Most debates about AI and employment rely on capability assessments: surveys of what large language models could theoretically handle, conducted by researchers who ask models to complete task descriptions. Anthropic took a different approach. Rather than asking what Claude can do, a research team led by economists Maxim Massenkoff and Peter McCrory looked at what Claude users are actually doing, and published the results in March 2026 in a paper titled "Labor Market Impacts of AI: A New Measure and Early Evidence."
The distinction matters. Theoretical exposure estimates tend to run high, because they count any task a model completes acceptably in a controlled setting. Observed exposure is lower, because it reflects the tasks users bring to the model in practice, how often those tasks succeed, and whether users treat the output as a finished product or a first draft. By grounding the analysis in millions of real Claude conversations, Massenkoff and McCrory produced something rarer in the AI-jobs literature: a measurement of AI's actual footprint in white-collar work today, not a forecast of what might happen in five years.
A New Exposure Measure
The paper's core contribution is what the researchers call "observed exposure." The methodology classifies each task a Claude user brings to the model on two dimensions: whether the model handles it without significant human correction, and whether the output is used directly rather than as a starting point for further human work. A task where Claude drafts a full response that the user sends without editing scores as automated. A task where the user takes Claude's output and rewrites it substantially scores as augmented.
The aggregate finding is striking: 68% of all Claude usage, across the full Claude.ai user base and API traffic, falls into task categories that a large language model can handle entirely on its own. That figure does not mean 68% of jobs are at risk today. It means that when users come to Claude, most of what they bring is within the model's current capability range. Across the API specifically, where tasks tend to be more complex and less conversational, the number stays roughly similar.
Key Findings: Labor Market Impacts of AI
- Claude tasks fully feasible for LLM alone68% of observed usage
- Computer programmers: AI task coverage75% (highest exposure)
- Entry hiring slowdown (ages 22-25, high-exposure roles)~14% since ChatGPT launch
- Current clear unemployment spikeNone detected yet
- Productivity gain estimate (labor)~1.2 ppt/year (conservative)
- Paper authorsMaxim Massenkoff, Peter McCrory (Anthropic)
Who Is Most Exposed
Computer programmers sit at the top of the observed-exposure ranking, with 75% of tasks they typically bring to Claude falling within what a model can handle competently. Customer service representatives are the next most exposed occupational group, followed by financial analysts and roles in office administration. The pattern tracks with where enterprise AI deployments have concentrated: coding assistance, customer-facing automation, and structured analytical tasks.
The exposure figures for legal, management, and creative roles are lower, not because AI is incapable of those tasks in isolation, but because the tasks users actually bring from those fields tend to involve more judgment, more context-dependence, and more verification requirements. Users in high-judgment roles are more likely to treat Claude as an augmentation tool than a replacement, and the data captures that.
The 2026 Agentic Coding Trends Report found a related pattern: 73% of engineering teams now use AI coding tools daily, but developers report being able to fully delegate only 0-20% of tasks. Observed exposure and delegation rates are different metrics, but both point to the same friction: the gap between what a model can do and what users are confident trusting it to do without oversight.
"A great recession for white-collar workers is absolutely possible." Peter McCrory, chief economist, Anthropic, Fortune, April 2026
Where the Signal Is Already Visible
The paper is careful about the unemployment data. As of the research period, there is no clear spike in unemployment rates for workers in the occupations most exposed to AI. The aggregate labor market has not yet registered the kind of dislocation that some forecasts predicted. But the researchers identify one early signal that is harder to dismiss: hiring of workers between 22 and 25 years old into high-exposure roles has slowed by approximately 14% since ChatGPT's public launch in late 2022.
That is not the same as layoffs. It looks more like a quiet narrowing at the entry point: firms are adding fewer junior roles in the functions where AI is most capable, before the economics of those roles have visibly changed. If that pattern continues, its effects will show up in cohort data over the next several years rather than in monthly unemployment numbers now. McCrory told Fortune that a white-collar recession is "absolutely possible" if productivity gains outpace firms' ability to find new work for displaced employees.
The context for reading this data has shifted considerably since the paper was published. Anthropic's IPO filing, submitted to the SEC on June 1, 2026, comes with revenue now running at $47 billion annualized, up from roughly $9 billion at the end of 2025. Essentially all of that revenue comes from Claude being used for exactly the tasks the labor market paper tracks: coding, writing, analysis, and customer interaction. The company's own growth is, in a narrow accounting sense, evidence for its own research findings.
What Firms Should Watch
The paper's methodology offers a more practical planning tool than most AI capability assessments. Observed exposure, by definition, reflects what is already happening in commercial deployments rather than what might happen under ideal conditions. A firm in financial services or software development can look at the 75% exposure rate for programmers and treat it not as a theoretical ceiling but as a reasonable baseline for what a well-deployed AI coding assistant will handle today.
The augmentation vs. automation split is also worth examining. Anthropic found that augmented usage, where Claude contributes to work the user then refines, is more common on consumer Claude.ai (52% of conversations) than pure automation (45%). That ratio is not fixed; as models improve and users gain experience, more augmented tasks tip toward automated ones. The Jevons paradox dynamic that Dario Amodei has discussed publicly suggests the total volume of work may grow even as individual tasks become more AI-intensive, though the distribution of that work across skill levels is the open question firms need to plan around.
For now, the clearest practical implication of the Anthropic labor market data is also the simplest: organizations should not wait for unemployment data to tell them that AI is reshaping their workforce. The reshaping has been visible in Claude's own usage logs for more than a year. The Anthropic Economic Index tracks it in near-real time. The question is not whether the shift is happening. It is whether firms are measuring it.
