Anthropic Economic Index launch and first findings

TL;DR

Anthropic launched the Anthropic Economic Index, publishing an initial report that quantifies how AI is used across U.S. occupations using anonymized Claude.ai conversation data, and open‑sourced the dataset for the research community.

What the Index measures

The Index analyzes real‑world AI usage rather than surveys or forecasts. It maps roughly one million Claude.ai conversations to the U.S. Department of Labor’s O*NET task taxonomy using Anthropic’s Clio system, then aggregates tasks into occupations and six high‑level occupational categories.

"Clio translates conversations with Claude (which are kept strictly private) into occupational tasks and occupations derived from O*NET…" – Anthropic blog

Key quantitative findings

AI adoption by occupation

Computer & Mathematical jobs account for 37.2% of Claude queries, despite representing only 3.4% of the U.S. workforce.
Arts, Design, Sports, Entertainment & Media represent 10.3% of queries, far exceeding their 1.4% share of workers.
Physical‑labor categories such as Farming, Fishing & Forestry appear in only 0.1% of queries.

Depth of AI use within jobs

Only ~4% of occupations have AI involved in ≥75% of their tasks.
Approximately 36% of occupations see AI in at least 25% of tasks, indicating moderate but widespread diffusion.

Wage correlation

Mid‑to‑high‑wage occupations (e.g., programmers, data scientists, copywriters) show the highest AI usage rates.
Both low‑pay (e.g., shampooers) and very‑high‑pay (e.g., obstetricians) jobs exhibit minimal AI interaction.

Augmentation vs. automation

57% of AI‑assisted tasks are augmented (AI collaborates with the user).
43% are automated (AI performs the task directly).
Sub‑breakdown: augmentation consists of Validation (2.8%), Task Iteration (31.3%), and Learning (23.3%); automation consists of Directive (27.8%) and Feedback Loop (14.8%).

Methodology at a glance

Conversation collection – Claude.ai Free and Pro plan conversations were harvested, with user privacy preserved.
Task matching – Clio matched each conversation to the most relevant O*NET task.
Occupational aggregation – Tasks were grouped into occupations, then into six broad categories (Computer & Mathematical, Arts & Media, Education & Library, Office & Administrative, Life Sciences, Business & Financial).
Supplementary data – Median U.S. wages from O*NET were merged to examine salary‑usage relationships.
Validation – The paper’s Appendix B details validation checks for Clio’s classification accuracy.

Limitations acknowledged by Anthropic

The dataset cannot confirm whether a conversation was work‑related; hobby or personal use may be mixed in.
The analysis does not capture how users ultimately employed Claude’s output (e.g., copy‑pasting code vs. manual editing).
Only Free and Pro plan users are represented; API, Team, and Enterprise users are excluded.
Classification errors are possible; see the paper’s Appendix B for error rates.
Claude’s lack of native image generation may under‑represent creative AI use.
Claude’s strong coding capabilities could bias the sample toward software‑related tasks.

Implications for policy and research

Occupational impact – The concentration of AI use in mid‑wage, knowledge‑intensive jobs suggests that AI is currently augmenting rather than fully automating work, implying a shift in task composition rather than wholesale job loss.
Augmentation dominance – A slight tilt toward augmentation (57%) indicates that AI is more often a collaborative partner, which may inform workforce training programs focused on human‑AI teamwork.
Data‑driven policy – Open access to the dataset enables economists and policymakers to test hypotheses about productivity, wage dynamics, and sector‑specific disruption.
Longitudinal tracking – Anthropic plans to repeat the analysis over time, allowing researchers to monitor trends such as increasing automation ratios or deeper AI penetration within occupations.

Open resources and community engagement

Dataset – Fully downloadable at https://huggingface.co/datasets/Anthropic/EconomicIndex.
Paper – The initial findings are detailed in the arXiv preprint https://arxiv.org/abs/2503.04761.
Feedback form – Researchers can submit comments and research ideas via https://docs.google.com/forms/d/e/1FAIpQLSfDEdY‑mT5lcXPaDSv‑0Ci1rSXGlbIJierxkUbNB7_07‑kddw/viewform?usp=dialog.

Future directions

Anthropic will continue to:

Update the Index with newer Claude conversation data.
Expand coverage to additional user tiers (API, Enterprise) where possible.
Refine the augmentation vs. automation taxonomy as model capabilities evolve.
Publish longitudinal reports that track changes in AI depth of use, wage‑task correlations, and sectoral adoption patterns.

Acknowledgements: The authors thank Jonathon Hazell, Anders Humlum, Molly Kinder, Anton Korinek, Benjamin Krause, Michael Kremer, John List, Ethan Mollick, Lilach Mollick, Arjun Ramani, Will Rinehart, Robert Seamans, Michael Webb, and Chenzi Xu for early feedback.