Data study · 23 September 2026 · 5 min read

How much time does AI save? By job

Anthropic data on 20 occupations: a programming task drops from 5 hours to 9 minutes, a nursing task only about 4 times faster. What the gap means for jobs.

Across 700 occupations in Anthropic's usage data, the typical task people bring to Claude would take a person about 44 minutes on their own and about 5 minutes with AI. That is roughly a tenth of the time.

The saving is not spread evenly. It is largest in programming, where a task goes from five hours to nine minutes, and smallest in nursing and teaching, where it is about four times faster rather than thirty.

Time per task, by occupation

Source: Anthropic Economic Index, June 2026 data release, API traffic, global, fields human_only_time_mean and human_with_ai_time_mean by occupation. Times are rounded.

How to read this table

These are estimates, from Anthropic's own analysis of the conversations, of how long the tasks people actually sent to Claude would take. They are not stopwatch measurements, and they describe the tasks people chose to hand over, not the whole job. A paralegal who uses AI to summarise a deposition still spends most of the week on work that never goes near a model.

They also come from API traffic, meaning developers and businesses calling the model directly rather than people chatting in the app. That skews toward the work that automates well.

What stands out

Programming tasks see the largest compression because a large share of them are long, well specified and checkable. Two occupations near the bottom, nursing and teaching, are ones where the task a person brings to AI is small next to a working day spent with patients or children.

Data entry and customer service are near the bottom too, which looks odd until you notice why: their individual tasks were already short. Saving a few minutes on a task that takes twenty is a smaller ratio, but at the volume those jobs run, it is the saving that shows up first in headcount. That is consistent with where 2026's AI related job cuts have been concentrated.

Faster tasks and job risk are different things

A job where tasks get 30 times faster can still grow if demand for the work rises. A job where tasks get 7 times faster can still shrink if it is all the same short task, repeated. Speed tells you where AI is doing the most work per request. It does not tell you what happens to the people doing that work.

That is why our index scores occupations from published exposure research rather than from a single usage signal. It covers 961 occupations, and each page lists the studies behind its score. For the latest example of how fast long tasks are compressing, see Claude Opus 5.5's 20 hour audit done in 3.

How much of your week looks like the top of this table is the question the assessment is built to answer. It takes about two minutes and needs no account.

Figures come from the AI Job Risk Index and were current when this was published. Scores change as new research is ingested, so the index is always the live version. See how scoring works. Informational guidance based on published research, not professional career or financial advice.