Anthropic map: AI can replace many white collar jobs, but real use lags capability
Mar 7th 2026
Anthropic’s new study maps what AI can do versus what it is actually doing and finds a large gap, with the highest risk falling on well paid, highly educated white collar workers as adoption grows.
- Anthropic introduced a new metric called observed exposure that compares what AI can do to how it is actually used in the workplace.
- AI models are theoretically capable of handling most tasks in business, finance, management, computer science, math, legal, and office administration.
- Actual adoption is far lower than capability, for example LLMs could cover 94% of computer and math tasks while Claude is observed doing 33% of those tasks.
- Workers most exposed are higher paid, more likely female, and nearly four times more likely to hold a graduate degree than the least exposed group.
- About 30% of workers in roles that require physical presence, like cooks and mechanics, have zero AI exposure.
- Researchers warn that if adoption accelerates the US could see a sharp white collar employment shock similar to a Great Recession for those sectors, though that outcome has not happened yet.
- Early signs include a 14% drop in job finding rates for young workers in exposed fields after ChatGPT, but there is no systematic rise in overall unemployment yet.