Estimating AI Productivity Gains from Claude Conversations
Anthropic's Estimating AI Productivity Gains from Claude Conversations (November 2025; n=100,000 conversations, model-estimated usage data) found that Claude estimated AI cut task completion time by about 80%, implying a 1.8% annual boost to US labor productivity growth if applied economy-wide.
Key findings
- 01Claude estimated that AI cut task completion time by about 80%; the median conversation showed an estimated 84% time savings.Estimates were generated by Claude for each conversation, then mapped to O*NET tasks.
- 02Tasks brought to Claude would take about 1.4 hours on average without AI, worth an estimated $55 in human labor.Costs use O*NET occupations and BLS wage data.
- 03Savings varied by task: about 90% for healthcare assistance tasks versus 56% for hardware issues, and 20% for checking diagnostic images.Where AI helps less, those tasks may become bottlenecks.
- 04Extrapolated economy-wide, the estimates imply a 1.8% annual increase in US labor productivity growth over the next decade.Anthropic says this is not a prediction, since it ignores adoption rates and future model improvements.
- 05Software developers account for 19% of the implied productivity gain, followed by general and operations managers (about 6%) and marketing specialists (5%).This reflects how often each occupation's tasks appear in Claude usage.
By the numbers
84%median estimated time savings per conversation
1.4 hrsaverage task length without AI
1.8%implied annual US labor productivity growth boost
What it means for you Draft
For executives at $10–100M companies
These are the AI model's own estimates of time saved inside a chat, not measured business results, and they leave out time spent checking the output. For a $10–100M company, they are a useful map of which kinds of tasks see the largest speedups, but not a basis for headcount or budget math.
For practitioners
Use the task-level variation as a starting list for pilots, then measure end-to-end time including review. The authors note estimates run long for short tasks and short for long ones.
Limitations
Medium trust.Large dataset and open method with validation, but estimates are model-generated and the publisher sells the AI being measured.
Time savings are estimated by Claude, not observed; validation on software tasks showed moderate correlation with actual times (Spearman 0.44 vs. 0.50 for human developers). Excludes time spent outside the conversation, such as verifying work, and covers only Claude.ai consumer-tier users.
About the publisher: Makes Claude; studies its own users.
Anthropic. "Estimating AI Productivity Gains from Claude Conversations." November 25, 2025. https://www.anthropic.com/research/estimating-productivity-gains
Related reports
Technology Report 2026Leaders expect a 95% developer productivity uplift in 1–2 years; current gains are 20–27%. Bain says AI moves bottlenecks into review and coordination.
2026 GenAI Code Security ReportAI models now write syntactically correct code nearly always, but about 44% of generated code still carried a known vulnerability.
State of AI Impact in Engineering: Q2 2026 ReportAI now writes 52.7% of code and AI spend rose ~28x in tech, but time on new features stayed flat and change confidence fell 6.1%.