Estimating AI productivity gains

Source: Anthropic Research Date: 2025-11-25 URL: https://www.anthropic.com/research/estimating-productivity-gains

Summary

Analysis of 100,000 Claude.ai conversations: Claude estimates time each task would take without AI vs. actual time with AI. Median 84% time savings per task. Management tasks at ~$133 cost, food prep at ~$8. Software development contributes 19% of projected economy-wide gains. Aggregated via Hulten’s theorem to 1.8% annual U.S. labor productivity growth if universally adopted — roughly double recent rates. Note: the January 2026 Economic Primitives report revised this to ~1.0pp after accounting for task success rates.

Implications

This is the economic measurement thread’s headline number — 1.8% annual productivity growth is a significant claim. The 84% median time savings is remarkable but median is the wrong statistic here; the right question is distribution across tasks and occupations. Software development’s 19% contribution is large given its share of total labor. The subsequent Economic Primitives paper’s 1.0pp revision (accounting for success rates) is important context — this paper’s numbers are pre-adjustment. The methodology of using Claude to estimate Claude’s own time savings has an obvious self-report bias worth flagging in any external citation.