Anthropic Economic Index: What It Measures and Can't Show About Your Job

What does the Anthropic Economic Index measure about AI use?

The Anthropic Economic Index (AEI) measures patterns in activity on Anthropic products by classifying Claude conversations against work tasks. It can show which kinds of tasks appear in the conversations included in a particular release, and how those patterns vary across the sampled activity. Read it as a version-specific indicator of Claude use, not as a forecast of what will happen to your job. It describes product activity through a task lens; it does not begin with an occupational census or an assessment of your own duties.

That distinction sets the limits. Conversation classifications do not establish how common AI use is among all workers, whether employers in your area have adopted it, whether a role has been redesigned, or whether anyone will be displaced. Those are separate questions requiring other evidence. Repeated measurements can still be useful as an early signal of changing activity on the platform; their value depends on keeping the product, period, and unit of observation in view. For a personal decision, an AEI pattern can suggest which recurring tasks to examine. It cannot tell you how much of your own work is changing or what move makes sense under your constraints. The useful question is therefore what this release records and how it was constructed, before treating any chart as a statement about employment.

What the first AEI snapshot actually counted

The initial AEI was a large collection of Claude.ai conversations, not a survey that selected workers and asked how they used AI. In [Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations (version 1)](https://arxiv.org/abs/2503.04761v1), Anthropic researchers describe approximately one million conversations from Claude.ai Free and Pro during December 2024 and January 2025. Anthropic’s launch explainer, [Introducing the Anthropic Economic Index](https://www.anthropic.com/research/the-anthropic-economic-index), gives the same approximate count and date context. The unit counted was conversation activity in those product tiers; a person could generate multiple conversations, so this was not a roster of one million distinct workers.

The researchers used Clio, Anthropic’s privacy-preserving analysis system, to classify conversation activity without exposing underlying conversation text in the published analysis. The pipeline inferred the task a conversation appeared to address, then linked it to the U.S. Department of Labor’s O*NET occupational task taxonomy. Tasks could then be grouped under occupations. Thus, occupational labels were a roll-up from classified product activity through a standard task vocabulary, not job titles reported by a representative sample of employees.

This method makes the first report interpretable as a map of observed and classified Claude activity. The version 1 paper reports that use was concentrated in a subset of tasks and occupations, with computer and mathematical work prominent in its occupational distribution. It also classifies interactions by whether AI use appeared more like automation or augmentation. These findings describe conversations and task assignments in that snapshot; they do not mean equivalent proportions of workers held those jobs, used AI, or had their tasks automated. A conversation assigned to a task indicates that the task appeared in product use, not its share in a worker’s week. The paper’s task and occupation distributions should be read as conditional on that classification path: conversation, inferred task, O*NET task match, then occupational grouping. The resulting chart answers where the analyzed Claude conversations clustered under the researchers’ scheme. It does not estimate how many workers in an occupation encountered the same task, or how much of their work Claude performed.

The version label matters because the unversioned arXiv record now presents a later, much larger analysis. Readers checking the first report should follow the version-pinned v1 paper; the contemporaneous Anthropic launch explainer independently makes the initial vintage and approximate count easy to verify. It corroborates the same producer’s release, rather than supplying a separate dataset. Later reports belong to an evolving series and should not silently replace the vintage behind an initial finding.

Scale and a standardized taxonomy make systematic comparison possible: the same classification frame can reveal concentrations across task categories. Clio also provides a privacy-preserving way to organize large volumes without publishing readable transcripts. These strengths concern processing and organization. The sample still comes from people who used Claude.ai Free or Pro during the specified window, and task assignments still depend on classification. Its denominator remains conversations, not workers or all workplace AI activity. For citation checking, preserve the v1 suffix when linking the first result, and use a later dated report when discussing a later dataset. A fixed version makes both the sample count and the method claim auditable against the right snapshot. That unit distinction is essential when reading occupational rankings.

Sources: Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations (version 1); Introducing the Anthropic Economic Index

Why does the AEI release vintage matter?

Two reports carrying the Anthropic Economic Index name can describe different slices of activity because the measurement design changed. Anthropic’s June 2026 report, “Anthropic Economic Index report: Cadences,” contrasts its continuous daily sampling with earlier seven-day samples. This changes the time resolution: daily slices make it possible to examine variation by hour, while a seven-day collection smooths activity across a longer window. A chart based on one cadence therefore answers a question about its own observation window. It should not be read as if every report sampled the same moments or had equal ability to reveal short-lived fluctuations.

The June report also adds an output classifier. Earlier reporting centered on conversations and the tasks they appeared to address; the newer design classifies outputs as well, so the analytical target now includes what Claude produced. That can distinguish dimensions of activity that a conversation-level view could not separate in the same way. It also means that a change in a category or share may reflect a changed classification target, not simply more or less of an unchanged behavior. When comparing a chart, identify whether its unit is a conversation, an output, or another reported quantity before interpreting its numerator or denominator.

Product boundaries changed too. “Anthropic Economic Index report: Cadences” reports Claude conversations separately from first-party API activity. The product channel matters because users, workflows and volumes entering through a consumer conversation interface need not match activity routed through an organization’s API integration. Combining or separating those channels changes what activity is visible and the denominator against which a share is calculated. A figure limited to Claude conversations cannot stand for all API activity; a figure that adds API data cannot be compared casually with a conversation-only figure. The index label alone does not specify which population of product activity a chart covers.

A practical comparison record should therefore include the report date, collection window and cadence, product boundary, classifier or classification target, and unit of analysis. Then check whether the compared values use the same denominator and whether the report describes a methodological break. If definitions differ, treat the charts as related observations with different scopes, not as points on a seamless series. The June 2026 design can make within-period timing more visible and separate product channels more legible. Those improvements increase temporal and product detail; they do not make platform telemetry representative of workers, employers, or all AI use. Better resolution is not the same as broader coverage.

Method changes can still improve future monitoring. Continuous sampling may reveal rhythms hidden by a weekly aggregate, output classification can add a new view of activity, and separate API reporting can clarify which product channel contributes to a pattern. The gain is informative when readers preserve the vintage alongside the result. For a trend claim, first establish that the reports use comparable windows, channels, classification targets and units; where they do not, describe the change as a difference between report designs and avoid attributing it entirely to a change in behavior.

Sources: Anthropic Economic Index report: Cadences

How does modeled AI exposure differ from observed use?

A task feasibility index and the Anthropic Economic Index measure different things. In “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” the International Labour Organization (ILO) estimates whether generative AI appears capable of performing or changing tasks, using task assessments and a modeling process. AEI, by contrast, records activity observed in Claude products and classifies that activity against work tasks. Here, exposure means a task’s modeled susceptibility to generative AI capability; it does not mean that a worker or employer has adopted a tool, or that a job will disappear. Feasibility and use are complementary evidence axes, not interchangeable scores.

The ILO’s route begins with tasks rather than platform traces. Its methods summary describes a representative sample of tasks from Poland, worker ratings about the potential effects of generative AI, expert assessment, and then model-generated scores mapped to ISCO-08 tasks. This design asks how a broad set of occupational tasks might be affected, including tasks that may not currently appear in a particular product’s activity. It is a structured estimate of technical feasibility informed by task judgments and modeling. Because the sampled task evidence and classification process shape the result, the index is not a direct observation of every workplace or a forecast for an individual worker.

AEI starts from the other direction: activity that occurred in Anthropic’s products. A task can receive a high feasibility assessment in the ILO framework while appearing rarely or not at all in the Claude activity represented in a particular AEI release. That gap could mean capability has not become a common workflow, the relevant users are outside the product’s observed activity, or the task is simply not captured in that sample. It does not by itself show that the feasibility estimate is wrong. Conversely, frequent classified Claude activity shows that users are applying the product to a task; it does not establish that the system can complete the task reliably without human review or that employers broadly use it.

Reading both axes together helps prevent two opposite errors. A feasibility score without observed use should be treated as a signal about technical possibility that merits attention, not evidence of current displacement. Observed use without a matching feasibility result can point to real user experimentation or assistance, but does not prove autonomous performance, dependable quality, or adoption across an occupation. The measures may diverge because one is modeled across task descriptions and the other depends on actual product activity, classification and release coverage. Neither number supplies the missing bridge from capability or usage to a changed allocation of work, staffing decisions, or a particular person’s employment outcome.

The ILO method can reveal a leading possibility before use is widespread; AEI can ground discussion in activity that users actually brought to Claude. Each contributes something the other cannot provide. For a career question, use feasibility to identify tasks worth monitoring and use product activity as evidence about one channel of present use. Then seek evidence about the intervening steps—reliable performance in the relevant workflow, organizational integration, and changed responsibilities—before treating either measure as a reason to make a major move. This interpretation keeps a technical rating from becoming a labor-market prediction and keeps a use trace from being mistaken for proof of broad capability.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure

What does Anthropic's observed-exposure measure add?

Anthropic’s report, “Labor market impacts of AI: A new measure and early evidence,” builds a composite intended to connect modeled task feasibility with observed use and then examine labor outcomes. It combines estimates of which tasks large language models could perform, Claude activity classified to occupational tasks, and a filter for whether that activity appears work-related. The work-context filter matters because a conversation can concern a task without representing productive work in an employment setting. The resulting measure therefore brings together a capability estimate and a platform-specific record of use, rather than treating either one as sufficient on its own.

The report also distinguishes what kind of use appears in the activity. It gives greater weight to use classified as automation than to augmentation, where Claude assists work that remains collaborative. That choice encodes a plausible labor-market hypothesis: replacing task performance could bear differently on labor demand than helping a worker complete or extend it. But it is an analytical weighting decision, not a directly observed count of jobs removed. Classification can be uncertain, and the same task may be automated in one workflow while supported in another. The score’s meaning depends partly on how these categories are defined and weighted.

A further step weights task-level results by the amount of occupational work time associated with each task. A task that is feasible and observed in Claude activity will contribute differently to an occupation-level index depending on its estimated place in that occupation’s task-time bundle. This makes the composite more labor-oriented than conversation shares alone: it connects product activity to tasks and then to the distribution of work within occupations. It still compresses several inputs into a single comparative measure. O*NET task descriptions and time weights cannot capture every employer’s workflow, while Claude activity does not cover all tools or all workers. The result is a constructed exposure indicator, not a direct measure of hours actually automated.

Anthropic then compares more- and less-exposed U.S. worker groups using labor-market data. Its early analysis finds no systematic increase in unemployment among workers in the more exposed groups since late 2022. That is a bounded result about group patterns over the period studied; it does not establish that no occupation, employer, or worker experienced disruption. The report also finds suggestive evidence of slower hiring into highly exposed occupations among workers ages 22–25. “Suggestive” is important: the pattern is not a settled causal estimate, and hiring can change for reasons beyond AI exposure. The evidence raises a question worth monitoring, rather than resolving what caused the observed difference.

The composite adds relevance to labor analysis because it joins feasibility, observed Claude use, work context, use type, and occupation task-time structure. That extra relevance does not make it a validated personal forecast. Its value is comparative and depends on modeling and provider-specific inputs; its outcome analysis is observational and group-level. A reader can use it to ask which tasks may merit closer attention, but neither an occupation’s score nor the early employment findings translate into an individual probability of displacement or a decision rule for leaving a job.

Sources: Labor market impacts of AI: A new measure and early evidence

What does independent workplace-use evidence add?

Worker surveys offer a different view from product telemetry: they ask people about AI use in their working lives, with a specified population, question, and period. In “From Exposure to Adoption: Generative AI in European Workplaces,” Henseke analyzes face-to-face interviews from the 2024 European Working Conditions Survey, covering more than 36,600 workers across 35 countries. The study reports an average generative-AI adoption rate of 12%, with country estimates ranging from under 3% to 25%. These are survey responses within a European workforce study, not a share of Claude conversations or a global estimate. The range also cautions against treating the average as a description of any particular country or workplace.

Henseke reports that adoption is associated with worker and job characteristics, including skills and job content, as well as employee influence, national digitalization, and training. Those relationships describe differences in reported adoption; they do not establish that any one characteristic caused a worker to adopt AI or that adoption changed their employment. The paper also uses a shift-share design to examine technology-related task restructuring. It finds no detectable average effect on worker-reported restructuring in the analysis. That result is bounded by the survey’s outcome, observation period, and design: it does not rule out changes in particular teams, tasks, or employers, and it is not evidence that adoption had no consequences of any kind.

A separate U.S. design appears in “The Rapid Adoption of Generative AI,” by Alexander Bick, Adam Blandin, and David J. Deming. Their nationally representative survey series reports that 27% of employed respondents said they had used generative AI for work in the previous week in late 2024. This weekly-use estimate has employed respondents as its denominator. The paper separately estimates that generative AI assisted between 1% and 7% of work hours. That range refers to hours, not the share of workers who used AI; it should not be read as another adoption rate. The authors also estimate reported time savings equivalent to 1.4% of total work hours, a third quantity with a different denominator and interpretation.

These studies should be compared as complementary measurements, not pooled into a single adoption figure. The European survey records reported adoption in a 2024 multi-country working-conditions interview and examines reported task restructuring; the U.S. series asks employed respondents about recent work use and estimates assisted time. They differ in geography, sampling, question wording, measurement period, and denominator. Neither is a direct replicate of AEI telemetry: the surveys capture workers’ reports across tools, while AEI classifies activity on Anthropic products. Their respective numbers answer different questions.

Together, the surveys establish that workers reported generative-AI use in defined settings, and they show why a measure of use should be paired with evidence about work context and outcomes. Self-reports can capture use outside a single provider’s products, while recall and survey design shape what respondents report. The European shift-share result offers a bounded test of reported task restructuring; the U.S. assisted-hours estimates describe time attributed to assistance. Neither demonstrates employer-wide integration, identifies how a specific role is changing, or establishes individual job displacement. For a career decision, these findings inform the scale and character of reported use; local workflow evidence is still needed to determine what is changing in one’s own work.

The contrast between adoption and restructuring is also analytically useful. A worker may report using a tool without reporting that their task allocation changed, while a task change may arise through several technologies or organizational decisions. Henseke’s study tests an average relationship under its shift-share design; it does not classify every reported use as a workplace intervention. The U.S. estimates similarly quantify reported use and assisted time, not whether assistance was formally integrated into a team’s production process. Keeping those outcome measures distinct prevents an adoption percentage from being mistaken for a measure of organizational transformation.

Sources: From Exposure to Adoption: Generative AI in European Workplaces; The Rapid Adoption of Generative AI

Why can the same job title hide different AI-related changes?

A task map reveals where work actually moves from input to decision, while an occupation-level average can only summarize a common bundle. The U.S. Department of Labor’s “Management Analysts, O*NET OnLine” profile lists information gathering, analysis, documenting recommendations, and consultation about procedures as distinct tasks. Those labels provide a useful starting vocabulary. They do not say how often a particular analyst performs each task or how the steps are arranged in a specific workplace.

Consider an explicitly hypothetical management-analyst assignment: review a set of operational records, identify a bottleneck, write a recommendation, and discuss a procedure change with its owner. Treat this as a way to inspect one’s work, not as evidence that an AI system has completed these activities in practice. At information gathering, ask whether records arrive in a consistent format, whether access is permitted, and how often missing or conflicting entries need human follow-up. A stable batch of well-labeled inputs presents a different kind of opportunity from a stream of partial records whose meaning depends on context.

At analysis, separate the repeatable operation from the judgment around it. A tool might help organize or compare consistent inputs, but the map should record what exceptions occur, how consequential a missed exception would be, and what verification is required before a result can inform a recommendation. Verification has a cost: someone must know the source data, check whether the output follows from it, and investigate anomalies. If that review takes as much attention as doing the initial analysis, assistance may save little time. If checks are narrow and reliable, the analyst may spend more effort interpreting implications instead.

At documentation, ask whether recommendations follow a stable template or depend on explaining trade-offs to a particular audience. Then trace consultation: who can challenge the recommendation, supply missing context, approve a change, and answer for its consequences? The accountable decision owner may remain central even when drafting or comparison becomes easier. Handoffs matter because an output that looks complete at one stage can still require negotiation, revision, or authorization at the next. Mark each handoff and identify who owns the next decision rather than treating a polished document as the end of the workflow.

Two people with the same management-analyst title could therefore face different change opportunities. One may regularly receive standardized data, perform repeatable comparisons, and pass a draft to a clearly identified reviewer. Another may spend more time resolving inconsistent inputs, eliciting requirements, and mediating procedure changes among teams. The same task labels cover both, but frequency, exceptions, verification burden, handoffs, and decision authority differ. The first workflow might offer more scope for tool-assisted throughput; the second may have more value in context gathering and coordination, even if some documentation is assisted. These are conditional interpretations of the map, not measured effects or predictions about either worker.

O*NET’s standardized vocabulary helps locate common tasks; the worker’s actual sequence supplies the detail needed to interpret them. To use the map, take one recurring assignment and record its input, steps, exceptions, checks, handoffs, and accountable decision owner. That concrete sequence can show where a task is bounded enough to examine and where responsibility or local context changes the work. It answers more about an individual workflow than an occupation label alone, while making no claim that every analyst or employer shares the illustration.

Sources: Management Analysts, O*NET OnLine

What can labor-market outcome studies say about one worker?

An employment-outcome model can describe or predict patterns among groups without yielding a probability for one worker or explaining what caused a particular outcome. “AI exposure predicts unemployment risk: A new approach to technology-driven job loss” analyzes historical U.S. unemployment-insurance outcomes organized by occupation and state over 2010–2020. Its units are occupation–state combinations across that period, not named employees followed through a specific AI deployment. The authors use controls and cross-validation to assess predictive relationships in the historical data. That design asks whether group-level information helps predict variation in outcomes; it does not directly estimate an individual’s chance of losing a job today.

The unit of analysis sets a hard boundary on interpretation. A state-level occupation group can contain workers at employers with different products, customer needs, staffing plans, and task mixes. Even if a model predicts that one occupation–state group has a higher unemployment outcome than another, that does not show that every worker in the first group faces the same conditions. A person’s role may include responsibilities that the occupation label averages away; their employer may also be expanding, contracting, reorganizing, or hiring for reasons unrelated to AI. Moving from a group pattern to one person’s forecast would require information and a validated individual-level model that this historical analysis does not provide.

Prediction also differs from causal explanation. A model may identify variables that help anticipate an outcome under its data and assumptions, yet that predictive association does not establish that AI exposure produced the unemployment pattern. Economic cycles, industry composition, regional shocks, organizational choices, and other technological changes can move together with measured exposure. Controls can address specified factors, but they cannot turn a historical predictive design into a controlled experiment or rule out every alternative explanation. Cross-validation tests predictive performance on held-out portions of the data; it does not establish the causal mechanism or guarantee that relationships from 2010–2020 will hold in a later labor market.

The study remains useful at the scale it examines. Occupation-by-state analysis can help detect broad differences worth investigating and can inform questions about where labor outcomes may vary. Its value does not disappear because it cannot answer a personal question. Rather, its result should guide attention toward evidence with a closer connection to the decision at hand. For an individual, that means checking whether a tool is repeatedly integrated into the actual workflow and whether work allocation, review, output expectations, or staffing have changed alongside it. These are practical questions to investigate, not a validated checklist or score.

Local evidence can be assembled from several observable sources. Is there an approved tool and a documented standard for when it should be used? Have tasks or handoffs shifted, and who now reviews the output? Has the expected volume of completed work changed, or has time moved toward checking and exception handling? Are hiring plans or staffing levels changing over a sustained period? Look for a pattern across these signals rather than treating a single announcement, trial, or isolated headcount change as proof. A pilot may not become standard practice; a workload increase may reflect demand; hiring can respond to budgets, seasonality, or a reorganization.

The proportionate conclusion is that group models can surface broad labor-market signals, while a worker’s next decision needs evidence nearer to their own employer and task allocation. Compare observed workflow changes with other plausible causes, and note what remains uncertain. If tools are used only occasionally and responsibilities remain stable, a dramatic inference would outrun the evidence. If integrated use recurs alongside durable changes in handoffs, review, workload, or staffing, the case for examining options more closely becomes stronger—but it still does not become a personal job-loss probability. The historical study can frame the question; local evidence must inform the response.

Sources: AI exposure predicts unemployment risk: A new approach to technology-driven job loss

What is one proportionate next step after reading an AEI result?

Choose one recurring work output and map its path: what information comes in, which steps turn it into a result, where exceptions arise, what must be checked, and who owns the final decision. Then ask the person responsible for the workflow what has actually changed: Is a tool approved and used repeatedly? Have handoffs, review, expected volume, or responsibilities shifted? If the answer is unclear, keep the next move small and gather a concrete example over time. If changes are already reliable and local, compare a bounded upgrade in your current role, an adjacent move, or a larger change. Test each against your salary needs, location, training time and cost, credential requirements, health, and family responsibilities; an appealing path that fails those constraints is not practical. The free task checker can help identify task-level change-pressure signals, but it does not provide a validated probability of displacement. For a broader comparison, the paid career roadmap lays out scenarios and a plan; it cannot guarantee employment or income. Use either as decision support after checking your own workflow, not as a substitute for that evidence. A pilot alone does not prove redesign.

Questions readers ask