On Feb 10, 2025, Anthropic announced the launch of the Anthropic Economic Index, a project designed to track how artificial intelligence affects labor markets and the economy over time. The Index’s inaugural paper presents novel, anonymized data and analysis based on conversations on Claude.ai, drawn from Free and Pro user interactions.
Data and methodology
The study analyzes roughly one million Claude.ai conversations. Anthropic used an automated tool called Clio to aggregate and classify conversations while preserving user privacy so that human researchers cannot view original chats. Each conversation was mapped to occupational tasks from the U.S. Department of Labor’s O*NET database (about 20,000 specific tasks), then grouped into occupations and broader occupational categories (for example, education and library; business and financial; and others).
Anthropic is open-sourcing the dataset used for the analyses and is inviting economists, policy experts, and other researchers to provide input and build on the findings.
Key findings
- AI usage is concentrated in computer and mathematical tasks: 37.2% of the analyzed queries to Claude fell into this category, covering activities such as software modification, code debugging, and network troubleshooting.
- The second-largest category was arts, design, sports, entertainment, and media, accounting for 10.3% of queries, largely reflecting writing and editing use cases.
- Physically intensive occupations—such as farming, fishing, and forestry—were minimally represented (0.1% of queries).
Depth and spread of use:
- Very few occupations saw AI used across most of their tasks: roughly 4% of jobs used AI for at least 75% of associated tasks.
- More moderate adoption was broader: about 36% of jobs used AI for at least 25% of their tasks.
Wages and AI use:
- By combining O*NET median U.S. salary data with AI-use rates, the study finds that both low-paid and very high-paid occupations show low AI usage. Higher AI use is concentrated in mid-to-high median wage occupations such as computer programmers and copywriters.
Automation versus augmentation:
- The analysis distinguishes between automation (AI directly performing tasks) and augmentation (AI collaborating with or enhancing human performance). Overall, 57% of tasks were classified as augmentation and 43% as automation. Augmentation cases included activities like validation, learning support, and iterative collaboration (e.g., brainstorming and refinement).
Limitations and caveats
Anthropic lists several important limitations to the study:
- The researchers cannot be certain that each Claude conversation that matched a work-related task was actually performed for work; some conversations may be personal or hobby-related.
- The data do not reveal how users employed Claude’s outputs—for instance, whether code snippets were copy-pasted into production or used only as references—so some instances labeled as automation could actually involve subsequent human editing (i.e., augmentation).
- The dataset covers only Claude.ai Free and Pro conversations, not API, Team, or Enterprise usage.
- Clio’s automated classification may mislabel some conversations; Anthropic documents validation steps in Appendix B of the paper.
- Claude cannot generate images directly (except via code), so some creative uses are not captured.
- Because Claude is positioned for coding tasks, code-related usage may be overrepresented in this dataset; Anthropic does not claim the sample is fully representative of AI usage overall.
Next steps and openness
Anthropic emphasizes that AI capabilities and labor-market dynamics can change rapidly, so it plans to repeat many analyses over time and regularly publish results and datasets as part of the Anthropic Economic Index. Longitudinal tracking will help monitor changes in the depth of AI use inside occupations and shifts in the balance between automation and augmentation.
The full dataset used in the paper has been released, and Anthropic provides a form for researchers to give feedback and propose further research directions.
Acknowledgements and job opportunities
Anthropic thanks contributors who commented on drafts and early findings, including 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.
Anthropic also advertises roles for researchers interested in studying AI’s labor-market effects, including Societal Impacts Research Scientist and Research Engineer positions.
Summary
The first Anthropic Economic Index report provides a detailed, usage-based snapshot of how AI is currently integrated into workplace tasks: usage is concentrated in software and writing-related tasks, augmentation slightly outweighs automation (57% vs 43%), and a minority of occupations use AI across most of their tasks (~4%) while a larger share (~36%) uses AI for a substantial minority of tasks. The data and methods are publicly shared to support further research and policy discussion.
