Anthropic Labor Research: Distinguishing Theoretical Automatability from Actual Exposure Using Real Claude Usage Data
In March 2026, Anthropic proposed the observed exposure metric, combining task theoretical capability, real Claude usage, work scenarios, and automation degree, and compared it with U.S. occupation and employment data.
Discussions on AI's impact on employment often equate model theoretical capabilities directly with actual substitution. This study uses real usage data to bridge this gap, finding that current actual coverage remains significantly below theoretical capability, and no systematic rise in unemployment rates is observed for high-exposure occupations.
The study combines O*NET tasks, Anthropic Economic Index usage data, and task-level theoretical exposure, assigning higher weights to work-related and more automated usage, then aggregates to the occupation level. Results show weak correlation between exposure and official employment growth forecasts, while providing only suggestive evidence of hiring slowdowns for young, high-exposure occupations.
Business and investment judgments can shift from static job replaceability ratios to real task adoption, automation methods, and diffusion speed; this is closer to operational changes than using benchmarks or expert predictions alone.
When formulating human resource strategies, track real usage and automation depth by task, rather than cutting positions based on model capability rankings; prioritize identifying areas with growing adoption but where organizational processes have not yet adapted.
Data only covers the Claude ecosystem with limited causal identification, requiring continuous cross-validation with other platforms, enterprise deployments, wages, job transitions, and long-term employment data.