If your job involves a screen, a keyboard, and judgment calls, this is the first AI labor market data set created from actual production use rather than theoretical benchmarks. Pooya Golchian breaks down what it’s really saying, from the precise exposure level of your occupation to what employment and hiring data through 2025 reveals, and how to adapt based on evidence rather than headlines.
Three years after ChatGPT launched, Anthropic economists Maxim Massenkoff and Peter McCrory published an analysis based on actual usage data from Claude, not surveys or capacity benchmarks. Its core concept, observed exposure, measures what AI is actually doing in professional workflows right now. The gap between that measure and each previous estimate is the study’s most important finding.
Key findings at a glance:
The following table directly reproduces Figure 2 of the article. The outer ring is the theoretical AI capability. The inner circle is what Claude is actually doing in the same 22 occupational categories. Read it to get the big picture before section-by-section analysis.
Why All Previous AI Risk Figures Were Probably Wrong
The dominant framework for measuring the threat of AI to jobs was built by Eloundou et al. in 2023. Their method asked whether a large language model (LLM) could reduce the time needed for a given task by at least 50%. Applied to every task in every occupation in O*NET (the Occupational Information Network, a US government database of standardized descriptions of job tasks), it produces alarming numbers. Computer and Mathematics occupations reached 94% exposure. Office and Admin reached 90%.
Thousands of articles have cited these figures. They are also measuring the wrong thing.
They measure capacity, that is, what models can do under ideal conditions with careful indications. They do not measure implementation, that is, what AI actually does in today’s professional workflows.
Anthropic researchers constructed a different measure using the Anthropic Economic Index, a data set of actual Claude API (application programming interface) calls and product usage. They filtered out professional and work-related contexts, separated automated channels from human-assisted augmentation (giving greater weight to the former), and mapped each usage pattern to O*NET task descriptions.
The result is observed exposure: a true measure of how deeply AI has penetrated real work, right now.
The gap between theory and reality
The gap between theoretical and observed exposure is consistently 50 to 65 percentage points. Computer and Mathematics occupations showed 94% theoretical exposure versus 33% observed. Office and Admin showed a theoretical 90% against an observed 25%. The implementation gap spans between 50 and 65 percentage points in all major categories.
This gap exists for several real reasons:
– Integration cost. Connecting AI to existing systems, data pipelines, and approval workflows requires significant engineering work.
– Responsibility and compliance. Legal, financial, and healthcare tasks often require human approval, regardless of AI capabilities.
– Trust and verification. Workers and managers require consistent, auditable results before delegating important decisions.
– Workflow inertia. Many organizations have not yet restructured their functions around the rise of AI.
The gap doesn’t mean AI will never close it. It means that the gap exists today and that theoretical forecasts dramatically exaggerate the current disruption.
Theoretical versus observed exposure by occupation category
Statistical analysis: what the full 22-category data set shows
The radar chart above covers all 22 occupation categories in the document. Running the entire data set through descriptive and inferential statistics yields findings that go beyond any individual data point.
Key statistics in the 22 C