Anthropic Uses Real Usage Data to Map AI’s Actual Impact on Jobs

When economists try to predict which jobs AI will disrupt, they typically rely on theoretical assessments—can a language model perform this task? But theory and practice often diverge. This week, Anthropic published research that bridges that gap, introducing a new framework that measures not just what AI could do, but what people are actually using it for in professional settings.

“We’ve developed a measure that combines theoretical LLM capability with real-world usage data, weighting automated rather than augmentative uses more heavily.” — Anthropic Research Team

The Observed Exposure Framework

The research, published in Anthropic’s latest Economic Index report, introduces observed exposure—a metric that tracks which occupational tasks are seeing actual AI automation versus theoretical possibility. The findings reveal a significant gap between what AI can do and what it’s currently being used for.

According to the data, tasks rated as fully feasible for an LLM alone account for 68% of observed Claude usage, while tasks requiring additional tools or software represent another 29%. Only 3% of usage falls into categories theoretically beyond AI’s current capabilities. Yet the coverage remains far from comprehensive—AI is far from reaching its theoretical capability ceiling.

Which Jobs Face the Highest Risk

Computer programmers top the exposure list at 75% coverage, reflecting Claude’s extensive use for coding tasks. Customer service representatives follow closely, with the research noting increasing automation in first-party API traffic. Data entry keyers round out the top three at 67% coverage, as document reading and data entry see significant automation.

The demographic pattern is striking: workers in the most exposed professions tend to be older, female, more educated, and higher-paid than average. This challenges assumptions that AI disruption will primarily affect low-skill, low-wage positions.

“Occupations with higher observed exposure are projected by the BLS to grow less through 2034. The question isn’t whether AI will affect employment—it’s how quickly and through what channels.” — Labor Market Analysis

Early Evidence and Open Questions

The research finds no systematic increase in unemployment for highly exposed workers since late 2022, when ChatGPT launched the current wave of generative AI adoption. However, there is suggestive evidence that hiring of younger workers has slowed in exposed occupations—a potential early signal of how firms are adjusting their workforce strategies.

The framework is designed to be revisited periodically, creating a longitudinal dataset that can identify economic disruption more reliably than post-hoc analyses. By establishing baseline measurements now, before meaningful effects have fully emerged, the research aims to capture the transition as it happens rather than reconstructing it after the fact.

The Productivity Puzzle

One of the paper’s key insights concerns the relationship between automation and augmentation. Tasks performed through API implementations—fully automated workflows—receive full weight in the exposure measure, while augmentative use cases receive half weight. This distinction matters because it helps distinguish between AI as a tool that enhances human productivity versus AI as a replacement for human labor.

The research acknowledges significant uncertainty about how these dynamics will play out. Past attempts to measure job vulnerability have often missed the mark—a prominent study on offshorability identified roughly a quarter of US jobs as vulnerable, yet a decade later most of those positions maintained healthy employment growth.

This article was reported by the ArtificialDaily editorial team. For more information, visit Anthropic Research.