Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations
Citation. 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., & Ganguli, D. (2025). Which economic tasks are performed with AI? Evidence from millions of Claude conversations. arXiv. https://doi.org/10.48550/arXiv.2503.04761
Summary. Handa et al. analyze more than four million Claude.ai conversations using a privacy-preserving system that maps conversations to O*NET tasks and occupations. The paper finds concentrated use in software development and writing, broader diffusion across many occupations, and a near-even split between augmentation-like and automation-like uses. The authors frame the method as an empirical measurement system for tracking actual AI use, while noting that the data come from one AI platform and cannot directly establish labor-market outcomes.
Key claims
- AI use was concentrated in software development and writing tasks, which together accounted for nearly half of observed usage in the Claude.ai conversation sample.
- AI use was nevertheless distributed across the occupational task space: about 36% of occupations showed AI use for at least a quarter of their associated tasks, while about 4% showed use across at least three-quarters of tasks.
- The authors classified 57% of observed usage as augmentation of human capabilities, such as learning or iterating on output, and 43% as automation, such as fulfilling a request with minimal human involvement.
- The paper treats real-world AI conversation data as a complement to exposure indices, surveys, and controlled productivity studies because it measures observed use rather than only predicted capability or self-reported adoption.
Evidence & limitations
- Evidence comes from more than four million Claude.ai conversations mapped to O*NET tasks and occupations through a privacy-preserving analysis pipeline.
- The paper is an arXiv preprint from Anthropic authors; it is strong descriptive evidence about use on one platform, not causal evidence about productivity, wages, employment, or durable organizational redesign.
- Platform, user-selection, and task-mapping limits matter: Claude.ai usage may not represent all AI tools, all workers, or all organizational settings, and mapping conversations to occupational tasks adds measurement uncertainty.
- The augmentation/automation distinction is useful but coarse; a conversation that looks automated in the transcript may still sit inside a larger human-governed workflow.