How to Read Task-Level AI Data Without Forecasting Your Job

What exactly is being counted when a source says workers use AI for a task?

Before reading an AI-at-work chart, identify its measure, population, time window, and task-mapping method. A platform trace records activity on that service; a worker survey estimates reported use among respondents; an exposure model estimates where capability may apply. These answer different questions. None alone tells you whether your employer adopted a tool or your own job is changing. The initial Anthropic Economic Index is a platform trace. Its February 2025 report analyzed about one million Claude.ai Free and Pro conversations, mapping them to U.S. Department of Labor O*NET task labels and grouping those labels into occupations. Its unit was conversations: the reported 37.2% for computer and mathematical work meant that share of conversations in the dataset, not the share of workers in those occupations using AI. Anthropic said it did not survey workers or forecast the future. It also cautioned that activity might not be work, task mapping could be wrong, and the sample was not representative of AI use generally. Read it as activity on one service, shaped by its users. “What Work Does Generative AI Do?” from the St. Louis Fed uses a worker survey. Its Real-Time Population Survey asked respondents which listed O*NET tasks they performed and, among workers using AI on the job, which tasks AI regularly helped them complete. The article describes nearly 14,000 workers across four quarterly waves from August 2025 to May 2026. A task adoption rate is the share of people who perform that task and report using AI for it. This denominator asks how many task performers report use. Limits include self-report, a U.S. population, and the ten most important O*NET tasks shown for each occupation. Comparison: Anthropic counts Claude conversations mapped to tasks, supporting claims about that sample; the St. Louis Fed counts task performers reporting use, supporting claims about reported prevalence. The peer-reviewed “The collaboration code” clarifies a further denominator: task conversation share measures usage intensity among tasks where AI use is already observed, not the initial decision to adopt. Thus, a task appearing in platform data does not show general worker use; survey-reported use does not establish local deployment, its share of your workweek, or displacement.

Sources: The Anthropic Economic Index; What Work Does Generative AI Do?; The collaboration code: how humans and AI work together across millions of conversations on job tasks

What does an exposure score add, and what does it leave out?

An exposure score estimates where AI capability may apply by a method. It can help identify duties to inspect, but does not show whether your employer adopted a system, whether it can complete the whole task reliably, or whether jobs or demand will rise or fall. Read it as a scoped comparison, not a personal forecast. The International Labour Organization’s 2025 paper, “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” combines worker task input, expert discussion, and model predictions. Researchers gathered perceptions about automation potential for 2,861 tasks, using a representative sample from Poland’s occupational classification and a survey of 1,640 employed people across occupational groups. Expert surveys and Delphi-style discussions informed an AI assistant that predicted scores for tasks in the international classification. The four gradients summarize modeled applicability, not workplace rollouts. The paper estimates that one in four workers globally are in occupations with some exposure and 3.3% of global employment is in its highest gradient. These global estimates are not the share of a person’s week AI could take over. An occupation is a bundle of tasks, not one uniform job. A score can flag that some component may be assisted or completed by current technology, while saying little about its frequency in a particular role, the checking required, or who remains accountable. It cannot establish whether adoption is economically worthwhile or permitted. The ILO paper says transformation is the most likely general impact because most occupations still have tasks requiring human input. It does not rule out displacement in particular settings. The U.S. Bureau of Labor Statistics’ 2026 “Artificial Intelligence (AI) exposure categories” shows how another index is assembled. BLS combines five sources: three theoretical measures and two based on observed AI use mapped to occupational tasks or activities. It converts each source’s results to percentile ranks, calculates separate medians for theoretical and observed evidence, then clusters occupations into four relative categories. “Very high” means relatively higher exposure than other occupations here, not an absolute amount of work exposed. The inputs measure different things: links between AI capabilities and occupational abilities, potential task-time reductions, or product interactions mapped to occupational activities. BLS crosswalks source classifications to its own occupation system and imputes missing ranks. Its documentation reports 211 imputed values among 4,155 occupation-source combinations, affecting 75 occupations. Coverage and aggregation can affect a category alongside workplace reality. Time matters. BLS says theoretical inputs describe capabilities available no later than mid-2023, while observed product data may lean toward early adopters. The categories combine automation and augmentation rather than distinguishing their effects. A high category cannot tell you whether a tool drafts material for a person to verify or removes a workflow. A low category does not promise an occupation will avoid change. BLS says the categories are not forecasts of employment growth, wages, productivity, adoption probability, or worker replacement. The ILO’s 2026 brief, “Workers’ exposure to AI: What indicators tell us – and what they don’t,” adds: exposure indicators describe technological susceptibility under a static view of tasks, but omit economic feasibility, institutional barriers, workflow change, and labor-market adjustment. They cannot predict displacement, productivity gains, or reskilling needs. Productivity may affect costs and demand in ways a score does not model. Use high exposure to identify a recurring duty in your work that resembles the source’s task description, then check local evidence. Treat low exposure as no guarantee of immunity. Neither result gives your personal odds. If tasks match, check approved local use and whether review, responsibility, workload, or staffing has changed. The score tells you where to inspect; your work context tells you what to do next.

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; Artificial Intelligence (AI) exposure categories

Why can two credible adoption measures disagree?

Two credible adoption measures can disagree because they observe different products, users, periods, task labels, and stages of work. Treat disagreement first as a boundary question: what exactly entered each measure, and what does its denominator represent? A platform trace records activity on a particular service. A worker survey records what respondents say they do. Neither is a census of employer deployment, and neither automatically describes the typical worker who performs a task.

The June 2026 Anthropic Economic Index report makes the product boundary visible. Anthropic says sessions increasingly involve long-running agentic work through Claude Code and Cowork, so chat transcripts no longer fully capture use. It changed its pipeline to sample more often, added a classifier for conversation outputs, and reported chat and Cowork separately from first-party API activity. Those are methodological updates, not evidence that every workplace adopted those products. Consumer chat, coding software, and API calls routed through an organization can reflect different users and workflows. A task label assigned to a conversation also compresses varied work into an occupational category; it may not reveal the user’s job or whether the output was accepted, checked, or used at work.

The St. Louis Fed’s “What Work Does Generative AI Do?” measures a different thing. It describes nearly 14,000 U.S. workers surveyed over four quarterly waves from August 2025 to May 2026. Respondents identify their occupation, say which of the ten most important O*NET tasks for that occupation they perform, and workers who use AI at work identify tasks AI regularly helps them complete. Its task adoption rate is therefore the share of surveyed workers who perform a listed task and report AI help on it. That denominator is not the share of all work hours spent on the task, nor the share of employers that installed a tool. It also depends on workers’ recall, survey responses, and the task options shown to them.

A peer-reviewed analysis, “The collaboration code: how humans and AI work together across millions of conversations on job tasks,” illustrates another denominator trap. It defines task conversation share as the fraction of conversations in its dataset mapped to each task, and explicitly says this measures usage intensity among tasks already observed in use. It cannot model the initial decision to adopt AI. A task with many conversations might be receiving repeated help from a smaller user group; the measure alone does not say how many workers tried it or how common the task is in their occupation. Likewise, a survey can count workers reporting use without measuring how many iterations, minutes, or completed work products that use represents.

The meaning of “coverage” can shift even within one provider’s data. Anthropic’s “Economic Index: New building blocks for AI use” distinguishes tasks seen in conversations from effective coverage adjusted using task frequency, duration, and observed success. These adjustments can change which occupations appear more covered: frequently repeated tasks, longer sessions, and successful interactions receive different weight from a raw count of distinct task labels. That is useful for asking a more specific question, but the provider’s time and success estimates remain estimates from selected product interactions. They are not an independent audit of work quality, total labor saved, or employer outcomes. Task selection is also not random: users may bring tasks they think a system can handle, or divide larger assignments into smaller requests.

So compare measures only after aligning their units, dates, geography, population, platform, definition of use, and denominator. Look for convergence across unlike evidence to identify a broad task worth examining; do not average incompatible rates into a personal risk score. A conversation share, a worker-reported task rate, and a time-weighted coverage estimate answer different questions. Even agreement among them establishes neither local adoption nor displacement. For an individual decision, the next evidence must come from the actual workflow: whether approved tools are used repeatedly, what work they assist or complete, how much checking remains, and who remains accountable.

Sources: Anthropic Economic Index, June 2026 report; Economic Index: New building blocks for AI use; The collaboration code: how humans and AI work together across millions of conversations on job tasks; What Work Does Generative AI Do?

How should you compare an occupation task with your actual work?

Use an occupation task list as a vocabulary and coverage check, then build your own inventory of recurring duties. Compare what the source calls a task with what you actually do, how often it occurs, what inputs it depends on, what counts as an acceptable result, and who checks or owns the result. A title is not a complete description of your work. O*NET makes a useful starting index because its task statements connect a task to an occupation and classify it as core or supplemental. Its documentation says a core task meets stated relevance and importance thresholds; supplemental tasks are less relevant, less important, or both. It records respondent counts where available and update dates, which help show what the occupation description represents. They do not turn the list into a diary of your week: O*NET describes occupation-linked duties, while your seniority, client mix, team design, local rules, and recently assigned responsibilities may differ. A wording match is a prompt to inspect, not proof of equal weight or workflow. The St. Louis Fed’s “What Work Does Generative AI Do?” illustrates why task labels are better than job titles for locating a signal, while also showing why they remain broad. Its U.S. Real-Time Population Survey asked respondents about tasks drawn from O*NET’s detailed work activities; across nearly 14,000 workers in four quarterly waves from August 2025 to May 2026, its task adoption measure was the share of people who performed a task and reported using AI for it. This is not the share of your time spent on a task or the share of coworkers using an approved workflow. Its measure cannot tell you how often a matched duty appears in your week or how much of it a system handles. Consider an illustrative report-preparation workflow. A source may use a label such as preparing reports or gathering information. Your version could include drafting a summary, reconciling source records, applying client context, checking figures, explaining exceptions, and signing off. These related pieces have different inputs, error costs, review needs, and accountability. If a tool can help draft a paragraph, that does not establish that it can reconcile the records or authorize the final report. Assess each remaining piece by tool access, evidence, review, and responsibility; do not assume it is beyond AI. This illustration does not claim an employer uses AI in this process. A compact comparison can keep the evidence honest. For each task label, note your corresponding activity; its frequency or approximate share of your work; the inputs and acceptable output; whether an approved local tool is actually used; and any observed change in total time, quality, checking, rework, or handoffs. Note who remains responsible and what is unknown. If a source includes an exposure category, record that separately from adoption: the BLS methodology page describes occupation categories as comparisons built from multiple inputs, not a measure of your employer’s deployment or your individual outcome. A task match cannot fill the local-evidence column. Use the inventory as a task audit, not a job-loss calculator. Frequency helps prioritize attention, but does not determine importance, accountability, or whether a workflow can change. A task that appears in an occupation dataset might occupy little time in your role; another duty important to your role might be missing or described under a different label. Use the occupation list to notice overlooked work, then your own task record to correct the picture. Keep the gaps visible instead of converting them into a score. The result is a sharper question for the next step: is this task only a match in an aggregate source, or is there evidence that its local workflow has changed?

Sources: O*NET Task Statements; What Work Does Generative AI Do?; Artificial Intelligence (AI) exposure categories

What local evidence makes an aggregate signal actionable?

Look for evidence that the workflow is changing where you work: an approved tool is in use, recurring tasks pass through it, handoffs or review have changed, output or workload is measured differently, or responsibilities and staffing are discussed or documented. A demonstration or public adoption chart shows that a capability or use exists. Neither alone establishes routine integration at your organization. An aggregate task match identifies a place to investigate; local evidence tells you whether your own process is moving.

Keep six steps separate. Capability means a system can produce or assist with an output under some conditions. Observed use means someone used it; the St. Louis Fed survey asks workers who perform listed tasks whether they use AI for them. Employer adoption means an organization has authorized or integrated a tool. Redesign means the sequence of work, review, or responsibilities changes. Demand concerns how much of a service or output is wanted and supplied. Displacement is a labor outcome, not a synonym for an earlier step. The ILO brief “Workers’ exposure to AI: What indicators tell us – and what they don’t” says exposure indicators do not account for economic feasibility, workflow changes, or how employment, wages, and demand adjust. A high exposure signal cannot fill in those missing links.

The OECD’s “Skills in the AI age” describes firm adoption as uneven across firms, sectors, and countries, with differences in resources, infrastructure, and skills. It identifies three channels that can coexist: automating existing tasks, creating new tasks or occupations, and improving productivity. Their balance shapes employment effects. The same drafting aid might reduce first-draft time, let a team handle more requests, shift effort toward checking, or contribute to redesign. Aggregate evidence does not establish which result will occur locally. Nor does reported use establish that output is accurate enough, approved for the data involved, or accepted as a substitute for the whole task.

A useful counterpoint is that workers should not wait for perfect labor-market evidence when routine work occupies a substantial part of their duties and local changes are already sustained. If an approved system repeatedly handles that work and managers have documented altered responsibilities, review ownership, workload expectations, or staffing, those are stronger planning signals than a broad occupation score. They still do not prove what will happen to a particular person. Staffing decisions may be unsettled or unrelated to a tool; changed work can also increase output or shift effort. Treat the evidence as a reason to plan options, not as confirmation of an individual outcome.

Some evidence is visible in ordinary work: an approved workflow, templates, changed handoffs, review queues, rework, or altered expectations for volume and turnaround. Other information, such as future budgets or confidential staffing plans, may not be available. Do not infer hidden intent from a tool announcement. Ask an operational question: which step should the tool support, who checks its output, what remains the worker’s responsibility, and how will quality and workload be assessed? This invites an explanation without treating adoption as a redundancy notice. If policy permits a trial, compare the whole task, including checking and correction, rather than only first-output speed.

Use a qualitative convergence rule. A task match in aggregate data warrants attention and a check of your actual duties. A task match plus approved local use warrants examining quality, time, handoffs, and accountability. Repeated workflow change together with evidence of altered responsibilities or demand warrants structured planning. That planning can begin with an upgrade to your current role, an adjacent move, or a larger change, compared against your experience, salary floor, location, available learning time, cost, health, and family needs. Exposure alone is not a reason for a costly pivot; strong local evidence is not a personal job-loss forecast.

Sources: OECD Skills Outlook 2026: Skills in the AI age; Workers’ exposure to AI: What indicators tell us – and what they don’t; What Work Does Generative AI Do?

When is a small experiment enough, and when is career planning warranted?

A task match usually supports monitoring, an approved low-stakes trial, or a focused workplace question. Consider structured career planning when repeated local changes affect a meaningful share of your work or responsibilities and a feasible alternative is worth comparing. Aggregate data tells you where to look; evidence from your workflow tells you whether to do more. Neither establishes your employment outcome.

Use three steps, without turning them into a score. First, if a task appears in a dataset but you see no corresponding workplace change, record the source, date and measure. Note whether the task recurs in your role, then revisit it later. This is monitoring, not dismissal: a dataset may flag a pressure point before you have enough local information to judge its effect, but it cannot fill that gap for you.

Second, if your organization permits a tool for the task, compare the whole old and new workflow. Record the input, required output, review, handoffs, rework and accountability. Compare quality and end-to-end time, including checking and correction. A faster first draft may not shorten the task if verification takes as long as drafting or the draft adds downstream work. If review is brief and reliable, saved time might support more output, deeper checking or another duty. These are possibilities, not outcomes an adoption chart can establish.

The Anthropic Economic Index primitives report shows why the unit matters: its provider-derived measures distinguish task coverage from estimates adjusted for frequency, duration and observed success. Product activity is not proof that a large share of work is completed successfully. The report also describes selection effects: users may bring tasks they expect to work and divide larger tasks into smaller steps. This is not an independent workplace audit. Before a trial, check tool approval, data rules and required human review, and use only authorized material.

Third, if change persists and duties, workload, responsibility, staffing or demand appear to shift, compare realistic paths: stay and learn a needed tool or review practice; move to an adjacent workflow that uses your existing knowledge; or consider a larger change with its own entry requirements and costs. The OECD's Skills in the AI Age explains that automation, new-task creation and productivity effects can coexist and shape employment effects. The same task use can therefore lead to different organizational responses; the report cannot establish which response your employer will choose.

Planning is more warranted when evidence converges: the task recurs substantially in your work, local use is established rather than merely demonstrated, responsibilities have changed over time, and a plausible alternative exists. The ILO brief Workers’ exposure to AI: What indicators tell us – and what they don’t cautions that exposure indicators do not predict displacement, productivity gains or reskilling needs. Local evidence improves your decision, but even documented redesign cannot settle future demand or your individual outcome. It is a reason to compare options, not a verdict to leave.

Compare options against your constraints. A course may teach a bounded skill; a project can test its application; a certificate may offer structure if valued in your target area. A degree takes a different investment and may fit a role requiring deeper foundations. None is automatic, and a short course alone does not establish readiness for machine-learning engineering. Start with the observed gap, then check prerequisites, fees, time, location, salary floor, health and family responsibilities.

Let the response grow with evidence: no local change supports monitoring; permitted use supports a bounded comparison; sustained redesign or role and demand evidence supports planning across staying, an adjacent move and a larger change. When work changes quickly, gather information before every detail is settled. Do not resign, take on training debt or abandon useful experience because an occupation-level measure looks high. The measure selects a question; your workflow and feasible alternatives shape the next move.

Sources: OECD Skills Outlook 2026: Skills in the AI age; Economic Index: New building blocks for AI use; Workers’ exposure to AI: What indicators tell us – and what they don’t

What should you conclude from a task-level signal?

Use a task-level signal to decide what to inspect next, not to estimate your personal odds of losing work. A match can focus attention on a recurring duty; evidence about your workflow and role is needed to judge whether it is changing. The U.S. Bureau of Labor Statistics’ “Artificial Intelligence (AI) exposure categories” combines theoretical capability and observed AI-use measures, ranks occupations,. BLS says the categories do not directly observe whether workers in an occupation used AI on the job. They are not forecasts of employment growth, adoption, replacement, productivity, or wages. High exposure is a reason to inspect tasks, not a personal forecast; low exposure does not certify that a role is unaffected or safe. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us – and what they don’t” explains that exposure estimates describe technological susceptibility under a static view of tasks. They do not establish whether adoption is feasible, how firms redesign workflows, or how demand adjusts. Productivity can lower costs and expand demand, potentially changing labor effects; that is not a promise for a particular workplace. Capability, use, organizational choice, and labor-market response are separate steps. Do not dismiss aggregate signals: repeated matches across credible sources can flag a task worth monitoring,. But repetition matters only when measures, dates, populations, and definitions are understood. Related platform charts are not necessarily independent confirmations. The OECD’s “Skills in the AI Age” describes employment effects as shaped by automation of existing tasks, creation of new tasks, and productivity effects, across firms and sectors. This explains why AI assistance on a task alone cannot show whether an organization will reduce labor, increase output, shift review duties, or redesign work. Look for approved, recurring use; changed handoffs and checking; and documented shifts in responsibility, workload, or staffing. A practical rule follows. If a signal matches a duty but you see no local workflow change, record the task and source, then revisit it. If approved use recurs, compare the whole task before and after: preparation, checking, rework, handoffs, and accountability. If sustained workflow change coincides with documented shifts in duties, workload, or staffing, begin structured planning. Compare staying and redesigning, an adjacent move, and a larger change against your salary floor, location, learning time and cost, health, and family responsibilities. The conclusion should change with evidence. Independent measures pointing to the same recurring task, credible local deployment, and sustained changes in responsibilities or demand justify faster planning. A matched task rare in your week, no local workflow change, or use that adds review while you retain responsibility supports a smaller response. For one bounded next step, choose a recurring task named by a source and ask: what do I do, how often, what must be checked, and has an approved local process changed? A free task-level checker can organize change-pressure signals and first actions,; it is not a validated displacement probability. If converging local evidence makes career scenarios useful to compare, a personalized roadmap can structure that comparison; it guarantees no employment or income. Let your next move match what you can verify where you work.

Sources: Artificial Intelligence (AI) exposure categories; Workers’ exposure to AI: What indicators tell us – and what they don’t; OECD Skills Outlook 2026: Skills in the AI age

What is one useful next step this week?

This week, choose one recurring task that appears in an adoption or exposure source and write a short baseline: what starts the task, what you produce, how often it recurs, how long review and correction take, and who remains accountable. Then check whether an approved tool is part of your local workflow. A task label in the BLS AI Exposure Categories can help you decide what to inspect, but BLS says its relative categories are not adoption probabilities or forecasts. The OECD Skills in the AI Age report also describes uneven firm adoption and multiple possible channels, including automation, new tasks, and productivity effects; it cannot tell you which channel your workplace will choose. If nothing has changed locally, keep the note and revisit it when you have new evidence. If the task is handled with an approved tool, compare the whole process, including checking, rework, handoffs, and responsibility, then ask who owns review and what has changed. Do not put sensitive material into an unapproved system. The free task checker can organize change-pressure signals and first actions, but it does not estimate your personal probability of displacement. If sustained local changes make a larger decision worth examining, the paid roadmap compares stay-and-redesign, adjacent, and larger-change scenarios against your constraints; it does not guarantee employment or income.

Sources: Artificial Intelligence (AI) exposure categories; OECD Skills Outlook 2026: Skills in the AI age