TL;DR

A new economics paper finds generative AI has spread across 80% of occupations and over 40% of tasks, yet in most of those occupations fewer than half of workers actually use it. Only around one occupation in six passes 70% adoption. No task at all exceeds 70%.

Why this contradicts the vendor numbers

The interesting fight here is methodological. Anthropic, Microsoft and OpenAI have each estimated occupational exposure by classifying their own chat logs. The authors — Alexander Bick of the St Louis Fed, Adam Blandin and Tyler Schumacher of Vanderbilt, and Harvard’s David Deming — instead asked workers, using Real-Time Population Survey data.

The results diverge, and the paper explains why. Chat-log classifiers map conversations onto broad, activity-shaped task descriptions such as editing written material. OpenAI’s data attributes around 15% of chats to that kind of editing. But in the US Labor Department’s O*NET occupational database, only 2.4% of workers hold jobs that include the task. The classifier is matching text to a label, not to a job. Exposure figures built that way, the researchers conclude, overstate how relevant the technology is to real work.

What the distribution looks like

Adoption clusters where you would expect: management and professional roles, especially in finance, business and computing. It is thinnest in personal services and work requiring physical or interpersonal presence. Four in five detailed occupations clear 20% adoption, but only about 15% — largely computer-oriented — get past 70%. Among tasks, just 2.8% exceed 50%.

As of May 2026, 45% of US adults aged 18 to 64 used generative AI for work and 55% used it outside work, with 62% using it at all. The strongest predictor of adoption the authors identify is prior experience: people who start in one domain tend to carry it into others.

Looking forward

This lands alongside a Gartner finding that only 22% of organisations have scaled AI across multiple business units — different method, same shape of answer. For UK firms, the practical reading is that headline exposure statistics describe reach, not depth, and the two are being conflated in a lot of business cases. The authors’ closing argument is the useful one: understanding why some workers adopt and others do not now matters at least as much as cataloguing which tasks the technology could theoretically do.