This site is a work in progress and has not been widely shared. Content may contain errors. Feedback is welcome.
This site is undergoing review. Some annotations were human-generated, some AI-generated — all are being verified.
Back to papers

Anthropic Economic Index Report: Economic Primitives

Appel, Massenkoff, McCrory, McCain, Heller, Neylon, Tamkin

2026arXiv pre-print
Adoption / usageInterdisciplinary
LLM / Generative AISoftware / codingWriting / contentHuman-AI collaborationAugmentation vs. substitution
View Repository
Summary

Handa, Tamkin, McCain et al. use privacy-preserving analysis of over 4 million Claude.ai conversations mapped to O*NET occupational tasks to document real-world AI usage patterns across the economy, finding concentration in software development and writing with mixed automation (43%) and augmentation (57%) patterns.

Main Finding

AI usage concentrates primarily in software development (37.2% of usage) and writing tasks (10.3%), with ~36% of occupations showing AI usage for at least 25% of their tasks; usage peaks in upper-quartile wage occupations; 57% of interactions show augmentative patterns while 43% demonstrate automation-focused usage.

Primary Datasets

Anthropic Economic Index

AI-focused

Claude usage data (privacy-preserving aggregation via Clio)

Secondary Datasets

O*NET

Other

OES/OEWS

Labor outcomes

O*NET task descriptions; BLS OEWS

Key Methods
Privacy-preserving conversation analysis via Clio system; hierarchical classification of 4+ million conversations to O*NET task categories using Claude; descriptive mapping of AI usage patterns across occupations, tasks, and skills
Sample Period
2024-2025
Geographic Coverage
International
Sample Size
4+ million Claude.ai conversations (1 million for primary task analysis, additional samples for other analyses) across ~20,000 O*NET tasks and 923 occupations
Level of Analysis
Task, Occupation, Individual
Occupation Classification
O*NET-SOC
Industry Classification
None
Replication Package
Partial
Notes
Anthropic Research. Maps real Claude usage to O*NET tasks; finds mostly augmentation (57%) over automation (43%); usage concentrated in software, writing, and analysis tasks. [Claude classification]: This is Anthropic's own research analyzing their Claude.ai platform usage. The paper uses the Clio privacy-preserving system (Tamkin et al. 2024) to analyze conversation patterns. Classification methodology involves hierarchical mapping of conversations to ~20K O*NET tasks using Claude itself as the annotation tool. The paper explicitly acknowledges it cannot observe how users actually integrate Claude outputs into their work, only what they request. Sample excludes Team, Enterprise, and API customers. Human validation of classifications was conducted (Appendix C). The paper provides descriptive statistics but makes no causal claims about AI's effects on productivity or employment.