How people actually use AI at work

Published: 23 October 2026
Topic: Market
Reading time: 2 min

Anthropic's Economic Index samples millions of real conversations to see what people use AI for. The picture is more collaborative, more concentrated and less reliable than the headlines suggest.

Most claims about AI at work are forecasts. The Anthropic Economic Index tries to measure instead, by classifying large samples of real, privacy-protected conversations with Claude.

Mostly working with it, not handing over

Its January 2026 report, based on a million Claude.ai conversations and a million transcripts from Anthropic's API in November 2025, found that 52 percent of Claude.ai use was augmentation, where a person works with the model, and 45 percent was automation, where the model is handed a task to complete. Augmentation had grown five points since August. Through the API, where businesses build AI into their own products, about three quarters of use was automation.

On Claude.ai, 46 percent of use was for work, 19 percent for coursework and 35 percent personal.

Concentrated on a few jobs

Use is far from evenly spread. The ten most common tasks made up 24 percent of Claude.ai conversations and 32 percent of API traffic, and the single most common task on both was fixing errors in software.

Faster, but not always right

The report estimates that AI speeds up tasks more as they get harder: by about nine times for work needing a school-leaver's education, and about twelve times for degree-level work. It also measures how often tasks succeed: 67 percent of the time on Claude.ai and 49 percent through the API, falling as complexity rises. Adjusting for those failures cuts the report's estimate of AI's potential contribution to annual US productivity growth from 1.8 percentage points to between about 1.0 and 1.2.

What it means for a business

Start where use is already proven: software, writing, analysis and support drafts.
Design for collaboration first. Most value today comes from people working alongside the model.
Plan for failure. If a third of tasks or more go wrong, review and fallbacks are part of the product.
Measure on your own tasks before you automate them end to end.

The useful headline is not that AI will do everything. It is that it already does some things very quickly, and some things wrong, and the work is knowing which is which.