---
title: "Which Economic Tasks Are Performed with AI? Evidence from Millions of Claude Conversations"
type: source
tags: [ai-adoption, labor-economics, onet, automation-vs-augmentation, anthropic, clio, task-framework]
authors: [Handa, Kunal; Tamkin, Alex; McCain, Miles; Huang, Saffron; Durmus, Esin]
year: 2025
venue: "arXiv:2503.04761 (cs.CY); Anthropic Economic Index"
kind: paper
raw_path: "raw/AI Capabilities & Adoption/Which Economic Tasks are Performed with AI (2025).pdf"
sources: []
key_claims:
  - "First large-scale empirical measurement of actual AI usage across economic tasks; analyses 4M+ Claude.ai Free/Pro conversations via privacy-preserving Clio system."
  - "Maps each conversation to the best-fit O*NET task; uses hierarchical traversal across ~20,000 O*NET task statements."
  - "Software engineering + writing tasks together account for nearly half of total Claude usage."
  - "~36% of US occupations use AI for at least 25% of their tasks; only ~4% of occupations show AI usage for ≥75% of tasks."
  - "Cognitive skills (Reading Comprehension, Writing, Critical Thinking) dominate; physical skills (Installation, Equipment Maintenance) and managerial skills (Negotiation) near zero."
  - "57% of conversations show augmentation (iteration, learning, back-and-forth) vs. 43% automation (direct task fulfilment)."
  - "AI usage peaks in upper-middle wage quartile (software roles) and drops at both wage extremes; similar pattern for barrier-to-entry (peaks at bachelor's-degree roles)."
  - "Framework: task-based economics (Autor-Levy-Murnane 2003; Acemoglu-Restrepo 2018) applied to real usage data rather than forecasted exposure (contra Frey & Osborne 2017; Eloundou et al. 2023)."
  - "Methodology limits: one platform, one-shot classification cannot capture downstream use of AI outputs; O*NET is static and cannot represent newly created tasks."
created: 2026-04-20
updated: 2026-04-20
---

# Which Economic Tasks Are Performed with AI?

## Summary
Handa, Tamkin et al. (Anthropic, February 2025) provide the first large-scale empirical measurement of *actual* AI usage across the US economy, replacing the dominant forecasting-via-exposure approach (Frey & Osborne 2017; Brynjolfsson et al. 2018; Webb 2019; Felten et al. 2023; Eloundou et al. 2023) with privacy-preserving measurement of real user behaviour. Published as the Anthropic Economic Index.

**Method.** Uses Clio (Tamkin et al. 2024) to classify 4M+ Claude.ai Free/Pro conversations from December 2024–January 2025 against the US Department of Labor's O*NET task taxonomy (~20,000 task statements). Hierarchical traversal handles the taxonomy size; aggregation thresholds enforce privacy. Each conversation is also labelled on an augmentation-vs-automation axis.

**Five main findings.**
1. **Where AI is used.** Highest usage in software engineering (software engineers, data scientists, bioinformatics technicians), writing-heavy roles (technical writers, copywriters, archivists), and analytical roles. Physical-manipulation roles (anesthesiologists, construction workers) show minimal usage.
2. **Depth within occupations.** Only ~4% of occupations have AI usage for ≥75% of their tasks; ~36% have usage for ≥25%. AI has begun to diffuse broadly but not yet replaced wholesale task portfolios.
3. **Skill profile.** Cognitive skills (Reading Comprehension, Writing, Critical Thinking) dominate AI conversations; physical skills (Installation, Equipment Maintenance) and managerial skills (Negotiation) barely appear. Mirrors human-complementarity theory.
4. **Wage and barrier-to-entry pattern.** AI usage peaks in upper-middle wage quartile (software roles), drops at both extremes. Similar inverted-U for required education: peaks at bachelor's-degree occupations, lower at minimal or extensive-training extremes. Contrasts with Webb (2019) prediction that peak exposure is in high-wage occupations.
5. **Automation vs. augmentation.** 57% augmentation, 43% automation. Most occupations exhibit a mix of both rather than pure replacement or pure collaboration.

**Limitations.** Single platform (Claude.ai, skewed toward technical users), cannot observe how outputs are used downstream, static O*NET cannot represent newly created tasks, classification is probabilistic.

## Connections
- Core bottom-up evidence for [[concepts/ai-adoption]] alongside enterprise-side [[sources/2025-korst-wharton-gen-ai-enterprise-adoption]] and Microsoft Copilot-side [[sources/2025-tomlinson-working-with-ai]].
- Anthropic-methodology sibling of [[sources/2026-shen-ai-skill-formation]] (both use internal data / Anthropic infrastructure; both co-authored by [[entities/alex-tamkin]]).
- Uses the same Clio system described in Tamkin et al. 2024 *(referenced-not-ingested)*.
- Complements [[sources/2024-xu-the-agent-company-benchmark]] by measuring actual task distribution rather than benchmark coverage.
- Feeds [[concepts/ai-adoption]] with the augmentation/automation distinction that contrasts with METR's developer-slowdown finding in [[sources/2025-becker-metr-ai-developer-productivity]].
- New entities: [[entities/kunal-handa]], [[entities/esin-durmus]].