is not causal: firms investing in AI may also be expanding for other reasons. This weighs against saying AI had already reduced aggregate postings in that period, but it does not show that every worker or task was unaffected. The measure matters. The study paired Lightcast online postings with the Census Bureau’s Business Trends and Outlook Survey and examined posting counts, not filled jobs, layoffs, wages, hours, or work quality. The authors say online postings contain limited wage information, so wages were not an outcome. Results end in November 2025 because Census revised its AI question wording, creating a break in the time series. Aggregate totals can hide fewer openings in one occupation alongside more in another, or changed requirements within jobs that remain open. Postings are one labor-market signal, not a direct count of workers’ experiences. The Fed note also reviews research with differing findings: some cited studies report little or no overall effect on jobs or postings, while other work identifies differences by occupation or experience level. These results need not conflict; they concern different populations, periods, measures, and outcomes. The authors do not claim to settle longer-term effects. Their backward-looking analysis is a snapshot and part of ongoing monitoring. It cannot establish that a particular occupation is protected or that an affected worker will find a transition easy. Keep the signals distinct. Capability asks whether a system can perform part of a task; exposure asks whether a task could be affected. Worker use records contact with a tool, while employer adoption depends on the survey’s firm-level definition. Labor demand concerns openings and hiring; displacement concerns workers losing paid work. A mismatch between worker and firm survey rates mainly raises questions about what each instrument captures. It does not establish job loss. For your own decision, look for repeated local evidence: duties removed or reassigned, role requirements changing in relevant postings, or a documented workflow redesign that changes who checks and owns the result. A single experiment or high national adoption figure is a reason to inspect tasks, not assume redundancy. Conversely, no aggregate decline in postings is not a personal safety guarantee. These U.S. data show no broad fall in online postings associated with adoption through the study period, while leaving room for concentrated effects. Match your response to what is changing in your role and workplace, and how strong that evidence is.

Sources: AI Adoption and Firms’ Job-Posting Behavior; Monitoring AI Adoption in the US Economy

What is the fair verdict on undercounting?

Some firm survey instruments can undercount AI use, but “undercount” needs a named instrument and an identifiable omission. The U.S. Census Bureau’s BTOS shows: from November 17, 2025, its question shifted from asking whether a business used AI “in producing goods or services” to use “in any of its business functions.” Census says cognitive testing found some respondents first answered no, then described AI in hiring, accounting, research, or software workflows. This shows how the earlier wording could miss support functions; it does not measure how many workers or firms surveys generally miss. Census reported a level shift alongside the wording change and created a new time series. Broader wording may improve coverage, but a comparison across the break mixes a change in measured activity with a change in reported use. The old question measured its narrower production category; the revised one deliberately measures more. For career decisions, ask what each rate counts rather than applying a guessed correction. Other gaps can reflect valid differences, not errors. “Measuring AI Uptake in the Workplace” reviewed 16 U.S. surveys fielded from late 2023 to mid-2024. Firm-level estimates ranged from about 5% to 40%; the note discusses respondent type, broad AI versus generative AI, recall period, examples, and firm-size weighting. Firm-weighted and employment-weighted estimates answer different questions: the latter gives larger employers more influence, but neither is automatically the share of workers personally using generative AI. “Monitoring AI Adoption in the US Economy” similarly compares firm, worker, and employment-weighted measures and cites units, sampling, framing, and information gaps as reasons for variation. Worker surveys add their own limits. Statistics Canada found generative AI use at work rose from 17% in September 2024 to 30% in July 2025 among workers aged 15 to 69; its prior-year question did not measure frequency or intensity. In “The Rapid Adoption of Generative AI,” a U.S. survey series found 27% of employed respondents used it for work at least once in the prior week by late 2024; 10% reported every-workday use, 17% some but not all workdays. Different recall windows and thresholds mean these are not competing measures of one hidden population; a trial is not routine integration. The verdict is conditional: wording and limited managerial visibility can leave worker activity outside firm responses; BTOS documents one route. But worker-firm percentage gaps cannot reveal how much use went unreported, how routine it was, or whether paid work was displaced. Say “undercount” only when you can name the survey, omitted category, and evidence. Treat rates as context; base a career move on recurring task changes and local evidence.

Sources: BTOS AI Core Question Updates; Measuring AI Uptake in the Workplace; Monitoring AI Adoption in the US Economy; Workplace artificial intelligence use: A profile of sociodemographic and job characteristics

Which career move fits the evidence you have?

Start with a current-role upgrade when a recurring task is changing and an allowed tool or adjacent skill can improve it. Test an adjacent move when your experience transfers but responsibilities are shifting. Plan a larger transition only when task evidence, destination requirements, and your constraints support the investment. Survey differences alone say little about your employer’s expectations or local options. For an upgrade, choose a recurring task such as preparing a first draft or synthesizing routine material, and check whether permitted AI assistance improves the work after review. A faster first draft is not a useful gain if verification erases the time saved or errors create unacceptable risk. Statistics Canada’s study of Canadian workers aged 15 to 69 found workplace generative AI use rose from 17% in September 2024 to 30% in July 2025. The study cautions that its past-year measure does not show frequency or intensity. It supports checking your actual task pattern, not assuming every worker in an adopting workplace uses a tool regularly. An adjacent move is worth testing when a changed workflow leaves work around the tool: clarifying a client’s situation, reviewing evidence, explaining tradeoffs, or taking responsibility for advice. These tasks require domain knowledge and judgment, but are not insulated from change. Try one bounded responsibility or project that uses your experience in another part of the workflow. Before investing in a course or certificate, compare its prerequisites and practice with current local postings or internal role descriptions. A supervised project or focused self-study with feedback can address a specific gap. A degree is a larger commitment, relevant when the destination requires its depth or credential. None of these formats alone establishes hiring value. Consider a larger transition when repeated changes remove core duties or role requirements shift, and a realistic destination fits after checking its entry requirements and local demand. The Federal Reserve’s 2026 analysis, “AI Adoption and Firms’ Job Posting Behavior,” found no negative aggregate posting effect in its studied period. Postings do not capture every employment outcome or rule out role-specific changes. The finding cannot decide an individual move. Compare salary floor, tuition, location, training time, health, family responsibilities, and access to approved practice. Prefer the smallest reversible experiment that produces useful evidence while preserving income and options. If documented redesign or removed responsibilities make delay costly, begin planning sooner; otherwise, test an upgrade or adjacent assignment before extensive retraining. The recommendation depends on your task evidence and constraints, not a national adoption gap.

Sources: AI Adoption and Firms’ Job-Posting Behavior; Workplace artificial intelligence use: A profile of sociodemographic and job characteristics

What should you track before spending or switching?

A survey gap is a reason to look closer, not to buy a course or leave a role. For two weeks, log five to ten recurring tasks that make up your work. For each, note frequency, whether an approved tool is available, whether you used it, what changed in the result, time spent checking or correcting, and who remains accountable. Add concrete changes in workload, handoffs, or expectations. Keep observed use, written policy, and manager awareness separate: a personal trial does not establish formal adoption, and a manager’s report cannot show every detail of your day. Do not put confidential material into a tool unless your organization permits it. This is a practical observation method, not a validated survey or forecast. The Federal Reserve’s “Monitoring AI Adoption in the US Economy” explains that firm and individual measures answer different questions and respondents differ in what they know or count. The U.S. Census Bureau’s “BTOS AI Core Question Updates” documents that some respondents did not recognize a production-focused question as applying to their business; discussion revealed uses in hiring, project management, accounting, research and development, and software workflows. This supports checking task details and question wording. It does not establish the amount of hidden use in your workplace. Choose one repeated, permitted task for a bounded comparison. Before trying assistance, list what a usable result requires: facts, format, error checks, and approval steps. Compare the ordinary and assisted processes on equivalent work, recording preparation, review, correction, and escalation time as well as drafting time. A quicker output that needs extra checking or leaves responsibility unclear is not a simple time saving. If quality holds and the process is allowed, you have evidence for a small role upgrade, such as taking on more review or improving a handoff. One good trial does not prove that a team has redesigned work. Look outside the log for concrete signals: a duty actually removed or reassigned, a new requirement in relevant job postings, or a documented team plan changing who does the work. Separate observations from predictions and ask whether the signal persists. Without repeated task change, preserve options and learn within the work you know. If changes recur, test an adjacent responsibility using your existing domain knowledge before committing to a longer course or larger pivot. Compare training with its prerequisites, cost, time, location, salary needs, health, and family obligations. Escalate faster if responsibilities or role requirements have already changed. The log cannot tell you whether a job is safe; it can help show whether the next step is an inexpensive experiment or a deliberate plan.

Sources: BTOS AI Core Question Updates; Monitoring AI Adoption in the US Economy

What is the next action?

Do not make a costly career move from the gap between firm and worker survey rates alone. Identify which recurring tasks are changing, whether tool use is permitted, and how much checking results require. A national figure cannot tell you whether your role is being redesigned. Choose a reversible next step: improve one task in your current role with an approved tool, or ask to try an adjacent responsibility. Compare the result for quality, verification time, and accountability. If your employer has documented a workflow redesign, removed responsibilities, or changed role requirements, you have a stronger reason to plan sooner than a short task log would provide. The Federal Reserve analysis, “AI Adoption and Firms’ Job Posting Behavior,” found no negative aggregate posting effect; job postings still cannot rule out changes in particular tasks or roles. Use the free task checker for transparent change-pressure signals, not a job-loss probability. To compare a stay-and-redesign path with adjacent or larger changes, the paid career roadmap compares those scenarios against your salary floor, location, training time, and family or health constraints. Let the survey gap change what you monitor, not dictate your occupation: match the scale of your investment to verified changes where you work.

Sources: AI Adoption and Firms’ Job-Posting Behavior