How to Read an AI Exposure Score Alongside Employer Adoption

What kind of signal is this score?

An AI exposure score estimates how much an occupation’s tasks overlap with specified AI capabilities or observed activity on a particular platform. It prompts inspection; it does not forecast outcomes for one worker. Its meaning depends on date, technologies, task descriptions, method, comparison group, and aggregation level. Capability estimates and measures incorporating recorded use answer different questions, even when their labels look similar.

The International Labour Organization’s 2025 global index illustrates a task-first approach. Its working paper combines task information, worker input, expert review, and AI predictions to estimate generative-AI exposure. A representative sample from Poland’s occupational task classification and 1,640 employed respondents produced 52,558 data points about 2,861 tasks, before extension to global estimates. These gradients describe potential exposure, not deployment at a particular employer. The ILO says transformation is more likely than full automation because occupations contain tasks requiring human input; this does not mean every exposed worker will experience the same redesign.

The U.S. Bureau of Labor Statistics takes a different approach for its projections. It combines five sources: three theoretical measures of whether AI could assist or complete occupational work, and two mapping observed Claude or Microsoft Copilot activity to occupational tasks. BLS normalizes results for relative comparison and groups occupations into categories. It cautions that these categories do not directly show whether workers in an occupation use AI on the job, or estimate adoption probability, job loss, productivity, or wage effects. The inputs differ in scope and date, and categories describe occupations as task bundles, not a reader’s own mix.

That mismatch matters. A role label can hide variation: one worker may spend much of the week drafting, another handling exceptions, judgment, or accountable review. An occupational category cannot resolve it. Record the score’s measure, date, technologies, tasks, and comparison method. Map those assumptions to recurring work you perform. Use the score to choose where to look; use workflow evidence to judge near-term relevance. Neither signal gives an individual job-loss probability.

Sources: Artificial Intelligence (AI) exposure categories; Generative AI and Jobs: A Refined Global Index of Occupational Exposure

Why can two credible scores give different readings?

Two credible AI exposure scores can disagree because “exposure” is not one standardized measurement. A score may estimate whether AI could speed up listed tasks, judge whether a system can perform parts of them, connect patents to occupational descriptions, or summarize activity observed on a particular platform. Approaches also use different task inventories, models, occupation classifications, dates, and rules for combining task ratings into an occupation-level result. A score is therefore interpretable only alongside its definition and method. Without those details, a measurement difference can look like a forecast about a worker. The Budget Lab’s February 2026 comparison examined seven occupational exposure measures, crosswalking their occupation codes so the measures could be compared. It found broad agreement overall, with more disagreement as measured exposure rose. The largest gap concerned the magnitude of exposure, rather than whether highly exposed occupations had any exposure at all. Computer programmers illustrate the gap: one measure placed them at the 99th percentile of exposed occupations, another at the 88th. Scores clustered more closely for plumbers, generally near the low end. This compares indices across U.S. occupation categories; it does not show that every programmer’s tasks are equally affected or predict employer adoption. Treat a high score as a prompt to inspect its task basis, not a rank of personal urgency. The source of the underlying observations can also shape a score. A May 2026 RISEI Lab page summarizes a working paper, “Who Uses AI? Platform Selection and the Measurement of Occupational AI Exposure,” which compares three platform-derived measures. The authors argue that platform users differ from the workforce, so conversation-log scores can combine task applicability with user composition. In the paper’s analysis, changing the platform input while holding the empirical design fixed shifted a post-ChatGPT employment coefficient by a factor of 1.9; reweighting estimates to Bureau of Labor Statistics employment shares attenuated them by 42 to 93 percent across specifications. These are reported working-paper results, not settled correction factors. The page identifies it as an arXiv preprint, so findings are provisional and platform-specific. This distinction matters when a score incorporates observed interactions. Such data can reveal what participating users do with a product, but the users are not automatically a representative sample of workers, and interaction is not the same as production deployment or substitution for labor. Conversely, a capability-based score may indicate that tasks are technically applicable even where an employer has not introduced a tool. Neither type alone establishes the reader’s local adoption level. If two credible scores differ, inspect which tasks and observations each includes, how it maps those tasks to occupations, and what its scale means. Give more weight to the task-level explanation than to a single rank. Agreement can support a broad signal to investigate; disagreement is uncertainty about magnitude or coverage, not a reason to average unlike scores or infer a job-loss probability.

Sources: Labor Market AI Exposure: What Do We Know?; Who Uses AI? Platform Selection and the Measurement of Occupational AI Exposure

What counts as employer adoption, and at what level?

Employer adoption is not one event. A business may report any AI use, a department may approve a tool, an employee may use it on a task, or a team may repeatedly run an integrated workflow with defined review and responsibility. These observations answer different questions. A company-wide rate says little by itself about whether a worker’s recurring duties have changed; evidence closer to the function, task, and workflow is more relevant. The Census Bureau’s May 2026 summary of its Business Trends and Outlook Survey shows what a business-level measure can establish. It describes a biweekly, representative survey of U.S. businesses, analyzing responses from December 14, 2025, through May 3, 2026. Businesses were asked whether they had used AI in the past two weeks and expected to use it in the next six months. Results vary by firm size and sector, so a national rate does not describe every employer or team. In November 2025, the core question changed from AI use in producing goods or services to use in any business function. That broader definition complicates comparisons across the change. New questions cover 15 business functions and operational changes such as training and workflow adjustments. These details help separate broad use from implementation signals, but do not identify a worker’s tasks. A ladder has four levels. First, a firm reports any use: a broad signal under a survey definition. Second, a function uses an approved tool: this narrows the setting, but may describe a pilot or occasional access. Third, an individual uses AI on work tasks: closer to lived experience, though informal use may not be permitted, repeated, or part of the team’s process. Fourth, a recurring production workflow gives AI a defined role and specifies who checks outputs, handles exceptions, and remains accountable. This is stronger evidence of integration, but does not establish that jobs or labor demand have fallen. The Bureau of Labor Statistics draws a similar boundary in Artificial Intelligence (AI) exposure categories: its occupational categories combine theoretical measures with observed AI interactions mapped to tasks, but do not directly observe whether workers in an occupation use AI on the job. Read an announcement at its own level; do not treat it as proof about your workflow. Ask where use occurs, who uses it, whether it is approved, how often it recurs, and whether it is a pilot or production work. Then check what changed: which steps moved, what review remains, who owns exceptions, and whether responsibility or workload shifted. Informal bottom-up use is a lead to verify. Until its frequency, permission, and process role are clear, it is not confirmed employer adoption affecting your job.

Sources: AI use by businesses is rising, but remains concentrated in certain sectors and functions; Artificial Intelligence (AI) exposure categories

Why does technical exposure fail to predict actual adoption?

A high exposure score says that AI may be capable of performing or assisting with some tasks in an occupation. It does not show that a particular employer can use it well, affordably, or within its rules. Adoption is a business and workflow choice as well as a technical possibility. A firm compares AI’s value on a task with its costs and the worker’s contribution. That comparison varies across tasks and can change as tools and processes change. The working paper “Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage” makes this distinction explicit. Its framework asks not only whether AI can do a task, but whether using it is worthwhile relative to a particular worker doing that task. The comparison includes AI’s productivity and user costs alongside worker productivity and wages. A system that can produce a draft may still be a poor substitute if staff spend substantial time checking facts, correcting errors, protecting data, or taking responsibility. Technical ability is one input, not the decision itself. The paper tests the gap using the representative German DiWaBe employee survey linked to worker and establishment information. Researchers compare workers’ reported AI use with prominent exposure measures and report a weak relationship. They then construct an occupation-level adoption index from their comparative-advantage framework. In their reported analysis, it accounts for 60 percent of cross-occupation variation in observed adoption, compared with 14 percent for an exposure-only model; the approaches diverge substantially for about 30 percent of workers. These are results from a working paper and a German survey context. They describe explanatory fit in that analysis, not a forecasting guarantee for another country, employer, occupation, or individual. The figures cannot be translated into anyone’s likelihood of keeping or losing a job. This helps explain why a score and a workplace can appear to disagree. Exposure measures often emphasize capability or task overlap. Actual adoption also depends on data access, compatibility with existing systems, reliability for the required standard, the time and expense of integration, and the cost of human review. Policy, accountability, and the consequences of an error can also affect whether an organization permits a use. The International Labour Organization’s research brief “Workers’ exposure to AI: What indicators tell us – and what they don’t” notes that exposure indicators use static task lists and omit economic feasibility and institutional barriers. It says these measures indicate technological susceptibility, not whether firms find automation profitable or how workflows and labor-market outcomes will adjust. For a worker, a high score alongside little local use is an open question about feasibility or organizational fit. It is not proof that the work is safe, because costs, tools, policies, and workflow design can change. It is also not evidence that deployment is imminent. Look for the barrier that might have shifted: Has the tool become reliable enough for the task? Is it approved and connected to the necessary information? Has the organization assigned review and accountability? Has use moved from a demonstration to recurring work? Update your view when conditions change; track whether responsibilities, time, review, or workload move. The score identifies where to inspect; evidence from recurring work tells you whether adoption has reached your tasks.

Sources: Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage; Workers’ exposure to AI: What indicators tell us – and what they don’t

How strong is the evidence that your work is changing?

Evidence becomes more relevant as it gets closer to recurring work. A company-wide announcement shows that leaders are considering or permitting a capability; it does not show that your function uses it. A function-level process may be a pilot. Repeated use on tasks you perform within a defined workflow is stronger evidence, especially alongside a shift in task ownership or review. This is an evidence ladder, not a measure of job-loss likelihood.

In “Large Firms With at Least 20 Employees Biggest AI Users,” the U.S. Census Bureau reports Business Trends and Outlook Survey responses on whether a business used AI in the prior two weeks and expected use in the next six months. It reviews U.S. data collected from December 14, 2025, through May 3, 2026. Census notes that in November 2025 it changed the question from AI use in producing goods or services to use in any business function. The figures describe reported business use, not whether a particular worker’s tasks have changed; the wording change matters when comparing periods.

Consider an illustrative knowledge-work process: a team uses a tool to draft summaries of incoming documents. That establishes neither that summaries are accepted without review nor that anyone’s role has shrunk. Check whether a person verifies claims against sources, handles exceptions, approves the final version, and answers for errors. Note changes across cycles in task ownership, review, time, output, errors, or staffing plans. Faster drafting might coexist with more checking or more documents handled; neither alone demonstrates reduced labor demand.

The Bureau of Labor Statistics says in “Artificial Intelligence (AI) exposure categories” that its occupational categories combine theoretical exposure with observed AI interactions mapped to tasks, but do not directly observe whether workers in an occupation used AI on the job. A score can point to tasks worth checking; it cannot confirm local use. Verify whether use is approved, repeated, and part of production work. A demonstration is weaker evidence than a recurring process with changed responsibilities.

Record the task and frequency, where AI enters, who verifies results, changes in review or ownership, and what remains unknown. Ask: Is the tool approved? Who is accountable for the final output? Decide whether repeated use has altered work you are responsible for, and what observable evidence supports that conclusion.

Sources: AI use by businesses is rising, but remains concentrated in certain sectors and functions; Artificial Intelligence (AI) exposure categories

What does adoption evidence establish about jobs?

Repeated use establishes that a task or workflow is being affected. It does not by itself establish that an employer needs fewer workers, wages will fall, or a particular worker will lose a job. A tool may assist with a task while a person checks the result; automating one step may leave exception handling and accountability with people. A redesigned process can change work without reducing roles. Each claim about employment needs evidence suited to that outcome. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us – and what they don’t” explains why an exposure measure cannot settle the employment question. The brief says indices rely on static lists of current tasks, differ in how they define exposure, and omit constraints such as economic feasibility and institutional barriers. They indicate technological susceptibility under a static view of work. They do not show whether firms find automation profitable, how workflows will change, or how employment, wages, and demand will adjust. The brief describes a possible demand channel: productivity gains may lower costs and expand demand, affecting employment in ways task overlap cannot capture. This is not a promise that new demand will preserve a particular job. It means labor demand cannot be read directly from a capability or use signal. If a team uses a system to prepare routine summaries each week, that is stronger evidence that summarizing has entered its workflow than a general company announcement. It still does not tell you whether output expanded, review work increased, responsibilities moved, vacancies were canceled, or staffing changed. Assistance describes how a person uses a tool. Task automation describes work performed without the same human step. Workflow redesign describes changes to steps or ownership. Output expansion concerns more or different work being produced. Labor demand concerns how much human work an organization or market needs. A staffing reduction is a specific employment outcome. These may be related, but none automatically follows from the one before it. Broader labor-market evidence offers context, not a substitute for local facts. In “What We Do and Don’t Know About How AI is Affecting the Labor Market,” The Budget Lab at Yale compared employment and wages in more-exposed and less-exposed U.S. occupations using synthetic differences-in-differences. The May 2026 analysis reports no statistically or economically significant effects so far under that approach. Its authors explain that the comparison is difficult because exposed and unexposed occupations differ in education, business-cycle sensitivity, and other ways; occupational averages could conceal jobs expanding while others contract. The finding weighs against claiming a clear aggregate effect in that analysis and period. It does not prove that no employer has reduced a role or that no worker has been affected. For a worker, look for outcome evidence alongside repeated use: changes in task ownership, review requirements, workload, output expectations, vacancies, or documented staffing plans. A pilot may justify learning how a process works; a repeated change to core responsibilities can justify reviewing skills and options. Neither is a personal job-loss forecast. Before a consequential move, check the facts against salary needs, location, training time and cost, health, and family responsibilities. Local redesign or staffing decisions can warrant faster planning, but the response depends on what is changing and which options are feasible.

Sources: Workers’ exposure to AI: What indicators tell us – and what they don’t; What We Do and Don’t Know About How AI is Affecting the Labor Market

How should you read the four score-and-adoption combinations?

Read the score as one axis and evidence from your recurring work as the other. “High” and “lower” are relative within a measure, not universal cutoffs; repeated use means more than a demonstration or one-off experiment. This comparison helps decide what to inspect next. It cannot calculate whether a person will lose a job. The Bureau of Labor Statistics’ “AI Exposure Categories” describes occupation-level relative exposure, not a worker’s actual on-the-job use. “Labor Market AI Exposure: What Do We Know?” finds greater disagreement among measures about exposure levels in highly exposed work, so unlike scales should not be treated as interchangeable.

| Score and local evidence | What it supports | What remains unknown | Useful update | |---|---|---|---| | High exposure; little local use | Some tasks merit inspection. | Whether the employer can integrate the tool or will adopt it. | A workflow begins recurring use, or an integration barrier changes. | | High exposure; repeated task use | The work is affected; examine moved steps, verification, and responsibility. | Whether workload, output, hiring, or staffing changes. | Review, ownership, time allocation, or staffing plans change. | | Lower exposure; concentrated use on a consequential task | A broad occupation label can miss a narrow task that matters locally. | The task’s role-wide importance and effects on adjacent steps. | Use recurs, expands, or changes who performs related work. | | Reported use; no workflow change | Someone may use AI, but that alone is weak evidence of a changed job. | Whether use is isolated or part of a process that could alter responsibilities. | Recurring use brings defined review, ownership, or output expectations. |

The counterintuitive case is the third row: a modest occupation-level score can coexist with meaningful change in one recurring task. Aggregation compresses variation among tasks and workers. The International Labour Organization’s “Generative AI and Jobs: A Refined Global Index of Occupational Exposure” treats occupations as task bundles and concludes transformation is more likely than full automation in its framework; that does not settle what happens to one task at one employer. Conversely, a high score prompts inspection; it does not prove deployment is underway.

When the axes disagree, let task-close evidence guide near-term attention and keep the score as a prompt to inspect work not yet touched by current use. Update your view when recurring use comes with an observable workflow change; assess staffing or labor-demand evidence separately. This is more useful than averaging incompatible scores or turning a category into personal odds.

Sources: Labor Market AI Exposure: What Do We Know?; Artificial Intelligence (AI) exposure categories; Workers’ exposure to AI: What indicators tell us – and what they don’t

What should a worker track before changing course?

Track one recurring task across several work cycles before changing course. Describe the task, how often it occurs, and the steps a person currently performs. Add the exposure score’s title or source, the date checked, the technology it covers, and whether it describes theoretical capability, observed platform activity, or both. The U.S. Bureau of Labor Statistics’ “Artificial Intelligence (AI) exposure categories” explains why this context matters: its occupational categories combine several sources, compare occupations relatively, and do not directly observe whether workers in a particular occupation use AI on the job. The category is not an estimate of adoption or worker replacement. For the task, record whether AI use is approved, who uses it, and whether it appears in a recurring production workflow or only in a demonstration, trial, or individual experiment. Note what the tool does, what a person still checks or decides, and who is accountable for the result. Separate the tool’s presence from a change to work: did drafting, searching, sorting, or summarizing move to the system; did review or correction become more demanding; did ownership, turnaround time, output, or workload change? Faster completion of one step does not tell you whether the whole job became smaller, more valuable, or more demanding. A dated note can be brief: task and frequency; current human steps; AI’s role; verification required; observed workflow change; and what remains unknown. Add the period observed and revisit the task after another representative cycle. This improves your local evidence; it is not a controlled study. A few observations can reveal a repeated pattern worth asking about, but cannot establish a labor-market trend or forecast employment. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us – and what they don’t” says exposure measures rely on task lists and omit factors such as economic feasibility and institutional barriers. They indicate technological susceptibility, not how a firm will reorganize work or how employment will adjust. Use the record to choose a proportionate next step. If use is sporadic and responsibilities look stable, keep observing and clarify policy before experimenting. If a tool repeatedly handles a meaningful task, consider a bounded upgrade: learn the approved workflow, check outputs against known requirements, and strengthen the judgment or verification the process still needs. If core responsibilities, workload, or staffing plans are changing, compare adjacent options against your salary floor, location, training time and cost, health, and family needs. Do not paste confidential work into an unapproved service or bypass review rules to gather evidence. The threshold is not a particular number of observations; it is enough repeated, task-close evidence to justify the next reversible action.

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

Which response fits the evidence and your constraints?

Choose the smallest response that matches what you can observe. A high exposure score can justify checking whether a task matters in your work; repeated use on a core responsibility can justify reviewing how that work is changing. Neither requires an immediate career change. If evidence is thin, observe and practice framing tasks or checking outputs against sources. For a task already assisted by an approved tool, a bounded upgrade may be enough. Learn the workflow, practice verification, and note whether it changes time, quality, review, or responsibility. A course can provide structure and feedback, but completing it does not establish readiness for a new role or prove employers will value the credential. A project tied to real work can demonstrate capability, subject to confidentiality and workplace policy. Choose learning that improves a capability you intend to use. Staying and upgrading may be least disruptive when the role still uses valuable domain knowledge and affected work can improve. Yet preserving the same title is not the goal. If recurring changes are moving core duties, compare adjacent roles that use existing experience with a different task mix. A worker who understands a regulated process might investigate quality review or exception handling, then verify local requirements before investing in training. This example does not establish role availability or security. A larger transition merits scrutiny because prerequisites, credentials, location, and time away from paid work can make it costly. Make the comparison against actual constraints. Set a salary floor; identify whether a move requires relocation, unpaid study, schedule changes, or new physical or caregiving demands. Include health, family, and time constraints. If a degree is required for a destination, a short course cannot replace it. If the aim is using AI better in an existing role, a degree may add unnecessary time and expense. Verify prerequisites before paying. The evidence supports this graduated approach. “Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage,” a Columbia working-paper page, reports a weak relationship between worker-reported AI use and prominent exposure measures in a representative German survey analysis; its proposed framework explains more cross-occupation variation in use in that model. This result is not a rule for predicting one employer’s choices. The ILO brief “Workers’ Exposure to AI: What Indicators Tell Us—and What They Don’t” says exposure indicators omit adoption constraints and do not establish labor-market outcomes. Check feasibility and recurring workflow evidence before costly retraining. Use a qualitative threshold: observe when use is tentative or detached from your tasks; upgrade when a recurring task is changing and a feasible improvement is clear; compare adjacent paths when documented changes repeatedly affect core responsibilities. This responds to change without turning a score, tool use, or company announcement into a personal forecast.

Sources: Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage; Workers’ exposure to AI: What indicators tell us – and what they don’t

What is the next useful step?

Pick one recurring task and check the exposure score’s method, scope, and date. Verify whether AI is repeatedly used in that workflow and note one observable change: who drafts, checks, approves, or owns the result. The U.S. Bureau of Labor Statistics explains that its occupational exposure categories do not directly observe workers’ on-the-job use and do not estimate job loss. Treat the score as a prompt to inspect, not a verdict about your position. If the task is not in regular use and responsibilities remain stable, keep observing. If repeated use changes review, ownership, or workload, consider a bounded upgrade or compare adjacent paths if core responsibilities are shifting. Test any commitment against your salary needs, location, available training time, health, and family constraints. The free task checker can organize change-pressure signals; the paid career roadmap can help compare paths when useful. Neither predicts or guarantees an employment outcome. Choose a next step justified by recurring work evidence and feasible in your life.

Sources: Artificial Intelligence (AI) exposure categories

Questions readers ask