The Anthropic Economic Index

economic indexlabor marketsO*NETClioautomation and augmentationusage data

Handa, Tamkin and colleagues map Claude.ai conversations to the U.S. Department of Labor's O*NET tasks with the privacy-preserving tool Clio, finding usage concentrated in software and writing tasks, about 36% of occupations with AI use in at least a quarter of their tasks, and 57% of conversations augmentative against 43% automative; the post launching the Anthropic Economic Index summarizes the paper and releases its data.

Kunal Handa, Alex Tamkin, Miles McCain et al. (Anthropic) · 2025-02-10 · original

The paper's introduction groups earlier evidence on how AI would affect work into three kinds: predictive models of which jobs are "exposed" to AI, controlled studies of productivity in particular tasks, and periodic surveys of users. In the authors' view none of these can track how AI is actually used across the economy as capabilities change (Section 1). This paper, "Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations" (arXiv:2503.04761, version 1 dated 11 February 2025), proposes a way to measure that directly: classify real conversations on Claude.ai by the occupational tasks they involve. The post of 10 February 2025 that accompanies it launches the Anthropic Economic Index, an initiative to repeat the measurement over time and release the data, and opens the dataset on Hugging Face.

The paper is the methodological base for the later Economic Index reports in this library (see "How it connects"). Section numbers below refer to the arXiv paper unless the post is named.

What the paper does

Tasks rather than jobs

The paper adopts the task-based framework from labor economics (Autor, Levy and Murnane, 2003, and later work cited in Section 2), which treats a job as a bundle of discrete tasks that a person or a machine might perform, such as debugging code or cutting hair. Technologies tend to take over or assist with some tasks within a job rather than whole jobs, so measuring use at the task level can show partial adoption that an occupation-level view would miss.

The task list comes from O*NET, the Occupational Information Network database maintained by the U.S. Department of Labor, which describes occupations through nearly 20,000 task statements, along with skills, wages and preparation levels. O*NET groups tasks into occupations and occupations into broad categories such as "Computer and Mathematical."

Clio and the data

The analysis uses Clio (Clio, Tamkin et al., 2024), Anthropic's system in which Claude reads and summarizes conversations so that researchers see only aggregates. Clio extracts attributes called facets, such as "What task is the assistant performing?" (Appendix B). For privacy and to reduce noise, a task is counted only if it appears in at least 15 conversations from at least five accounts (footnote 5, Appendix B.1).

The data are Claude.ai Free and Pro conversations; business customers (Team, Enterprise and all API use) are excluded (footnote 4). Section 4.1 describes the samples as 7-day snapshots, and Appendix E gives the details: the task, wage and barrier-to-entry, and automation analyses each use 1 million conversations from 16 to 23 December 2024; the skills analysis uses 500,000 conversations from 10 to 17 January 2025; and the model comparison uses 1 million conversations from a longer window, 15 December 2024 to 4 January 2025.

The cluster-based robustness check in Appendix G uses a separate sample of about 2.8 million conversations from 28 November to 18 December 2024. The abstract describes the study as analyzing "over four million Claude.ai conversations."

Mapping a conversation to a task

O*NET's task list is too long to fit in Claude's context window, so the authors classify over a hierarchy (Appendix B.1, Figure 9):

  1. Build a task hierarchy. Each ONET task statement is embedded with the all-mpnet-base-v2 sentence model; the embeddings are grouped by k-means into neighborhoods averaging 40 tasks; Claude proposes higher-level task names for each neighborhood, deduplicates them, assigns each lower task to a parent (with the order of options shuffled), and renames the parents to fit their contents. The result has three levels: 12 top-level tasks, 474 middle-level tasks and 19,530 base-level ONET tasks.
  2. Screen. Claude 3.5 Haiku decides whether a conversation possibly involves an occupational task.
  3. Descend the tree. For relevant conversations, Claude picks the best option at each level until it reaches an O*NET task.
  4. Aggregate to occupations. When a task belongs to several occupations, its conversation count is divided equally among them.

Allowing up to three tasks per conversation gave "qualitatively similar" results, so the paper reports single assignments (Appendix B.1).

Automation and augmentation

Separately, Clio classifies each conversation into one of five collaboration patterns (Section 3.4, Table 1). Two count as automative, where the AI performs the task with little human involvement:

  • Directive: the user delegates the whole task with minimal interaction.
  • Feedback Loop: the AI completes the task with the user mainly relaying feedback from the environment, such as error messages.

Three count as augmentative, where the AI works with the user:

  • Task Iteration: the user and AI refine an output together.
  • Learning: the user seeks understanding or explanation.
  • Validation: the user asks the AI to check their own work.

What it finds

Where usage concentrates

  • By category. Computer and Mathematical occupations account for 37.2% of conversations, and Arts, Design, Entertainment, Sports, and Media for 10.3%, reflecting writing and content generation (Section 3.1). Education, science, and business occupations are also well represented; occupations requiring physical labor are least present. The post gives Farming, Fishing, and Forestry as 0.1% of queries and charts each category against its share of U.S. workers (Figure 3 in the paper); the chart's description in the post gives 3.4% of workers for Computer and Mathematical, against its 37.2% of conversations.
  • Software and writing together. The abstract states that software development and writing tasks "together account for nearly half of all total usage."
  • Skills. Of O*NET's 35 occupational skills, cognitive skills such as Critical Thinking, Reading Comprehension, Programming and Writing are most often exhibited by the model, and physical skills such as Installation, Equipment Maintenance and Repairing least often (Section 3.2, Figure 5). The authors caution that Active Listening, the second most common skill, probably reflects Claude's default habit of rephrasing and asking clarifying questions.

Depth within occupations

The paper asks what fraction of each occupation's tasks shows any AI use (Section 3.1, Figure 4):

Share of an occupation's tasks with AI use Share of occupations
at least 25% about 36%
at least 50% about 11%
at least 75% about 4%

Examples: Foreign Language and Literature Teachers (75% of tasks, including course planning but not grant writing), Marketing Managers (50%), Physical Therapists (25%, research and patient education but not hands-on treatment). The paper reads this as AI being used for specific tasks within occupations rather than across whole jobs. In Section 4.2 it also reports that about 57% of occupations show use for at least 10% of their tasks.

Wage and preparation

Usage peaks in the upper quartile of median wages, with occupations such as Computer Programmers and Web Developers, and is lower at both ends of the wage scale (Section 3.3, Figure 6). ONET's Job Zones* classify occupations by the preparation they need, from 1 (little or none) to 5 (extensive, usually an advanced degree). Table 2 (Appendix D.2) compares each zone's share of conversations with its share of occupations as a representation ratio. Zone 4, typically requiring a bachelor's degree, is the only zone used more than its share: about 54% of conversations against about 36% of occupations, a ratio of 1.50. The ratio is 0.40 for Zone 1 and 0.58 for Zone 5, while Zones 2 and 3 sit closer to their baselines (0.78 and 0.94). The authors suggest that physical work explains low use at both ends, and that regulation and access to sensitive information (for example in healthcare) may also limit use in the most specialized roles (Appendix D.2).

Automation versus augmentation

Across conversations, 57% are augmentative and 43% automative (Section 3.4, Figure 7). The paper's text gives only the two totals; the description of the post's chart (its image alt text) gives the breakdown as Task Iteration 31.3%, Learning 23.3%, Validation 2.8%, Directive 27.8% and Feedback Loop 14.8% (57.4% and 42.6% in total). Directive conversations are mostly writing and content generation; Feedback Loop conversations are mostly coding and debugging; Validation is the smallest category and is concentrated in translation (Section 3.4). The authors note that users may edit outputs outside the chat, so the true augmentative share "may be even higher," and that most occupations show a mix of both patterns.

Model differences

Comparing Claude 3 Opus with Claude 3.5 Sonnet (new), in a sample of about 54% Sonnet and 46% Opus conversations (Appendix B.6), Opus is used relatively more for creative writing and educational content and Sonnet more for coding (Section 3.5, Figure 8). The paper connects the Opus pattern to user observations about its "character and writing style," citing Anthropic's post on Claude's character.

Comparison with forecasts

Section 4.2 sets the results beside earlier predictions. Webb (2019) predicted the highest exposure around the 90th wage percentile, whereas usage here peaks lower. Eloundou et al. (2023) predicted that 80% of U.S. workers could have at least 10% of their tasks affected by language models, against the 57% of occupations observed here; the paper calls this "lower than predicted but potentially trending toward their forecast," notes that predicted healthcare use has not appeared, and that scientific use is higher than expected.

Validation and robustness

  • Human checks (Appendix C), on 150 conversations each, drawn from conversations users had submitted feedback on: the task assignment was judged acceptable for 95.3% at the top level, 91.3% at the middle level and 86% at the O*NET level; the collaboration-pattern label was judged the best one for 90.7%.
  • Dataset composition (Appendix B.7): after screening, non-work conversations were 23% of the dataset and coursework 5 to 10%, depending on the confidence level; many non-work conversations still mapped to occupational tasks, such as personal nutrition planning to dietitian tasks.
  • Accounts versus conversations (Appendix B.2, Figure 10): counting unique accounts instead of conversations changed occupational-category results little.
  • Clusters instead of conversations (Appendix G): assigning Clio's anonymized clusters to tasks reproduced category shares well (minimum correlations 0.95 Pearson, 0.95 Spearman, 0.82 Kendall), occupations reasonably (0.70, 0.58, 0.46) and individual tasks weakly (0.47, 0.22, 0.18). The authors attribute the weak task-level match to overlapping ONET task statements and many-to-many mappings. The caption of Figure 25 notes that only a small fraction (under 20%) of the roughly 20,000 ONET tasks are recovered, and direct assignment gave an average of 4.8 tasks per occupation (Figure 27).

The prompts used

The paper has no principles in the constitutional sense, but its measurements are produced by prompts printed in full in Appendix F, so their wording defines the categories:

  • Screening asks: "Does the conversation possibly involve an occupational task?" with a Yes or No answer (F.1). The word "possibly" sets a permissive threshold.
  • Task mapping tells Claude its job is "to identify which task is performed by the assistant" (F.2), so the unit is what the model does, not who the user is. The paper notes the consequence: a conversation classified under dietitian tasks may come from a non-dietitian (Section 3.1).
  • Collaboration patterns are defined in one line each and Claude is told to return "None" when unsure and to "Use 'None' liberally" (F.3).
  • Skills asks for "ALL skills exhibited by the assistant," with the 35 skill names shuffled (F.4). Only 0.1% of conversations were labeled None (Appendix B.4).

The post adds that Clio lets this run "without human researchers being able to see the original conversations" (post, footnote 1).

Limitations and open questions

The authors list these limitations (Section 4.1 and the post's "Caveats"):

  • Sample. Only Claude.ai Free and Pro conversations over short windows (7-day snapshots, per Section 4.1); API and enterprise use and other providers may differ. Claude outputs only text, so users needing images or video are missing. Because Claude is marketed for coding, coding may be overrepresented, and the post says the authors "don't argue" the data represent AI use generally.
  • Work or not. A writing request may be for a job or for a hobby novel; classification does not show whether the user is a professional.
  • Classifier noise. Claude's reading of tasks may differ from O*NET's intent, and accuracy falls at finer levels.
  • Query complexity. A request for an omelette recipe does not show culinary expertise, so use may be overestimated for some tasks.
  • O*NET itself. The database is static, U.S.-centric, cannot include new tasks AI creates, and contains many near-duplicate task statements.
  • Use of outputs. The paper cannot see whether outputs are copied, edited, checked or discarded, so automation and augmentation are inferred from the conversation alone.

Later Index reports revisited several of these: Insights from Claude 3.7 Sonnet (March 2025) dropped the occupational screen and released a bottom-up taxonomy to capture use outside O*NET, and the September 2025 report says the share of directive conversations rose from 27% to 39% relative to this first sample.

Research questions

  1. How much Claude.ai use falls outside O*NET's task list? The paper notes O*NET cannot include new tasks. Comparing the share of conversations that fit poorly at the base level with a bottom-up clustering of the same conversations would measure the gap.
  2. How stable is the automation and augmentation split under different prompt wordings? The five patterns are defined in a single prompt. Reclassifying the same sample with paraphrased definitions, and measuring agreement, would show how much the 57 to 43 split depends on wording.
  3. Does automation in the chat match what users do with the output? The authors cannot see edits outside the window. A consented study pairing conversations with the resulting documents or code would show how often "directive" use is followed by human revision.
  4. How representative are Claude.ai users of the workforce? The paper compares conversation shares with workforce shares but not user demographics. Linking usage to survey data on users' actual occupations would show how far task classification tracks real jobs.
  5. Why is Job Zone 5 underrepresented? The paper proposes physical work, regulation and data sensitivity. Comparing cognitive-only tasks in Zone 5 occupations (such as literature review by physicians) with similar tasks in Zone 4 would separate these explanations.
  6. Does model choice shift the occupational mix? Section 3.5 finds different task profiles for Opus and Sonnet. Tracking task shares across model launches, as the March 2025 report began to do, would show whether capability changes or user self-selection drive the difference.

How it connects

References

  • Handa, K., Tamkin, A., McCain, M., Huang, S., Durmus, E., Heck, S., Mueller, J., Hong, J., Ritchie, S., Belonax, T., Troy, K. K., Amodei, D., Kaplan, J., Clark, J., and Ganguli, D. (2025). Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations. arXiv:2503.04761v1. https://arxiv.org/abs/2503.04761
  • Anthropic. (2025, February 10). The Anthropic Economic Index. https://www.anthropic.com/news/the-anthropic-economic-index
  • Dataset: https://huggingface.co/datasets/Anthropic/EconomicIndex
  • Works cited above only as the paper describes them (Autor et al., 2003; Webb, 2019; Eloundou et al., 2023; Tamkin et al., 2024) were not read for this note.