Why AI Estimates Miss the Mix Behind a Job Title

What does an occupational estimate say when one title covers different weeks?

An occupational AI estimate describes a standardized pattern attached to a job category; it does not tell you how one person spends a particular week. Three different things can otherwise blur together: scores assigned to separate tasks, a summary calculated for an occupation, and the mix of duties or time that varies among people sharing a title. A role-level score can help compare occupations or focus attention, but it is not a personal work diary.

The distinction matters because methods preserve different details. The U.S. Bureau of Labor Statistics’ “Artificial Intelligence (AI) exposure categories” combines theoretical capability measures with observed-use measures mapped to occupational tasks or activities, then reports relative categories. The International Labour Organization’s “How might generative AI impact different occupations?” describes a 2025 approach that considers both average task exposure and variation across task scores. That added spread can show when an occupation contains unlike task-level signals, but it still does not measure your hours. To decide what an estimate means for you, first inspect its task list and aggregation, then compare them with a representative week before choosing a career response.

Sources: Artificial Intelligence (AI) exposure categories; How might generative AI impact different occupations?

What does one occupation code actually contain?

A standardized occupation profile is a map of work associated with a code, not a record of one employee’s calendar. O*NET’s “Scales, Ratings, and Standardized Scores” explains that tasks can carry ratings for relevance, importance and frequency. Relevance is the share of job incumbents who said a task applied to their job; importance rates how much the descriptor matters to the occupation; frequency reports how often incumbents said the task occurred, using broad categories such as frequently, occasionally and rarely. These are useful clues about the occupation-level pattern, but they are not a statement that every incumbent performs every listed task, nor do they add up to a personal allocation of paid hours. Importance and frequency answer different questions: a task can matter greatly without being done every day, while an often-repeated task need not define the full responsibility of the role. The distinction is visible in O*NET’s current “Accountants and Auditors” profile, code 13-2011.00. Its 30 task statements range from processing invoices and reconciling accounts to preparing audit findings, investigating deficient controls, advising clients, and conferring with company officials about financial and regulatory matters. The O*NET 31.0 “Task Ratings” data dictionary describes these as occupation-specific task and duty ratings, with task records connected to an occupation code and rating scale. The profile’s breadth makes the code informative as a shared reference point; it does not tell us how many hours any particular accountant spends on each item. Imagine two accountants whose duties both fit this code. In one illustrative week, routine invoice processing and reconciliations might recur, with audit review occurring only occasionally. In another, the week might center on reviewing controls, explaining findings and advising decision-makers, while transaction processing is handled elsewhere. These are hypothetical bundles, not observations reported by O*NET. The point is that the same profile can be relevant to both while matching their actual work in different proportions. A title can also fit imperfectly when a role is specialized, senior, shaped by its industry, or organized around a particular employer’s workflow. That is a reasoned limitation of matching an individual job to a broad occupation classification, not evidence about how common any mismatch is. So read the task list beneath the title before reading an occupational estimate as a description of your own work. Mark which statements genuinely recur in your role, which are occasional, and which are absent; keep importance, frequency and time spent separate in your notes. The resulting comparison is a practical check on fit, not a new validated score. If several central duties are missing from the profile, first question whether the occupation code captures your work well. If it does, use the listed tasks to identify which parts of your week merit closer examination.

Sources: Scales, Ratings, and Standardized Scores; O*NET 31.0 Task Ratings; Accountants and Auditors

How does an occupational estimate compress task evidence?

An occupational estimate compresses evidence in stages. A method starts with task statements or broader work activities, assigns a measure to each item, links them to occupation codes, then combines or classifies the results. The final label is shaped by which tasks are included, what the measure means and how aggregation works. A capability score, observed-use measure and occupational category may all be called “exposure,” but they do not describe the same thing. O*NET documentation shows why occupational ratings should not be read as shares of one worker’s week. Its scales distinguish importance, relevance and frequency. Importance rates how important a descriptor is to an occupation; relevance is the proportion of incumbents who say a task applies to their job; frequency reports how often it occurs, with categories such as frequently, occasionally or rarely. None directly measures the paid hours a particular worker spends on the task. A task can be important without occupying most of a week, and frequency does not specify duration. That matters when an AI measure is attached to occupational tasks. Task count, importance, frequency and time share are separate quantities. If a method averages task scores equally, a rarely performed task could contribute as much as one that dominates a worker’s week, unless the method applies weights. Weighting by importance would answer a different question from weighting by frequency or time. “Average exposure” is incomplete unless the reader knows what is averaged, what receives weight and whether the unit is a task, activity or occupation. Methods differ in how they turn item scores into occupational results. The current BLS methodology combines five external sources, three theoretical and two based on observed use, maps results to its occupational classification, converts scores within each source to percentile ranks, calculates separate medians for theoretical and observed evidence, then clusters occupations into four relative categories. BLS explains that percentile conversion compares an occupation with others in the same source. The result is a composite relative grouping, not a count of the worker’s tasks AI can perform or a measure of their time allocation. The process can be summarized as: task or activity definition → capability, applicability or observed-use measure → occupational mapping → weighting, normalization or aggregation → score or category. Each step can discard detail. Mapping may combine activities under a broader code; normalization can replace an absolute value with a relative rank; aggregation can hide whether a result comes from many moderately scored tasks or a few higher-scored ones. An occupational average can conceal whether scores are concentrated or broadly distributed. Mean and dispersion are distinct summaries: the mean locates the center of task scores, while dispersion describes their spread. Neither statistic supplies personal weekly shares. Occupational categories make comparisons across many jobs possible, while a worker’s diary captures a narrower period that may not represent the occupation. The trade-off is scale against individual detail. BLS warns that its categories depend on source coverage, crosswalks, normalization and classification choices, and are not worker replacement estimates. The ILO’s 2025 framework uses task-level scores and a mean-plus-variability approach; the next section examines what that adds. Before comparing headline numbers, look for task definitions, the unit mapped to the occupation, the weighting rule and any normalization. Those details show what a number retains and what it leaves out.

Sources: Scales, Ratings, and Standardized Scores; Artificial Intelligence (AI) exposure categories; How might generative AI impact different occupations?

How does BLS turn unlike evidence into one occupation category?

The Bureau of Labor Statistics (BLS) combines three theoretical measures of occupational AI exposure with two measures based on observed AI use, then groups occupations into four relative categories. Its 2025–35 supplement summarizes how a U.S. occupation compares with others in those inputs. It does not show how one worker divides a week among tasks or estimate that worker’s chance of job loss. “Artificial Intelligence (AI) exposure categories” explains both the calculation and its limits. The inputs begin with different questions. One theoretical source asks whether AI applications relate to abilities required at work; two others assess whether a large language model could reduce time spent on occupation-specific tasks. The observed sources map Claude use and Microsoft Copilot activity to occupational tasks or work activities. BLS notes that theoretical measures also differ from each other: one framework, for instance, distinguishes tasks directly exposed to a language model from those requiring additional software. Observed activity is different again. BLS says it does not directly observe whether workers in a named occupation used AI on the job. These inputs broaden the evidence without making the measures equivalent. BLS maps source results onto the Employment Projections classification, based on the 2018 Standard Occupational Classification. Within each source, it calculates an occupation’s percentile rank among covered occupations. The 0-to-1 scale makes unlike results easier to combine, but ranks only position within a source, not an absolute share of tasks exposed. Some sources do not cover every occupation, so BLS imputes missing ranks after normalization. Its procedure predicts a missing rank from other available source data and the occupation’s two-digit SOC major group. Across 4,155 possible occupation-source combinations (831 occupations times five sources), 3,944 were observed and 211 imputed, affecting 75 occupations. BLS lists source coverage, crosswalks and imputation among factors on which categories depend. The agency then calculates two medians for each occupation: one across the three theoretical source ranks and one across the two observed current-evidence ranks. A clustering algorithm groups occupations into “Low,” “Moderate,” “High” or “Very high” relative exposure. The paired dimensions combine potential capability with mapped real-world activity, but the final category does not retain separate task scores or a worker’s allocation of hours. BLS cautions that the categories are relative and are not forecasts of employment growth or decline, wages or productivity; they are not estimates of adoption probability or worker replacement. They also do not distinguish automation from augmentation. The theoretical sources describe capabilities available no later than mid-2023, and the observed-use sources may lean toward early adopters of particular models. These limits matter when reading a category as an account of every workplace. Use the category to identify broad occupational context, then check the matched occupation code, the underlying evidence dimensions and whether the listed activities resemble the work at hand. It supports comparisons among coded occupations, not conclusions about one worker’s week or an employer’s workflow plans. Use it to decide what to inspect next.

Sources: Artificial Intelligence (AI) exposure categories

What does the ILO's task-variability gradient add?

The ILO's 2025 framework adds a useful question to an occupational average: are the task scores clustered near that average, or spread across very different levels? In its global assessment of 436 detailed occupations in the International Standard Classification of Occupations (ISCO-08), each task received a potential GenAI automation score from 0 to 1. The ILO's “How might generative AI impact different occupations?” plots each occupation by the mean of its task scores and their standard deviation, a measure of how dispersed those scores are. “Generative AI and jobs: A 2025 update” describes the refined assessment as covering nearly 30,000 tasks and combining task-level data with expert input and AI predictions. Consider two hypothetical occupation profiles with a similar average task score. In one, the listed tasks have scores spread widely: some score higher, others lower. In the second, the scores sit closer together around the same average. An average alone could make these profiles look alike. A dispersion measure can distinguish them. The ILO uses the combination of mean and variability to define exposure gradients. Its description characterizes Gradient 2 as moderate occupational exposure with high task variability, a mix of more exposed and less exposed tasks. Gradient 4, by contrast, represents high exposure with low variability: high potential is more consistent across listed tasks. That is the method's added value: it makes one form of within-occupation heterogeneity visible. It does not measure how often a specific worker performs each task, how many paid hours each one takes, or how responsibilities are divided among people sharing a title. Standard deviation here is calculated across scores assigned to task statements within an occupation. It is not a measure of variation between workers, weeks, employers, or task importance. Nor does it say whether a higher-scored task is a central part of a particular person's job. A task list with dispersed scores could coexist with one worker spending most of the week on the lower-scored work, while another role-holder spends more time on higher-scored work; those are logical possibilities, not findings in the ILO classification. The ILO's gradients therefore improve an occupational comparison, not an individual assessment. They can tell a reader that an occupation-level mean may conceal tasks with unlike scores, and point toward the task statements that deserve closer inspection. They cannot supply the worker's personal task mix. The practical reading is to inspect both the average and the spread when the method reports them, then compare the underlying task descriptions with recurring duties in a representative week. Keep the two records distinct: one is an occupational model of potential task exposure; the other is a rough account of what the worker actually does. This distinction makes the estimate more useful without turning dispersion into a personal exposure score or a prediction of job change.

Sources: Generative AI and jobs: A 2025 update; How might generative AI impact different occupations?

Why can a capability estimate and an observed-use estimate disagree?

Capability and observed-use estimates can disagree because they answer different questions. A capability measure asks how assessed AI abilities match an occupation’s tasks. An observed-use measure starts with activity recorded in a product or dataset, then maps it to occupations. One models potential applicability; the other records use in one setting. Neither measures employer adoption or one worker’s weekly duties. The U.S. Bureau of Labor Statistics explains this distinction in “Artificial Intelligence (AI) exposure categories.” Its 2025–35 comparison combines three theoretical measures with two evidence measures. The observed inputs use Claude and Microsoft Copilot activity mapped to occupational tasks or work activities. BLS converts each source to percentile ranks, calculates separate theoretical and observed medians, then groups occupations into four relative categories. This is not a count of workers using AI on the job. Since inputs differ in concept, mapping and coverage, a gap is a reason to inspect evidence, not proof one estimate is wrong. “Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations” shows what observed-use research can establish. The 2025 paper analyzes more than four million Claude.ai conversations mapped to U.S. Department of Labor O*NET tasks and occupations. It reports software development and writing together account for nearly half of usage in the dataset; the paper cautions that it covers one platform and has data and method limitations. This is mapped Claude.ai activity, not a representative census of workers, employers, paid hours or organizational adoption. Platform users, tasks they bring, conversation-to-task classification and the O*NET crosswalk shape what appears. The abstract does not establish the direction or size of each effect. A capability estimate has a different boundary. “The OECD AI exposure measure: Mapping the OECD AI Capability Indicators to occupations,” maps capabilities across nine cognitive, social and physical domains to occupational requirements and constructs an AI Capability Gap index. The OECD describes it as a forward-looking measure for a five-to-ten-year view. Its multidomain scope differs from activity on one language-model platform, including robotics and machine vision as well as language and reasoning. The OECD says actual effects depend on adoption, regulation, organizational change and social choice. Its measure cannot show that a capability works reliably in a particular workplace or that a task has transferred from worker to system. When estimates diverge, compare dates, technology scope, platform or population, task definitions, occupational crosswalks and aggregation. Keep “capability assessed” and “use observed” separately. A high capability signal with little recorded activity may indicate a gap between technical fit and measured use; platform activity does not prove broad adoption or independent task completion. Ask which recurring tasks at work fit the measure and whether an approved workplace workflow uses AI for them. These estimates are clues about different parts of the picture, not personal job-loss odds.

Sources: Artificial Intelligence (AI) exposure categories; Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations; The OECD AI exposure measure: Mapping the OECD AI Capability Indicators to occupations

What can task-score variability tell a worker—and what can it not?

Task-score variability can tell you whether an occupation’s listed tasks receive similar AI exposure scores or a mixture of higher and lower scores. It cannot tell you which tasks fill your week, how much time you spend on them, or whether two people with the same title do the same work. The International Labour Organization’s “How might generative AI impact different occupations?” explains that its 2025 framework compares each occupation’s mean task score with the standard deviation across its tasks. The mean describes the center; the spread describes variation across task scores.

Consider two hypothetical task profiles with a similar mean. In one, scores sit near that average; in another, high and low scores are farther apart. A mean alone can make these profiles look alike, while spread signals that the second occupation combines unlike task assessments. The ILO uses both dimensions to distinguish gradients, including occupations with moderate exposure and high task variability from those with high, more consistent exposure. This helps locate tasks to inspect, but does not show their frequency, hours, or removability. The illustration assigns no values or actual occupations.

Variation across task scores is not variation across workers. The first concerns assessed tasks in an occupation classification; the second concerns how people with the same title divide duties, time, and responsibility. An occupational estimate may expose the first; it does not provide a worker’s diary. Neither mean nor spread tells whether work involves independent judgment, client discussion, exception checking, or approval of consequential decisions.

The O*NET “Accountants and Auditors” profile makes the difference between task content and allocation concrete. It lists invoice processing and reconciling discrepancies alongside examining controls, preparing audit findings, advising clients, and discussing financial or regulatory matters. This standardized profile shows breadth; it is not an AI score or anyone’s weekly allocation. An accountant who mainly processes invoices may have a different recurring task mix from one focused on audit review and client advice, though both may match the same code. These are illustrations, not measured patterns in O*NET.

A role-level spread can serve as a prompt: if scores differ widely, inspect the higher-scoring task statements and ask whether they match recurring duties. If scores cluster, that suggests consistency among the method’s task assessments, not uniformity among workers. O*NET’s “Scales, Ratings, and Standardized Scores” documents occupational ratings, not individual hours or task percentages. Specialization, seniority, industry, employer process, or title wording could also make a standard profile fit a particular position unevenly. That is a limitation to check, not a quantified estimate of mismatch.

For a personal decision, use occupational variability to decide where to look, then use a representative-week inventory to describe your work. Mark recurring tasks, frequency, verification, approval, judgment, and exceptions. Keep the notes qualitative: they are a practical self-review, not a validated exposure score. Task-score spread can reveal a mixed occupational profile that an average would hide; only reviewing your own work can show whether that mix resembles your week.

Sources: How might generative AI impact different occupations?; Accountants and Auditors; Scales, Ratings, and Standardized Scores

Where does occupational exposure stop and workplace change begin?

An exposure result concerns technical potential or measured use under a defined method. Employer adoption, workflow redesign, labor demand and worker displacement require separate evidence. A technically assistable task may still involve trusted inputs, checking, approval, exceptions or accountability. A task score describes a relationship between a technology measure and a task description; it does not show that an organization has changed who does the work, how much work is available, or how staffing decisions will follow.

The International Labour Organization’s research brief, “Workers’ exposure to AI: What indicators tell us – and what they don’t,” says exposure indicators estimate whether AI systems can substitute for people in specific tasks, and vary by method. It identifies static task lists, subjective judgments, and omitted economic or institutional constraints. Exposure does not establish whether firms find automation profitable, how workflows will change, or how employment, wages and demand will adjust. Productivity gains might lower costs and expand demand, so task automation alone does not determine net employment direction. It does not establish that gains will occur locally.

The U.S. Bureau of Labor Statistics draws a similar line in “Artificial Intelligence (AI) exposure categories.” Its 2025–35 classification combines theoretical measures with observed AI-use measures mapped to occupational tasks. BLS says those observed measures do not directly show whether workers in a given occupation used AI at work. It explicitly says the relative categories are not forecasts of employment growth or decline, wages, adoption probability, productivity, or worker replacement; they also do not distinguish automation from augmentation. They offer consistent comparisons across U.S. occupations, but say nothing about a particular employer’s plans.

Observed-use evidence differs. The accessible abstract for “Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations” reports analysis of more than four million conversations mapped through U.S. Department of Labor O*NET tasks and occupations. It describes activity on one platform while noting data and method limitations. Such evidence can show mapped task activity in that service’s data; it cannot establish how common use is across workers, whether employers adopted it, whether tools completed work in production, or whether jobs were removed. Capability and platform-use measures answer distinct questions,; disagreement calls for checking evidence type.

The practical inference is modest but useful: occupational exposure is better for identifying where to ask about process change than for estimating whether a role will disappear. If a decision depends on local conditions, check workplace policy and workflow changes, and consult relevant local labor-demand evidence separately. Use the comparison as context and investigate your own tasks and decisions. Neither a high category nor observed tool activity settles what an employer will do or what will happen to an individual worker.

Sources: Workers’ exposure to AI: What indicators tell us – and what they don’t; Artificial Intelligence (AI) exposure categories; Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations

How should you compare an occupation estimate with your own week?

Start with the occupation code and estimate date, then compare its task statements with work you actually repeat. The code identifies the standardized profile; the date helps show when its evidence was assembled. Neither tells you how your week is divided. Treat this as a practical self-audit, not a validated exposure measure or labor-market forecast. In O*NET’s “Scales, Ratings, and Definitions,” ratings characterize tasks within occupational profiles. They can explain why a task appears there, but are not a record of your hours. The U.S. Bureau of Labor Statistics’ “Artificial Intelligence (AI) exposure categories” methodology also produces occupation-level categories from mapped sources, not an individual work diary. If the profile seems wrong for your role, note that mismatch before interpreting the category. Keep only source tasks that plausibly describe your work. From an ordinary week, list recurring tasks and mark them daily, weekly, occasional, or seasonal. These labels organize attention; they are not measured shares. If the week was unusual because of a deadline, audit, client issue, or leave, repeat the notes for another ordinary week. Seasonal work may need a separate snapshot. For each task, ask whether a tool might draft, search, classify, summarize, or complete a bounded step. Then note the surrounding work: supplying context, handling exceptions, checking outputs, getting approval, and taking responsibility. A recurring, low-consequence digital step may be a place to investigate an approved assistive workflow. An occasional review with serious consequences may require careful verification even if it happens less often. Frequency alone does not settle importance or suitability for automation. Keep notes qualitative. Do not turn frequency labels into percentages, combine them into a personal “AI score,” or assume an occupation category describes your duties. Follow workplace rules on approved systems, confidential information, and review before trying a tool; those rules and the consequences of error shape any sensible trial. If the profile differs from two ordinary weeks, check the code, specialization, and task mapping before considering a skill upgrade or adjacent move. Once the profile fits, the notes can identify recurring work to examine. They cannot establish employer adoption, future demand, or your odds of losing a job.

Sources: Scales, Ratings, and Standardized Scores; Artificial Intelligence (AI) exposure categories

What is a realistic next move when the estimate misses your task mix?

A mismatch between an occupational estimate and your week is a reason to check the mapping before changing careers. Confirm that the occupation code and task statements describe your work. If they do not, identify a closer profile or record the mismatch; do not force your week into a title-level category. The ILO brief “Workers’ exposure to AI: What indicators tell us – and what they don’t” explains that exposure indicators describe technological susceptibility under a method, while leaving adoption and labor-market outcomes unresolved. If a matching task is recurring and bounded, check whether workplace rules permit a small, reviewable assistive step. Keep human checking and responsibility visible. If a core responsibility is shifting, consider a nearby skill that builds on your experience, then weigh learning time and cost against salary, location, health and family constraints. Consider a larger change only after that comparison. A poor occupational match is not evidence that your role is safe or threatened; improve the task description before drawing conclusions. The free AI Proof Work task checker organizes task-level change-pressure signals and first actions; these are not validated personal displacement odds. The optional career roadmap compares stay-and-redesign, adjacent and larger-change paths against your constraints and provides a 30/60/90-day plan, without guaranteeing work or income. Verify the mismatch, then choose one permitted next step.

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

Questions readers ask