Can You Compare AI Exposure With Adoption in Your Occupation?

What is the fairest comparison for one task you actually do?

Yes, but use the free map as a starting point, not as a personal task score. Its methodology describes a map of 342 U.S. occupations, drawing on Bureau of Labor Statistics Occupational Outlook Handbook data, with rough estimates of how current digital AI could reshape each occupation. It explicitly says exposure is not replacement and that the scores are not validated probabilities or hiring assessments. The map can orient you toward an occupation-level question; it cannot tell you how much AI is used for one duty in your own job. The St. Louis Fed’s “What Work Does Generative AI Do?” offers a different kind of evidence: reported use among workers in an occupation, and reported use among workers who perform a particular task. Keep the two measures in separate columns. Exposure estimates describe potential for change; adoption measures describe reported use in a defined population and period. Neither says whether a task is fully automated, whether use is permitted at your workplace, or whether anyone’s job will disappear. Start closer to the work you actually do. Write down one recurring task in plain language, then note its input, output, frequency, and what must be checked or approved. Look for a survey measure whose task wording is genuinely similar. The Fed analysis asks workers who perform a named task whether they use AI for it, rather than treating every worker in an occupation as doing every task. Its index remains population evidence, not a measurement of your personal workflow. If no task-level measure fits, an occupation-level reported-use rate can add context, but label it as such rather than treating it as your task’s adoption rate. For example, imagine a reporting role that includes drafting a summary, reconciling figures against source records, and approving the final report. “Prepare research reports” might resemble the drafting task, but it does not necessarily cover reconciliation or approval. The same person may use a tool for an early draft while retaining responsibility for checking numbers and signing off. This is an illustration, not a reported case. Its point is that a job title or one task label compresses activities with different inputs, error costs, and accountability. An occupation average can be relevant while still missing the question that matters to you. A useful comparison ladder runs from your actual task, to a similarly worded task in a worker survey, to reported use across the occupation, to an occupation exposure estimate, then to firm surveys and platform logs. Moving outward changes the unit and often the population: a firm response describes an organization, while activity on one platform describes that service’s users and classifications. These sources can inform one another, but they do not form a single scale from “safe” to “at risk.” Begin with the closest task match, mark each step outward, and return to your local workflow to check access, verification, and human approval. The map is broad orientation; a task comparison is more relevant when its wording and population fit.

Sources: Methodology; What Work Does Generative AI Do?

What does the map's exposure score mean, and what does it leave out?

The live map’s exposure score is a rough estimate of how much current digital AI could reshape an occupation. It is a prompt to inspect work, not a measured rate of worker use. Capability asks whether a system can assist with or perform an activity; exposure compares that capability with work associated with an occupation. Neither says whether workers or employers use a tool now. The map’s Methodology says it covers 342 U.S. Occupational Outlook Handbook occupations, using rough large-language-model (LLM) estimates alongside Bureau of Labor Statistics data. It calls scores generated estimates, not validated probabilities or hiring assessments. The US Job Market Visualizer repository describes a tool for exploring BLS data visually, not a serious economic publication. The score is an estimate, not an observed outcome.

Observed worker use asks whether workers report using AI, and for which task. Organizational adoption asks whether an employer introduced or permitted a system. Redesign concerns changed duties, review, or responsibility. Labor demand concerns how many workers employers or customers need as costs and demand shift. Displacement is a realized reduction or loss of work. These are distinct questions, not outcomes predicted by the score. Capability may exist without a suitable workflow; exposure may be broad while use remains limited; use may support a worker rather than remove a task. A high estimate does not mean a task is reliably automated or an employer adopted automation. A low estimate cannot establish immunity: this occupation-level view of current digital AI is not an inventory of specialized duties or future capability.

The Methodology lists omissions: demand elasticity, regulation, employer adoption, new work created when costs fall, and social preference for human workers. These can affect whether technical possibility becomes workplace change. For example, a system might draft routine material while a person verifies facts, handles exceptions, and owns the decision. Whether this saves time, expands output, changes staffing, or is used depends on the setting. This is an illustration, not a reported employer plan. The score does not measure local conditions, verification cost, or error consequences. It cannot tell whether you have an approved tool or whether your role includes the task represented by its occupation label.

Do not confuse this map with the U.S. Bureau of Labor Statistics’ AI exposure categories. BLS combines five external sources into relative categories for detailed occupations: three theoretical measures of capability and two mapping observed AI interactions to tasks or work activities. BLS says these measures do not directly observe whether workers in an occupation used AI on the job. Exposure does not imply job loss, productivity gains, automation probability, or wage effects. The categories compare occupations within source measures; they are not outcome forecasts. The products differ in method and scale, so the map score is not an official BLS category. An official classification can still be exposure rather than an adoption count.

Use the map as orientation: it can identify an occupation and task bundle for closer inspection beside adoption evidence. The map labels an occupation; a worker survey may ask about a task; a firm survey describes a company. None describes your workflow automatically. Write down one recurring task, its input and output, frequency, checks, and who remains accountable. Seek the closest reported-use evidence, noting differences in geography, date, task definition, or population. A mismatch is a reason to ask about access, policy, task fit, or verification, not proof that a job is disappearing. The estimate alone cannot establish that your tasks will be automated or justify a career decision.

Sources: Methodology; US Job Market Visualizer; Artificial Intelligence (AI) exposure categories

How close is the U.S. task survey to your actual work?

The St. Louis Fed task index is closer to your work than its occupation index when you need to ask whether a particular activity is already being done with AI. It asks workers who perform a named task whether they use AI for that task. That is more relevant than asking only whether people with your occupation use AI somewhere in their jobs. It remains a population estimate, not a reading of your own job description, tools, permissions, or work quality. Ask “How closely does its task label match an activity I perform?” rather than “Does the index describe me?” The analysis, “What Work Does Generative AI Do?,” draws on nearly 14,000 respondents to the Real-Time Population Survey across four quarterly waves from August 2025 through May 2026. The survey covers U.S. adults ages 18 to 64. Researchers identify each respondent’s occupation, then show the ten activities O*NET rates as most important for that occupation. Respondents indicate which activities they perform; workers who use AI at work are then asked which tasks AI regularly helps them complete. O*NET contains roughly 2,000 detailed activities, so the ten-item selection is a practical sample, not a complete inventory of anyone’s duties. The indexes also have different denominators: occupation use is the share of workers in that occupation reporting AI use for their jobs; task use is the share of workers who perform a listed task and report using AI for it. A task-use rate does not mean that the same share of all workers, or of all hours spent on the task, is assisted. Consider an illustration: a U.S. operations analyst drafts a weekly report, reconciles inconsistent figures, checks sources, and explains exceptions to a manager. A survey task such as “prepare research reports” may overlap with drafting but miss reconciliation, source checks, or accountability for the explanation. Comparing the whole job with the drafting label is too broad; comparing only drafting still requires checking whether the survey wording describes the same input, output, and review standard. This is not evidence about analysts’ or employers’ AI use. It shows why a task label can guide a question without transcribing a local workflow. Gaps also arise when a task is new, specialized, regulated, or routinely bundled with review that its label leaves unstated. Start by describing the task plainly: its input, required result, frequency, and the consequence of error. Find the closest listed activity and read its wording before comparing rates. If it fits, the task index offers U.S. context about reported use among people who say they perform that activity. If it partly fits, name what is missing instead of treating the rate as yours. The occupation index adds context about use somewhere in the broader occupation, but does not resolve the task-level gap. Neither index establishes whether your employer adopted a tool, permits its use, finds it reliable in your workflow, or expects your role to change. Those require local facts about access, policy, review, and responsibility. Outside the United States, this evidence is a comparison method, not a local adoption estimate. The sample, occupation classifications, O*NET task list, workplace rules, and AI access are U.S.-specific. A similar task elsewhere may involve different workflows, regulation, language, or responsibilities. Look for a suitable local survey and check that it measures the same unit and period; if none exists, say the local figure is unavailable. Do not present a U.S. occupation average as a fact about your workplace. In any geography, use the task index where its wording aligns, state uncertainty where it does not, and treat a mismatch as a prompt to inspect your process, not a verdict on your role. Check the publication date before relying on it.

Sources: What Work Does Generative AI Do?

Why can occupation use look broad while task use stays shallow?

Because the measures count different things: many occupations can include some AI use even when few workers use it on any one listed task. In “What Work Does Generative AI Do?,” the St. Louis Fed reports that more than 80% of occupations had at least one in five workers using AI at work, while fewer than 3% of listed tasks had adoption above 50%. No listed task exceeded 70%. This is widespread presence across occupational groups alongside shallow use across most task categories. It does not show that most work in each occupation has shifted, or that change is unimportant wherever task adoption is low.

The difference starts with the denominator. An occupation adoption rate asks what share of workers in an occupation report using AI for their jobs. A task adoption rate asks what share of people who perform a particular task report using AI for that task. Workers in one occupation may divide their work among research, drafting, coordination, client communication, checking, and decisions. Some may use AI for drafting, others for research, and others not at all. At the occupation level, each user counts; at the task level, use is distributed among different entries. The occupation figure can therefore be broad even when no one task has high adoption.

The St. Louis Fed analysis reports that at least 20% of workers use AI in more than 80% of occupations, but at least half use it in only 40% of occupations. For tasks, more than 40% had adoption above 20%, while fewer than 3% crossed 50%. These thresholds describe different distributions: an occupation can clear a modest-use threshold even though no particular task is commonly assisted. They are results for the study’s U.S. population and observation period, not timeless rates for every workplace. The article’s finding that workers in high-adoption occupations use AI for different parts of their jobs supports a picture of distributed use, not uniform change to a role.

A task count still does not measure how much work changed. A task used by many people may take little time; a less common use may affect a consequential step. The share does not tell us task frequency, time saved, output quality, review burden, or who remains accountable. “AI helps complete it” also does not mean the task is automated: assistance may leave substantial human work in place. Conversely, low reported use does not prove a task is insulated from future change. The survey measures reported use in its window; it does not settle future employer deployment, work redesign, or staffing.

When comparing an occupation signal with adoption evidence, ask what each number counts rather than whether they match. An occupation rate establishes that use is present among workers in that group. A task rate shows whether people performing a similarly named activity report using AI for it. Neither identifies your task mix, frequency, access, checking needs, or accountability. Start with one recurring activity you perform and compare it with the survey’s task wording. If the wording does not fit, retain that mismatch as uncertainty instead of using the occupation average as a substitute. Broad occupational use warrants inspection of particular tasks; it does not establish broad task adoption or predict an individual job outcome.

Sources: What Work Does Generative AI Do?

Why do survey percentages disagree even before you compare occupations?

Survey percentages disagree because “use” can mean different events. One questionnaire may ask whether an employed person used generative AI at least once in the previous week; another asks whether a worker used it during the previous year; a third asks a manager whether anyone in the company uses it. A task survey asks among people performing a named task, how many say AI regularly helps them complete it? Each answer has a different population, recall period, and denominator. They describe activity under their own definitions, not a common adoption meter.

The Rapid Adoption of Generative AI, published in Management Science, reports nationally representative U.S. surveys. By late 2024, 27% of employed respondents had used generative AI for work at least once in the preceding week. Separately, it estimates 1% to 7% of all work hours were assisted. One figure counts people crossing a weekly threshold; the other measures hours. A person using a tool once and one using it repeatedly can both count in the first measure. Neither figure identifies which occupation’s tasks changed, and this period is not current prevalence.

Statistics Canada’s Workplace artificial intelligence use: A profile of sociodemographic and job characteristics uses the Canadian Survey on Working Conditions, collected from September 2024 to July 2025. It reports 22% of workers aged 15 to 69 used generative AI in their main job or business during the previous 12 months on average; the share rose from 17% in September 2024 to 30% in July 2025. This annual-recall measure does not count task performers or hours. The article says its analysis is primarily descriptive and does not capture intensity or frequency. Comparing its 22% with the U.S. weekly 27% as competing estimates erases geography and question wording.

The OECD report Generative AI and the SME Workforce surveyed more than 5,000 small and medium-sized enterprises in seven countries in late 2024, interviewing each firm’s most informed respondent, often an owner or manager. It found generative AI was used by a respondent or colleague in 31% of SMEs. That is a company-level indicator, not the share of workers or tasks using a tool. It cannot stand in for a particular employer or a worker’s own access. Country, industry, firm-size mix, and respondent knowledge can shift the result before an occupation comparison begins.

The San Francisco Fed’s Measuring AI Uptake in the Workplace reviews earlier employee- and firm-level surveys. Its March 2025 summary reports firm estimates from 5% to about 40%, and worker estimates from 20% to 40%; it identifies measurement considerations as part of the variation. These are earlier-study ranges, not a current universal rate or pooled estimate. A firm may report formal or known use while workers report their own activity, including use not captured by a company-level question. Check what each survey counted before choosing between numbers.

Before interpreting a gap, write down six labels: geography, observation date or window, unit counted, definition of AI, denominator, and task wording. “U.S. workers using generative AI at least once last week” and “Canadian workers who used it at work during the prior year” fail several matches, so keep their percentages as separate context. Occupation and task rates also differ: workers in one occupation may perform different duties. A firm rate answers whether reported use exists in an organization, not how common it is among workers in one occupation. Platform logs count activity visible on a service, not all workplace use or a representative workforce unless the method establishes that link.

Do not rank unlike percentages or treat a newer statistic as a better match merely because its date is recent. Prefer the closest measure: same country where possible, relevant time window, same unit, transparent AI definition, and task wording that fits work performed. If labels do not match, state the mismatch and use the figure as broad context. A gap can prompt a local question about access, policy, or task fit; it cannot by itself show faster occupational adoption, a worker’s task changing, or jobs being removed.

Sources: Measuring AI Uptake in the Workplace; The Rapid Adoption of Generative AI; Workplace artificial intelligence use: A profile of sociodemographic and job characteristics; Generative AI and the SME Workforce

What local conditions can explain a mismatch?

A mismatch between an occupation-level exposure estimate and reported AI use does not identify its cause. It points to a local question. The Federal Reserve’s 2025 household well-being report separates workers’ use of generative AI from reports that their employer encourages it. Among U.S. workers, 25 percent said they had used it in the prior month, while 19 percent agreed that their employer encouraged use. These answer different questions: personal use can occur without reported encouragement, and encouragement does not establish use for a particular duty. Neither measure establishes formal deployment across an organization. The report also offers a clue about autonomy. Thirty percent of employees who often or always controlled how to complete tasks reported using generative AI, compared with 20 percent of employees with less control. This survey comparison does not show that autonomy caused the difference. Workers with more control may also have different roles, tools, policies, or tasks. Treat the finding as a reason to consider discretion in your workplace, not as a diagnosis or forecast. Imagine your occupation appears highly exposed on a broad map, but you rarely use AI to prepare client summaries. Several explanations could fit: your organization may not provide an approved tool; client records may be restricted; your role may leave little time to learn a workflow; or a survey’s “drafting” task may differ from your duty to check sources, reconcile records, and approve the summary. These are questions to investigate, not evidence of a hidden adoption plan. Start with the exact work. Do you perform the task described by the survey, and how often? A label such as “prepare a report” can mean summarizing public information in one role and reconciling sensitive records in another. Is a tool available and permitted for those inputs? Restrictions on confidential, personal, regulated, or customer data can matter even when a system seems technically capable. Do not put work data into an unapproved service to find out. Then ask what must happen after a draft appears. Does someone need to check facts, trace sources, reconcile exceptions, explain a decision, or approve the result? How costly could an error be, and how much review time is required? Is the work screen-based, or does it depend on a customer conversation, physical inspection, or context missing from the prompt? A tool may help with one part while a person retains judgment and accountability. A broad exposure score and survey rate do not measure these local costs. The St. Louis Fed’s “What Work Does Generative AI Do?” analysis keeps the comparison grounded: its task measure counts workers who perform a listed task and report using AI for it; its occupation measure counts workers in an occupation who report job use. Neither resolves your access, policy, task frequency, or review requirements. If the occupation signal and matched task-use figure differ, note the mismatch and one plausible local explanation. Do not turn it into a score about employer intentions. The next useful question may be modest: “Is there an approved tool for this task, and what must I verify before using its output?” If policy is unclear, ask through the normal workplace channel before testing. Where a permitted low-stakes workflow exists, note its inputs, output, review time, and errors under existing rules. One trial cannot establish broad adoption or future staffing effects; it can show whether this task merits further review. If the work is rare, access is unavailable, or no close task comparison exists, monitoring may be enough for now. A mismatch should direct attention to a condition you can clarify, not become a career alarm.

Sources: 2025: Employment and Job Quality; What Work Does Generative AI Do?

What can a firm survey or platform log add to the comparison?

Firm surveys and platform logs add context, but answer different questions. Firm surveys report organization-level use and perceived staffing effects; platform logs record activity on one service. Neither tells you how often a particular worker uses AI for a named task. Keep the unit attached to every result. The OECD report *Generative AI and the SME Workforce* illustrates the company-level view. It draws on a representative 2024 survey of more than 5,000 small and medium-sized enterprises in Austria, Canada, Germany, Ireland, Japan, Korea, and the United Kingdom. Each respondent was best placed to describe company technology use, often an owner or manager. The survey found generative AI was used by a respondent or colleague in 31% of SMEs. Among SMEs using it, 83% reported no effect on overall staff need; 6% reported an increase and 9% a decrease. This does not establish use or staffing effects in a large corporation, particular occupation, or your team. “No effect on overall staff need” is also not a count of duties unchanged: it is a broad company assessment, not a task measure. A worker-task survey narrows the lens. In *What Work Does Generative AI Do?*, the St. Louis Fed reports distinct occupation and task adoption indexes based on workers’ reports. Its task measure covers listed tasks respondents perform; its occupation measure captures workplace use across an occupational group. A firm respondent may summarize colleagues’ tools, while a worker reports personal experience; the observations do not share a denominator. Company use does not show which duties in your role involve AI; low task use does not rule out broader deployment. Platform records offer a narrower view: what people do through one service. The San Francisco Fed page for the working paper *What Work Does Generative AI Do?* says platform chat-log estimates differ systematically from survey-based O*NET task shares, chiefly because chat data overclassify basic, generic tasks relative to the O*NET framework. Logs are not a census of work: they show activity visible to that platform and its classification method, not necessarily other tools or offline duties. Such records cannot stand in for an occupation’s adoption rate without evidence linking users to that workforce. Keep these sources in separate columns for population, unit, period, and reported outcome. Label the OECD findings “SMEs, organizational use and staffing assessment”; task-survey evidence “workers performing named tasks”; and platform logs “activity on one service.” Compare figures only when populations and questions align. Do not use the OECD staffing responses to promise a worker that no jobs are changing, or to predict cuts elsewhere. It records respondents’ views at that time, not later decisions or other employers. For your own comparison, use firm and platform evidence as context around the closest available task measure. A company-wide survey establishes reported organizational use, not your task’s frequency. Platform logs describe recorded interactions, not all work in your occupation. Keep those limits beside the checker’s occupation-level exposure signal. These are differently scoped clues, not one adoption scale: firm reports frame organizational conditions, task surveys locate reported use more closely, and platform logs describe activity within their service. None alone establishes whether your employer will redesign a role or whether your employment is at risk.

Sources: Generative AI and the SME Workforce; What Work Does Generative AI Do?; What Work Does Generative AI Do?

How should the checker result change your next step?

Start with one recurring task: note its input, output, frequency, and error consequences. Compare it with the closest reported-use measure, checking task wording, geography, and period. The St. Louis Fed’s “What Work Does Generative AI Do?” reports occupation use and use among workers performing a named task, but samples ten high-importance O*NET activities per occupation. The live map gives an occupation-level exposure estimate, not task adoption or your employment outlook. Treat the two signals as prompts. Higher exposure and matched use suggest checking output quality, review time, and accountability. Higher exposure with lower use calls for checking tool access, workplace rules, or a task-label mismatch. Lower exposure with higher use is a reason to examine reported assistance and workflow changes. Low readings on both axes offer no urgent conclusion. The Bureau of Labor Statistics’ “Artificial Intelligence (AI) exposure categories” says exposure does not imply job loss or automation probability. If policy permits, examine one low-stakes task with an approved tool and record review effort and errors. Otherwise, ask what tools and data rules apply. A closer task match, local access facts, or an actual workflow result could change your view. Treat uncertainty as a reason to monitor, not to make a career move from this comparison alone.

Sources: What Work Does Generative AI Do?; Artificial Intelligence (AI) exposure categories

Questions readers ask