Claude estimates 1.8% annual productivity growth in the U.S.

Anthropic estimates that widespread use of today’s AI models could raise U.S. labor productivity by 1.8% annually over the next decade. This figure is not a forecast. It is an estimate based on how people use Claude today and on the assumption that the technology reaches the entire economy.

The analysis examines 100,000 real, anonymized conversations from Claude.ai. For each one, Claude estimated how long a professional would take to complete the task without AI assistance and how long it would take with the model’s help.

According to those estimates, the tasks analyzed would require 1.4 hours of human work, or about 84 minutes, and AI would reduce that time by around 80% on average. In economic terms, each conversation represented tasks with an estimated labor cost of approximately $55.

The difference depends on the work

AI does not speed up every task equally. Claude estimates much larger time savings for work involving reading, writing and organizing information than for activities requiring a physical presence or direct interaction.

• Management tasks would take about 2.0 hours without AI.
• Legal tasks would take around 1.8 hours.
• Educational tasks would take 1.7 hours.
• Food preparation tasks would require between 30 minutes and one hour.

In some cases, the difference is especially large. Anthropic cites educational planning tasks that, according to Claude, would take a person 4.5 hours but were completed in 11 minutes with the model. It also estimates savings of 87% when drafting invoices, memos and similar documents, and 80% for financial analysis tasks.

But the effect is more limited in some jobs. Healthcare shows estimated savings of 90% for certain tasks, while hardware problems come in at 56%. Reviewing diagnostic images, for example, would produce savings of just 20%, because professionals can already perform that task quickly without AI.

The calculation for the entire economy

Anthropic cross-referenced these estimates with U.S. labor data, average wages and O*NET, a well-known classification of professional tasks. It then applied a standard economic method to calculate what would happen if the improvements spread across the economy over ten years.

The result: today’s models could roughly double the pace of labor productivity growth recorded since 2019, which has been 1.8% annually. The largest effects would be concentrated in:

• Software development, which would account for 19% of the total estimated improvement.
• Management and operations, at about 6%.
• Market research and marketing, at 5%.
• Customer service, at 4%.
• Secondary education, at 3%.

Restaurants, retail, construction and parts of healthcare show a smaller impact in the data. Not necessarily because AI cannot help in those sectors, but because few tasks associated with them appear in the Claude.ai conversation sample.

An estimate, not a definitive measurement

The study has one central limitation: Claude estimates how much time Claude saves. The model cannot see everything that happens after a conversation, such as reviewing responses, correcting errors, adapting a document or checking that the result works in practice.

Anthropic also acknowledges that its sample does not represent every use of AI. People tend to use Claude for tasks they believe the model can handle, leaving out less common work or tasks that require tools and a physical presence.

To check whether the estimates related to reality, the company compared them with 1,000 programming tasks with recorded completion times. Claude Sonnet 4.5 achieved a correlation of 0.44 with actual times, compared with 0.50 for the developers themselves. The result suggests that its calculations are useful for comparing tasks and identifying trends, but not for precisely measuring every case.

The most important point lies elsewhere: speeding up some tasks can turn others into bottlenecks. If AI prepares a report in minutes, but the worker still needs hours to travel, review data or get approval, those slower steps carry more weight in the final workflow.

That is why the real impact will not depend only on how many minutes a model saves. It will also depend on whether companies reorganize their processes around that capability. Anthropic will continue tracking these changes through its Economic Index, but for now, 1.8% should be read as a scenario based on current usage, not as a promise about the future.