Estimating AI productivity gains from Claude conversations (www.anthropic.com)

Anthropic used a privacy-preserving analysis of 100,000 real Claude.ai conversations to estimate task-level time savings and their macroeconomic implications. For each anonymized chat they asked Claude to say how long the task would take with and without AI, mapped tasks to O*NET occupations and BLS wages, and aggregated medians. Claude’s estimates imply tasks averaged ~90 minutes without AI and that Claude speeds up individual tasks by ~80%, with an average implied labor cost of about $55 per task. Savings vary widely by domain (e.g., legal/management tasks ~2 hours without AI, food prep ~30 minutes; healthcare assistance sees ~90% speedups, hardware troubleshooting ~56%), and extreme examples include reducing a 4.5‑hour curriculum task to 11 minutes. Extrapolating these task-level gains using standard productivity accounting yields a headline estimate that current-generation models could raise U.S. labor productivity growth by ~1.8% per year over the next decade—roughly double recent run rates—though Anthropic stresses this is not a forecast because it omits adoption rates, off-chat human work (validation, integration), and future model improvements. They validated Claude’s time estimates against JIRA tasks: Claude Sonnet 4.5 gave meaningful, directionally useful correlations (Spearman ρ≈0.44; log Pearson r≈0.46) that are slightly below developer self‑estimates, and noted a compression bias (overestimating short tasks and underestimating long ones). The method provides a scalable lens for tracking AI’s evolving impact while highlighting occupation-specific bottlenecks and important caveats about unmeasured downstream labor.