AI Exposure at Work: Tasks, Jobs, and Your Next Move

What does AI exposure mean for tasks versus jobs?

AI exposure is a signal about tasks: it estimates how far a defined AI capability may be able to perform or alter particular activities that make up work. It is not a verdict about a whole occupation, workplace, or individual worker’s future. A job title groups activities together, and people with the same title can have different responsibilities and workflows. An occupation-level estimate therefore summarizes potential change across a bundle; it cannot establish that the job will disappear or that a worker will lose employment. Use it to identify recurring tasks worth examining, then check whether the work around them is actually changing.

A task is a particular activity; an occupation is a standardized category that collects many activities under one label. In “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” the International Labour Organization describes an assessment built from task-level data to estimate exposure to generative AI. Such an index can map where potential task change may be concentrated. Its result depends on the capability and task framework assessed; it does not report how a particular employer organizes work. The ILO’s later brief, “Workers’ exposure to AI: What indicators tell us—and what they don’t,” cautions that exposure indicators should not be read as predictions of displacement. Together, these sources support using exposure as a prompt for investigation, not an individual risk score.

For a worker, the useful question is not simply, “Is my occupation exposed?” Ask instead: Which activities recur in my week, which are changing in practice, and what review, context, coordination, or accountability still sits with a person? A high exposure label may justify inspection and preparation. By itself, it does not justify concluding that the role is replaceable, committing immediately to retraining, or deciding to leave. Treat the estimate as an early map of potential task change, to be checked against actual work and local evidence.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; Workers’ exposure to AI: What indicators tell us – and what they don’t

What does an exposure score actually measure?

An exposure score estimates how far a defined AI capability might perform or change tasks listed for an occupation. It measures technical susceptibility under a particular method, not what happens inside one employer. It does not directly measure whether a tool works reliably in a live workflow, whether a firm has adopted it, how much work will be delegated, or whether employment will change. Treat it as a map of tasks worth inspecting, not a verdict about a job or worker. Researchers first define the technology and describe work as tasks. They assess how well that technology could handle each task, using choices that may include expert judgment, worker reports, or model predictions. Task-level assessments are mapped to occupational categories and summarized. The result depends on which tasks are included, what “could perform” means, whose judgments count, and how scores are combined. Two indices can disagree without one being useless: they may measure different capabilities or apply different assumptions. The International Labour Organization’s 2025 working paper, “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” makes these choices visible. It updates an index specifically for generative AI, not all AI or automation. Researchers gathered perceived automation potential for 2,861 tasks from 1,640 employed respondents, sampled from Poland’s occupational classification. They compared these responses with an international expert survey and Delphi-style discussions, then used the resulting knowledge to develop an AI assistant that predicted scores for tasks in ISCO-08 occupational descriptions. This is a structured estimate based on task evidence and judgments, not direct observation of firms replacing workers. The ILO’s companion brief, “Generative AI and jobs: A 2025 update,” describes four increasingly exposed gradients. They reflect mean exposure and task-score variability. Similar averages can conceal different task patterns. The category result cannot tell a worker which activities dominate their week. The brief reports a mean automation score of 0.29 in 2025, compared with 0.30 in 2023, while the standard deviation fell from 0.30 to 0.14. It notes changed scores for some tasks amid expanded generative AI abilities in voice, image, and video generation. These figures describe this index’s framework and revision. They are not percentages of work automated, observed adoption rates, or headcount forecasts. Its scale is not a universal meter. The 2026 ILO brief, “Workers’ exposure to AI: What indicators tell us – and what they don’t,” says indicators vary by method, rely on static descriptions of existing tasks, embed judgments, and omit factors such as economic feasibility and institutional barriers. A capability estimate cannot show whether an employer can integrate a system, whether outputs meet required standards, or how costs and demand will respond. A generative-AI index also does not settle what robotics, conventional software, or future tools may do. When reading a score, check five things: the capability assessed; the task taxonomy and occupation coding; the date and population; whether results are task-level, averaged, or grouped into gradients; and the judgments or predictions used. Then compare it with your recurring tasks. The score can direct attention; your workplace’s process, quality requirements, and task allocation determine what is changing for you.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; Generative AI and jobs: A 2025 update; Workers’ exposure to AI: What indicators tell us – and what they don’t

Why can the same occupation contain very different exposure?

An occupation-level exposure estimate groups tasks under a standardized job category. It cannot guarantee that every worker with the same title does the same work. One person may draft or process documents; another may resolve unusual cases, coordinate with clients, review work, or make accountable decisions. Task shares vary with seniority, customers, regulation, and responsibility. A title points to a broad pattern, not a complete job description. The International Labour Organization’s “Generative AI and jobs: A 2025 update” makes variation part of its method. It summarizes analysis at the six-digit occupational level covering nearly 30,000 tasks, and describes four exposure gradients based on mean exposure and task variability within an occupation. The brief reports overall automation scores slightly lower than in 2023, while variability was also lower. This is a modeled assessment of generative AI against a defined task framework. It does not observe every employer’s workflow, and its gradients are not personal risk bands assignable from a title. The ILO’s 2026 brief, “Workers’ exposure to AI: What indicators tell us – and what they don’t,” says exposure patterns show substantial heterogeneity within occupational groups, while indices differ in method and definition. Its summary notes that measures rely on static task lists and embed judgments, while omitting adoption conditions such as economic feasibility and institutional barriers. The same occupational average can coexist with different combinations of addressable and less addressable work. A category can show exposure without telling us which worker performs which tasks, how often, or with what checks. Four distinctions help make the signal usable. Frequency asks how often an activity recurs. Time share asks how much of the week it occupies. Consequence asks what an error costs. Coupling asks whether it stands alone or depends on context and handoffs. These are practical questions, not a validated formula or a way to calculate displacement odds. An occasional, easily checked task can matter differently from a daily bottleneck whose output feeds later decisions. Imagine a records role where routine document classification is a visible, repeatable activity. An exposure indicator might make that task worth examining. The role could also involve mismatched records, missing context, source-document checks, and explaining exceptions. If classification becomes faster, work might shift toward checking; alternatively, a standardized workflow might reduce some steps. The estimate cannot tell which design an employer will choose, whether a system is reliable enough, or how much review remains. This is an illustration, not a reported case. Task coupling is easy to miss when a score is read as a list of independent activities. Changing one step can remove a handoff, create a checking obligation, move a bottleneck, or leave the rest untouched. Effects depend on where a task sits in a workflow. Exposure indicators identify potential susceptibility within classified work; they do not map local inputs, approvals, exceptions, or responsibility. Use occupation research to decide where to look, then inspect recurring work. Write down activities that fill a normal week, including coordination and review. For each, note frequency, time share, consequence of error, and dependence on other work. Compare that list with the exposure signal. A mismatch shows why a standardized category may be a poor proxy for your role. The task list gives you a basis for checking workflow evidence without mistaking a common title for a complete audit.

Sources: Generative AI and jobs: A 2025 update; Workers’ exposure to AI: What indicators tell us – and what they don’t

What steps stand between capability and a changed job?

Technical capability is only the first step. A workplace must also find the system reliable and economical enough, fit it into a process, decide who checks exceptions and errors, and choose what to do with the time or output gained. Those choices can produce assistance, task substitution, new review work, changed demand, or no meaningful adoption. An exposure estimate describes a possible opening for change, not its result. The ILO brief “Workers’ exposure to AI: What indicators tell us – and what they don’t” makes this distinction directly: indicators capture technological susceptibility under a static view of tasks, while omitting whether firms find automation profitable, how workflows change, and how employment, wages, or demand adjust. It treats exposure as a signal of transformation, not a forecast of displacement. Between capability and work redesign are practical gates. A tool must handle the specific inputs and outputs well enough for the task. A plausible draft or classification may still need correction; where errors carry material cost, review time and accountability affect whether using it is worthwhile. An employer must also integrate the tool with existing systems, information access, permissions, and handoffs. It can stall when data, process fit, or exception ownership is missing. If a tool is adopted, work can change in more than one way. With augmentation, a worker uses the system inside a task, perhaps to prepare material that they check, interpret, and adapt. With substitution, the system performs some work previously done by a person. Both can coexist: routine first-pass work may be delegated while exception handling, verification, or communication remains with staff. That describes task allocation, not net staffing. An employer might use saved time for other work, raise output expectations, reduce a handoff, add review duties, or change headcount. Demand can also move in more than one direction. If productivity lowers costs, an organization might expand output or serve work that was previously uneconomic; it might also need fewer labor hours for the same output. Neither result follows automatically from exposure. The ILO brief notes that indicators omit productivity gains that could lower costs and expand demand, while cautioning that this possibility does not make an exposure estimate a job-growth forecast. The U.S. Bureau of Labor Statistics explains a related boundary in “Artificial Intelligence (AI) impacts on employment projections.” Its projections cover about 800 occupations over a ten-year horizon and describe expected outcomes under specified assumptions, rather than forecasting exactly what will happen. BLS says technology adjustments are generally conservative when evidence is convincing, and that uncertainty about AI’s labor-market effects remains high. They inform aggregate demand, not whether one employer adopted a tool or how one worker’s job will change. For an employee, the signal becomes more decision-relevant when local evidence appears: work standards change, output targets rise, review ownership shifts, staffing plans are communicated, or tasks move between roles. A trial alone does not establish lasting adoption. Look for what has actually changed in the workflow and whether it recurs. If expectations, responsibilities, and staffing remain stable, a score alone does not justify treating displacement as imminent. If they change together, investigate how the role is being redesigned and which valued responsibilities remain.

Sources: Workers’ exposure to AI: What indicators tell us – and what they don’t; Artificial Intelligence (AI) impacts on employment projections

How do observed use and local adoption differ from exposure?

Usage evidence answers a narrower question than exposure does: what people are doing with a particular system, in the observed setting. It does not establish that an employer has adopted that workflow across a team, changed staffing, or reduced demand for a role. Local evidence is more useful for deciding what to do next, but it still needs interpretation. A changed output target, review duty, workflow owner, hiring requirement, or staffing plan is a sign to investigate; none on its own proves that jobs will be lost. The June 2026 Anthropic Economic Index report illustrates both the value and the boundary of platform data. It analyzes sampled Claude conversations across products and reports task and output patterns, including differences in how much users delegate on different product surfaces. It also links usage to a survey of about 9,700 respondents. Within that sample, people whose Claude sessions were more automated reported higher current and anticipated exposure, while respondents in more automated use patterns also expressed more optimism about future work outcomes. These are patterns among Claude users, not a representative estimate of all workers or employers. The report explicitly says its linked survey is not representative of the general population; its sample over-represents computer and mathematical occupations and management relative to U.S. employment. The association cannot establish whether delegation changes a worker’s outlook, optimistic users delegate more, or other factors shape both. Most of all, platform activity shows interactions with one provider’s tools, not a company-wide adoption decision or a resulting employment change. Keep four signals separate when assessing a concern. Exposure research identifies task families worth inspecting. Platform or worker-use evidence shows that some users apply a tool to particular activities. Employer evidence concerns the actual process: who now produces, checks, approves, and owns the output, and whether targets, required skills, or staffing have changed. Labor-market evidence tracks broader employment or wage patterns, which may reflect many forces. The U.S. Bureau of Labor Statistics’ employment projections, for example, cover about 800 occupations over a ten-year horizon and describe outcomes under stated assumptions; BLS says they are not forecasts of what will happen, and notes high uncertainty about AI’s future effects. Such projections provide context, not a diagnosis of one workplace. For an individual decision, record specific changes and their dates rather than relying on a job title or a colleague’s tool use. Check workflow documents, manager communications, vacancies, review expectations, output targets, and staffing plans for corroboration over time. Ask whether a new tool is optional assistance, a required step, or a reason to reassign responsibility; note who handles errors and exceptions. This local checklist is a practical way to gather evidence, not a validated prediction rule. One changed task may be a trial; repeated changes across responsibilities and targets provide a stronger reason to discuss expectations or compare options. Treat exposure as a map for where to look, usage as evidence of what some users do, local process changes as evidence of what this employer is doing, and employment data as a broader outcome check. Escalate from concern to a career pivot only when the pattern and your constraints justify it.

Sources: Anthropic Economic Index: June 2026 report; Artificial Intelligence (AI) impacts on employment projections; Workers’ exposure to AI: What indicators tell us – and what they don’t

What can employment projections add—and what can’t they settle?

Employment projections show the direction and scale of expected aggregate occupational demand under stated assumptions. They cannot establish that AI alone caused a projection or tell an individual what will happen at one employer. Read exposure and demand side by side: an exposed task can sit inside an occupation projected to grow, while a projected decline may reflect several forces rather than tasks directly transferred to software. Exposure concerns susceptibility of tasks to a defined capability; projections estimate employment across occupations and an economy over time.

The U.S. Bureau of Labor Statistics’ 2024–34 projections illustrate the difference. Its July 2026 summary, “Artificial intelligence, information technology, and employment, 2024–34,” projects data scientist employment to rise 33.5 percent and customer service representative employment to fall 5.5 percent. The table also projects growth for medical records specialists and decline for medical transcriptionists. These contrasts show that occupations involving information can have different projected directions. They do not establish AI as the cause of each result or show how every worker will be affected. BLS discusses productivity and technology adoption as factors, not as a calculation of personal outcomes.

A projection is conditional, not a promise. In “Artificial Intelligence (AI) impacts on employment projections,” BLS says its ten-year projections cover about 800 occupations and use historical trends alongside assumptions about the labor force, economy, industries, and occupations. They describe what would be expected if those assumptions hold; actual values may differ. BLS says uncertainty about AI’s labor-market effects remains very high, adjusts for emerging developments conservatively when evidence is convincing, and updates projections annually. Its AI case-study interpretation uses information available as of June 2024. Treat the projection as a dated planning reference, not a precise forecast of one employer’s redesign.

Exposure and demand are different quantities. An exposure index starts with tasks and estimates their susceptibility to a specified capability. An employment projection estimates net occupational headcount after multiple expected influences. Even if a task can be automated, an employer might use the capacity to expand output, shift effort to other work, or reduce staffing; an exposure index cannot determine which. A projected decline also says nothing by itself about remaining workers’ duties, pay, or hours. Aggregate totals do not trace any one person’s transition or establish job quality.

Skill-demand research adds another lens. The OECD working paper “Artificial intelligence and the changing demand for skills in the labour market” examines skills in occupations that do not require specialized AI expertise. Its abstract says management and business skills are among the most demanded in highly exposed occupations, and the share of vacancies in those occupations requesting at least one emotional, cognitive, or digital skill rose by eight percentage points over the period studied. It also reports panel-establishment evidence that demand for these skills may be beginning to fall. The mixed findings call for attention to measures and context, not a claim that one skill bundle is permanently rising or falling. Vacancy patterns are not task exposure, headcount projections, or guaranteed openings for a particular worker.

Outside the United States, use BLS figures only as directional context. Check your national statistical agency’s occupation definitions, assumptions, projection horizon, and labor data. Then examine openings where you can realistically work, required skills and credentials, salary fit, and transition time and cost. Exposure points to tasks worth inspecting; national projections add broad demand context; local evidence tests whether a next role fits your circumstances. None alone predicts personal displacement.

Sources: Artificial Intelligence (AI) impacts on employment projections; Artificial intelligence, information technology, and employment, 2024–34; Artificial intelligence and the changing demand for skills in the labour market

How should a worker turn task exposure into a proportionate move?

Start with recurring work, not the occupation label. List three to five activities that take meaningful time or affect the result. For each, note whether inputs and outputs are digital and repeatable, what a correct result requires, how costly an error would be, how much context or trust is involved, and who is accountable. An exposure estimate can help choose what to inspect; it cannot show how your employer has divided the work or what it will do next. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us – and what they don’t” explains that exposure indices use static task lists and do not establish feasibility, workflow change, or resulting demand and employment.

Imagine a knowledge-work process that turns documents and meetings into a recommendation. Producing a first draft or grouping routine material may be a candidate for AI assistance. But someone still has to check whether sources support the summary, reconcile conflicting records, spot exceptions, understand stakeholder needs, and take responsibility for the recommendation. This is an illustration, not a report about a particular employer. The draft, check, exception, and decision remain separate activities even when one job title bundles them. Exposure in one activity does not settle what happens to the whole role.

For each activity, ask what would make assistance workable: Are inputs available in usable form? Can results be checked against a source or standard? Can errors be caught before they affect a decision? Is sensitive information permitted in the workflow? Does review take less effort than doing the task without assistance? The ILO’s 2025 “Generative AI and Jobs: A Refined Global Index of Occupational Exposure” describes an index using task-level inputs, expert discussion, and model predictions, based on a representative sample of tasks in Poland’s occupational classification. It estimates potential exposure across defined tasks; it does not test your workflow or certify a tool’s reliability. The ILO’s 2026 indicators brief adds that feasibility and institutional barriers are outside exposure scores.

A useful test is small, reversible, and bounded by quality and privacy requirements. Choose one repeated task where experimenting is allowed, keep a human check, and compare a baseline with a two-to-four-week trial if the task occurs often enough. Decide what to observe: time to usable output, corrections, missed details, review effort, workload, or quality against the same standard. A faster output that needs more verification or causes downstream rework may not save effort. If there is no safe way to inspect quality or protect information, map the process or clarify requirements instead of testing on live work.

Measure checking as well as production because assistance and substitution can occur in different parts of a workflow. Anthropic’s June 2026 Economic Index report analyzes sampled use of its own products and finds patterns that vary by product and output, . These records show activity on one provider’s services, not general employer adoption or staffing outcomes. For your own trial, keep quality, review responsibility, and total effort in view alongside speed. If production becomes quicker, checking and exception handling may become more important, but only if the organization assigns and values that work; do not assume new review duties will protect a role.

Use what you learn to compare three paths. Staying and redesigning may fit when workflows change but your experience remains useful; build skill in a relevant tool while strengthening context, verification, coordination, or accountable judgment. An adjacent move may reuse your subject knowledge with a different task mix; check real role requirements, location, pay range, credentials, and transition cost. A larger change may warrant investigation if important duties or demand appear to be shifting over time and the route fits your finances, health, and family. A broad exposure score alone cannot rank them.

Tie learning to a defined gap. If the trial shows you need to evaluate outputs, handle data, or automate a repeated step, a focused course, guided practice, or project may be more proportionate than a degree. If your goal is software practice or machine-learning engineering or research, deeper formal study may fit that different objective. A certificate can document completion; alone it does not establish workplace capability, hiring demand, or readiness for a specialized role. Set a review date and record what improved, what required review, and any local change in duties, targets, staffing, or required skills. Prefer the smallest reversible step that resolves your key uncertainty; consider a larger plan when repeated workplace evidence and your income, geography, time, health, and family constraints support it.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; Workers’ exposure to AI: What indicators tell us – and what they don’t; Anthropic Economic Index: June 2026 report

What is the next useful action?

Write down three to five recurring tasks before deciding what an occupational exposure label means for you. For each, note how often it recurs, what evidence suggests it is changing, and what still requires judgment, context, verification, or accountability. Separate an exposure estimate from observations such as changed workflow, review expectations, output targets, or responsibilities. “Workers’ exposure to AI: what indicators tell us – and what they don’t,” an ILO brief, explains that indicators based on current task descriptions do not establish adoption or account for workflow and economic adjustments. Treat the score as a reason to inspect, not a personal forecast. Choose one reversible step that fits the uncertainty. If workplace rules allow it, test a bounded workflow and record quality, rework, time, and review effort. If expectations are unclear, ask what work is changing and who checks results. If evidence points to a sustained shift, compare an adjacent role using your domain knowledge before committing to retraining. Set a review date after a few weeks. Then decide whether to stay and redesign, explore the adjacent route, or investigate a larger change. Respect income needs, location, learning time, health, and family responsibilities. The free task checker organizes task-level change-pressure signals and possible first actions; its result is not a validated probability of displacement. If you need to compare paths against your constraints, the paid career roadmap can structure scenarios and a 30/60/90-day plan. It does not guarantee employment or income. Start with a task you can verify, then revisit your decision when evidence or constraints change.

Sources: Workers’ exposure to AI: What indicators tell us – and what they don’t

Questions readers ask