What AI Exposure Misses About Adoption Within Occupations

What does an exposure score measure—and leave out?

An exposure score estimates overlap between an occupation’s described tasks and a defined set of AI capabilities. It maps potential contact, not each worker’s actual tasks. A high score identifies tasks worth examining, but cannot establish that a particular worker performs them, has approved access, gets reliable results, or has changed responsibilities. Exposure helps decide where to look; it is not a personal adoption measure or a forecast of job loss. A score matching occupation activities to capabilities summarizes the group, but people with the same title may have different assignments, task frequency, information access, or sign-off duties. A relevant task may occupy little of one worker’s week and much of another’s. Moving from an occupation-level signal to an individual conclusion assumes these differences do not matter. The St. Louis Fed’s “What Work Does Generative AI Do?” makes the denominator issue visible by measuring reported use at two levels. Its nationally representative U.S. survey drew on nearly 14,000 workers across four waves from August 2025 through May 2026. The occupation measure is the share of workers in an occupation reporting AI use at work. The task measure is the share of workers who perform a listed task and report using AI for it. Respondents saw the ten O*NET activities rated most important for their occupation, not a complete inventory of each person’s work. These adoption measures answer different questions from exposure scores. Keep six stages separate. Capability asks what a system can do under specified conditions. Exposure maps that capability against work tasks. Adoption means workers report using AI; organizational rollout concerns whether an employer provides or permits it. Task redesign means duties or accountability actually change. Labor demand concerns hiring or employment patterns; displacement is work lost through substitution. Evidence at one stage does not prove the next. The working paper summarized by Columbia Business School, “Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage,” reinforces that boundary. Using a representative German employee survey linked to worker and establishment information, it finds a weak relationship between prominent exposure measures and worker-reported use. Its proposed model accounts for 60% of cross-occupation variation in observed adoption, compared with 14% for exposure alone. Those are model-fit results for that study, not causal estimates or a formula for any individual. Read a ranking as a prompt to inspect your task bundle; it cannot predict individual adoption, redesign, or employment.

Sources: What Work Does Generative AI Do? | St. Louis Fed; Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage | Columbia Business School

What evidence shows adoption varies within occupations?

Worker-level evidence shows that exposure and occupation describe only part of who uses AI at work. In the St. Louis Fed’s “What Work Does Generative AI Do?”, occupation adoption means the share of workers in an occupation who report using AI for their jobs; task adoption means the share of workers who perform a listed task and report using AI for it. The distinction matters: a task rate is conditional on people actually doing that task, while an occupation rate averages across workers with potentially different assignments. The Fed analysis draws on nearly 14,000 workers in a nationally representative U.S. survey, interviewed across four quarterly waves from August 2025 through May 2026. Respondents identified their occupation, then saw the ten activities O*NET rates as most important for that occupation. They indicated which activities they performed and, among those using AI at work, which tasks AI regularly helped them complete. Adoption is widespread but shallow: at least half of workers use AI in only 40% of occupations, while fewer than 3% of listed tasks have AI use reported by more than half of their performers. These findings describe U.S. occupation and task denominators, not every task in a worker’s day. The study also finds that exposure scores can rank occupations or tasks with higher adoption reasonably well, yet do a poor job of predicting whether a given individual adopts AI. Within most occupations and tasks, some workers report use while most do not. Work type indicates where use may be more common, but not which coworker uses AI. The task list is selective: respondents were asked about ten high-importance O*NET activities, not to provide a complete work diary. It measures self-reported use, not verified tool logs, employer-wide rollout, productivity, changed responsibilities, or job loss. A European comparison points to context around exposure. “From Exposure to Adoption: Generative AI in European Workplaces,” a working paper using the 2024 European Working Conditions Survey across 35 countries, reports adoption ranging from under 3% to 25%. Occupational exposure strongly predicts uptake, but the exposure-to-adoption relationship is steeper where workers have relevant skills, jobs include more abstract tasks, employees have greater organizational influence, and countries have more digitalization and workplace training. These associations in a different population and period do not establish that any condition causes an individual to adopt. The study’s country rates should not be averaged with the Fed’s U.S. figures: definitions, labor markets, and periods differ. Together, the studies support a limited but useful conclusion: variation within occupations is real, and exposure alone cannot explain it. A job title groups workers, but it does not tell you who handles routine first drafts, who works through unusual cases, how often either task occurs, or who has authority and time to use an approved tool. That missing detail is why a broad occupation signal can guide what to inspect, but cannot diagnose a specific workplace. To understand your own situation, check the tasks you actually repeat and the conditions around using and verifying AI; the population evidence establishes the question, not the personal answer.

Sources: What Work Does Generative AI Do? | St. Louis Fed; Generative AI at Work: From Exposure to Adoption across 35 European Countries

Why can two workers face different costs for the same task?

The same task can make AI worthwhile for one worker and not another. Adoption is a choice under constraints: expected help must outweigh setup, prompting, checking, correction, risk and time, relative to what the worker can already produce. An exposure measure can indicate that a task may be technically affected. It does not show whether using a tool improves a worker’s complete result after the extra work of directing and verifying it. The working paper “Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage,” using Germany’s representative DiWaBe employee survey linked to worker and establishment information, finds a weak relationship between worker-reported AI use and prominent exposure measures. Its explanation is economic: feasibility is not the same as profitability. A tool’s potential productivity gain must be weighed against its user costs, while the worker’s own productivity is considered alongside pay. If a worker can already complete a task quickly and reliably, an AI draft that needs extensive repair may offer little advantage. If another worker faces a slower first pass and can check the output efficiently, the balance may differ. The study operationalizes this idea at task level, estimating worker productivity relative to pay, mapping exposure measures to AI productivity, and inferring task-specific user costs from observed adoption. Its resulting occupation-level index accounts for 60% of cross-occupation variation in observed adoption in the study, compared with 14% for an exposure-only model; the researchers also report substantial divergence between approaches for about 30% of workers. These figures describe explanatory fit in this German dataset and model. They do not mean that comparative advantage causes a stated share of one person’s behavior, supply a personal adoption formula, or predict displacement. Imagine two analysts with the same job title. One prepares standardized first drafts from consistent inputs. Another spends more time resolving unusual cases, reconciling conflicting records and taking responsibility for the final recommendation. A tool might provide a useful starting point for both, yet the review burden and error costs differ. The first analyst may be able to check a draft against a repeatable template; the second may need context the tool does not have and must defend the result. This is an illustration, not study evidence. Worker experience may also affect the cost calculation. “What Work Does Generative AI Do?” from the St. Louis Fed draws on nearly 14,000 U.S. workers surveyed over four quarterly waves from August 2025 to May 2026. It reports that workers with at least six months of generative AI use adopt it for more work tasks, an association the authors say is consistent with costly learning or experimentation. That pattern makes practice a plausible way to lower setup costs, but it does not prove tenure causes adoption: people who persist may differ in other ways, and the survey is not a controlled test of training. For a task you perform, ask both whether AI can help and whether it improves your complete task after setup, checking, correction and risk are counted. Compare that effort with your current way of working, and include who must trust or sign off on the result. Exposure identifies a possibility to examine; comparative advantage explains why the same possibility can have different value across workers. Neither settles whether an employer will adopt it or whether a job will change.

Sources: Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage | Columbia Business School; What Work Does Generative AI Do? | St. Louis Fed

Which workplace conditions change the exposure-to-use link?

Skills and discretion can help a worker act on a task where AI appears useful, but access, rules, data suitability, training and workflow determine whether use is allowed and workable. Research associates some conditions with adoption; it does not show that one factor explains every difference between coworkers. A worker who cannot use a tool safely or officially should clarify the rules and task before assuming a skill deficit or buying a course. A 2026 working paper, “From Exposure to Adoption: Generative AI in European Workplaces,” analyzes 2024 European Working Conditions Survey data covering more than 36,600 workers in 35 countries. Its abstract reports that occupational exposure predicts uptake, while worker skills, abstract task content and influence over organizational decisions steepen the exposure-to-adoption relationship; country digitalization and workplace training do so as well. These findings suggest that exposure is more closely connected to reported use where enabling conditions are present. They are statistical relationships, not proof that training or discretion causes an individual to adopt AI. It cannot diagnose a particular employee. Workers may differ in assignments, time to learn, choice of method and ability to judge outputs. But skill cannot compensate for a rule that bars client material, a workflow with no approved tool, or a task whose errors are too costly to check. Permission alone does not make a tool valuable if it performs poorly or adds more review effort than it saves. This points to context, not a universal checklist. The OECD’s 2025 report “Generative AI and the SME Workforce” describes organizational friction. It surveyed 5,232 small and medium-sized enterprises in seven countries in late 2024, interviewing the person with the best overview of company technology, often an owner or manager. Among firms not using generative AI, 57.3% cited poor fit with their work, 54.1% cited copyright, legal or regulatory concerns, 52.5% cited concern about information entered into tools, and 49.8% cited employees lacking the right skills. These are firm-reported reasons, not measured causes of individual behavior. Data, legal or task-fit constraints may remain even when a worker is capable. The OECD report says employee participation in AI-related training was not common among SMEs already using generative AI. This indicates a gap between some use and systematic preparation, but does not establish how much training a particular worker needs. Separately, the OECD’s “AI and Skills” synthesis describes employer-funded training as a common response and reports an association between training and positive worker-reported outcomes. It draws together studies with differing samples and dates, so supports treating training as one part of adoption, not a guarantee of better performance or job security. Identify the recurring task, then check for an approved tool and suitable data, who can authorize a trial, how outputs must be reviewed, and whether the team has time or training. If policy prohibits use or the task is a poor fit, ask about approved practice or workflow redesign before paying for broad retraining. If access is permitted but you lack a specific skill, choose a small learning step tied to that task and its verification needs. These checks do not explain why a coworker adopts; they help distinguish a learnable gap from a workplace constraint or a sensible decision not to use AI.

Sources: Generative AI at Work: From Exposure to Adoption across 35 European Countries; Generative AI and the SME Workforce; AI and Skills

Why do exposure scores, surveys and product logs disagree?

Exposure, reported use, perceived capability and product activity are different measurements, so apparent disagreement need not be contradiction. Exposure maps occupational tasks to a defined technical capability. Surveys record what people say they do or believe; product logs record activity on one service. None is a universal adoption rate, and none alone shows that a job has been redesigned or eliminated. The St. Louis Fed’s “What Work Does Generative AI Do?” uses nearly 14,000 U.S. workers surveyed across four waves from August 2025 to May 2026. It measures reported job use by occupation and reported AI use among workers performing listed O*NET activities. Respondents see ten important O*NET activities for their occupation and identify which they perform and where AI helps. This links use to tasks, but is not a full work diary or a measure of performance quality. The Fed finds exposure scores help rank adoption across occupations and tasks but poorly predict an individual worker’s adoption. This identifies a measurement gap; it does not explain why a particular coworker uses AI. Pew Research Center’s “Workers’ exposure to AI” asks a different set of questions in an October 2024 U.S. worker survey: how much work respondents say is done with AI, whether they have heard of workplace use, and how much of their job they think AI could do. Pew reports 16% said at least some work was done with AI, while 31% of non-users said at least some could be done with AI. That is reported use versus perceived applicability, not technical exposure; interpretations of “AI” and “can be done” vary, and awareness is not a workflow audit. The older survey is not a current adoption estimate. Anthropic’s “Anthropic Economic Index report: Cadences” links a survey with sampled Claude sessions from mid-May to early June 2026. The linked sample includes about 9,700 respondents, excluding people with fewer than five sessions. It compares reports with activity among these users, not workers generally. Anthropic says its survey is not representative: computer and mathematical occupations are about 30% of respondents, versus 4% of U.S. employment. Logs omit other tools and offline work; a work message proves neither employer adoption nor role change. The German DiWaBe analysis in “Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage” links worker-reported use with worker and establishment data. “From Exposure to Adoption: Generative AI in European Workplaces” analyzes 2024 survey data across 35 countries and reports worker, job and country correlates. These settings broaden the evidence, but their outcomes should not be pooled as if they sampled the same people or asked the same question. When a chart seems to disagree with your workplace, check who was counted, what they were asked, when they answered, and whether the measure records potential, perception, reported use or product activity. Use the figure as context for examining your own tasks, not as an explanation of coworkers or a forecast about your job.

Sources: What Work Does Generative AI Do? | St. Louis Fed; Workers’ exposure to AI; Anthropic Economic Index report: Cadences; Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage | Columbia Business School; Generative AI at Work: From Exposure to Adoption across 35 European Countries

Can adoption change a job before its title changes?

Yes. A job can acquire new tasks, lose old ones or redistribute review work while the job title stays fixed. But a record that someone used AI does not show that their employer redesigned a workflow, changed staffing, or displaced anyone. Those are separate outcomes, and each needs its own evidence. One possible path is task expansion: a worker uses an AI tool to help with an activity usually associated with another occupation, then returns to it often enough that it becomes part of the routine. OpenAI’s “How workers are unlocking new ways of working” analyzes more than 1.5 million work-related ChatGPT messages from April through July 2026. In a consistently observed group of roughly 6,200 users, the share of occupation-specific AI activity involving previously used cross-occupation tasks rose from 13.1% in April to 25.9% in July. This is evidence of recurring activity within that product’s observed conversations, not a representative estimate of workers or proof that employers formally reassigned duties. The distinction matters because recurring tool use can mean several things. A person might use AI to draft material for a task already in their role, to take on a new adjacent task, or to speed up a step while keeping the same responsibility for checking and approval. Only the second necessarily expands the task mix, and none by itself establishes that review duties, output expectations or headcount changed. Other evidence cautions against treating adoption as a synonym for restructuring. The working paper “From Exposure to Adoption: Generative AI in European Workplaces” examines worker survey data across 35 European countries and reports adoption rates ranging from under 3% to 25%. Its shift-share analysis found no detectable effect of adoption on worker-reported task restructuring. That is a specific null result for the paper’s measure and design, not proof that no jobs changed; the authors describe the evidence as consistent with an initial integration phase. The results can coexist: some tool activity recurs even as measurable reported task restructuring remains undetected in this analysis. Staffing is another step removed. In its 2024 survey of 5,232 SMEs across seven countries, the OECD asked a knowledgeable company respondent about generative AI’s effect on staffing needs. Among SMEs using it, 83% reported no effect on the overall number of staff needed; 9.1% reported a decrease and 5.5% an increase. The survey did not measure the size of those changes, so the figures cannot establish a net employment effect. This is why a use event or new task should not be read as a staffing decision: workload, contractor use, staffing needs and individual employment are distinct outcomes. For a worker, stronger signals are observable changes over time: recurring assignments that were previously outside the role, who now checks or signs off on the output, whether expected volume or turnaround changes, and whether managers alter hiring or redistribute work. None alone proves displacement. The useful conclusion is narrower: adoption can precede a title change, yet the path from use to routine task, changed responsibility and staffing outcome is not automatic. Track the actual allocation of work and accountability; treat headlines about AI use as prompts to observe, not verdicts about your position.

Sources: How workers are unlocking new ways of working; Generative AI at Work: From Exposure to Adoption across 35 European Countries; Generative AI and the SME Workforce

What should you check before making a career move?

Start with a representative week of recurring work, not the most dramatic task in your job description. For each task, note its input, the judgment or transformation you add, who relies on the result, and what an error would mean. An occupation-level exposure measure can point to tasks worth inspecting, but the St. Louis Fed’s “What Work Does Generative AI Do?” distinguishes workers in an occupation from workers who perform a listed task. Neither inventories your week. Frequency and stakes distinguish routine tasks from occasional high-consequence work. Next check access. Is a tool approved, may the data be entered, and what rules govern client information or review? If access is restricted, check policy before reading non-use as a skill gap. The OECD’s “AI and Skills” synthesis places skills and training alongside worker voice, privacy and accountability. A course cannot grant access or settle policy; seek clarification or workplace support. For an allowed task, compare the whole workflow with and without assistance. Count setup, instructions, checking, correction, handoff and sign-off, not only the time to produce a first draft. Define an acceptable result and check whether you can spot plausible errors. Experience may help review, but does not make every output safe. The St. Louis Fed evidence measures reported use, not the quality or net value of each use. A low-stakes work sample under workplace rules tests fit more directly than an exposure label. Keep sensitive work out of unapproved experiments. Then look for observable change: different assignments, review duties, expected volume, turnaround or accountability. One colleague’s use does not show team redesign. Revisit these signals over time. If the task changes but you remain responsible for review, identify whether your gap is tool practice, evaluation, policy knowledge or time. Match one next step to that gap and your goal. To use AI in your current field, try a permitted task through a small project that includes verification and feedback. For an adjacent role, compare its tasks and entry requirements with your experience before paying for training. A course offers structure; a project demonstrates applied work; a credential may matter when employers or local rules require it; self-study can test interest. Compare prerequisites, feedback, time and cost. None guarantees readiness. Consider a larger change only after finding a real mismatch or opportunity, weighed against salary needs, location, health, family duties and time to learn. Move from task evidence to a reversible step, then reassess. If no practical gap appears, defer costly retraining; if one does, address it before treating exposure as a reason to leave.

Sources: What Work Does Generative AI Do? | St. Louis Fed; AI and Skills

What is the next proportionate step?

Treat an occupation exposure result as a reason to inspect one recurring task, not as a verdict about your job. Check whether your employer permits the relevant tool and data use; if it does, compare the complete task before and after a small, reviewable trial. Include setup, checking, correction and accountable sign-off in that comparison. If access is not permitted, ask what approved process, training or safeguards apply before experimenting.

Then watch for workplace evidence that changes the decision: assignments moving, review duties changing, output expectations rising, or responsibility being redistributed. A single use event does not establish redesign or displacement. The St. Louis Fed’s “What Work Does Generative AI Do?” finds that exposure alone does not definitively predict reported adoption; its U.S. worker measures cannot tell you what will happen to your role. Mandated rollout or restructuring is an important exception: your immediate task may be to clarify new responsibilities, training and safeguards, rather than choose whether to try a tool. If you want a structured task-level starting point, the free checker reports transparent change-pressure signals, not validated odds of displacement. A larger career decision can wait until your task or workplace evidence identifies a real gap; one bounded next step is enough for now.

Sources: What Work Does Generative AI Do? | St. Louis Fed

Questions readers ask