The U.S. government just gave every job in America an official AI exposure score, straight from actual Claude and Copilot usage data.

Here's how the Bureau of Labor Statistics actually built it.

"BLS combined five external data sources to create a four-category classification of relative AI exposure for every detailed occupation for which projections are produced." That's 831 occupations total -- essentially every job category the government formally tracks for its employment projections. This is part of BLS's 2025-35 Employment Projections release, published August 27, 2026, and it's the first time a federal agency has attached an official AI exposure rating to its full occupational dataset.

Two of the five sources use real usage data rather than expert guesswork or task-description modeling. "Anthropic calculates an observed exposure measure applying Claude usage from Claude.ai conversations and Anthropic Application Programming Interface (API) traffic." Microsoft contributed real Copilot usage data too. That distinction matters: a lot of earlier "AI exposure" research (including well-known academic frameworks) scored jobs by having a model or a panel of experts guess which tasks could theoretically be automated. Building part of this score from actual, observed usage logs -- what people are really asking these tools to help with, in real jobs -- is a meaningfully different kind of evidence.

"'Very high' relative AI exposure generally indicates that, compared to other occupations, a larger fraction of an occupation's tasks can be completed or assisted by AI technology, and that LLMs have been observed performing some of the occupation's tasks." In other words, "very high" isn't a guess about the future -- it requires evidence that language models have actually already been observed doing pieces of that job's work.

"'High' or 'Very high' relative AI exposure does not necessarily mean employment will decline." BLS is explicit that an exposure category is not a forecast, a wage estimate, or a prediction of job loss. A job can be highly exposed to AI assistance and still grow in headcount -- exposure measures how much AI touches the work, not what happens to the workforce as a result.

This is the first time a federal agency has published an official, data-backed exposure ranking for every job it tracks, built in part directly from real AI usage logs rather than surveys or guesswork. Earlier attempts to measure this -- including well-cited academic frameworks built on expert panels rating job tasks against AI capability lists -- relied on judgment calls about what AI theoretically could do. Grounding part of an official government dataset in what people are actually observed doing with Claude and Copilot in real workplaces is a different, and arguably more defensible, kind of evidence. The full occupation-by-occupation data is available as a downloadable file on BLS's site.

Source: https://www.bls.gov/emp/ (U.S. Bureau of Labor Statistics, 2025-35 Employment Projections, AI exposure categories)