Anthropic labor market report shows where AI is taking over jobs in healthcare

Anthropic released a labor market report on March 5 that tracks how often specific job tasks are completed by AI, based on millions of real interactions with its Claude models. Researchers Maxim Massenkoff and Peter McCrory used the data to identify where AI is already handling work and found a widening gap between capability and deployment. Healthcare roles show moderate theoretical exposure to AI, but real-world use remains low, constrained by regulation, integration complexity, and clinical oversight.

The headline finding is that the job market remains largely stable. But the data shows early shifts in administrative and information-heavy roles, including several that are central to how health systems operate. Five findings show where those shifts are beginning to take shape:

Medical record specialists rank among the most AI-exposed jobs

Anthropic’s rankings place medical record specialists at 66.7 percent observed exposure, behind only computer programmers (74.5 percent), customer service representatives (70.1 percent), and data entry keyers (67.1 percent). The leading automated task for medical record specialists is compiling, abstracting, and coding patient data, the kind of structured text processing that current LLMs handle well and that faces fewer clinical liability barriers than direct care roles.

Three of these four occupations are common in health system revenue cycle and health information management departments. The Bureau of Labor Statistics projects customer service representative employment to decline 5 percent from 2024 to 2034. For CIOs overseeing these functions, Anthropic’s data adds a usage-grounded layer to what BLS projections have been signaling.

Clinical roles lag as real-world AI use trails theoretical capability

Computer and Math occupations show 94 percent theoretical capability but just 33 percent observed coverage; for healthcare categories, the observed share is lower still. Massenkoff and McCrory attribute the gap to legal constraints, software integration requirements, and human verification steps.

In healthcare, those barriers are particularly steep: regulatory oversight, EHR interoperability challenges, and clinical verification demands all limit how quickly theoretical capability converts to real-world automation. The report frames the gap as one that will narrow as tools improve, meaning current low exposure in clinical functions likely understates the change ahead.

Exposed workforce demographics align with healthcare administrative roles

Using Current Population Survey data from the months before ChatGPT’s release, Anthropic found that workers in the top quartile of observed exposure are 54.4 percent female (versus 38.8 percent for unexposed workers), earn an average hourly wage of $32.69 (versus $22.23), and hold graduate degrees at nearly four times the rate of their unexposed counterparts. The exposed group is also 11 percentage points more likely to be white and almost twice as likely to be Asian.

This profile inverts the typical automation narrative and maps closely onto health system administrative and informatics workforces, where women and credentialed professionals fill a large share of coding, billing, data analysis, and clinical documentation roles. The populations most exposed are often among the most experienced administrative staff.

Entry-level hiring declines in AI-exposed occupations

The report found no statistically significant increase in unemployment among highly exposed workers. What they did find: among workers aged 22 to 25, the rate of starting new jobs in exposed occupations dropped roughly 14 percent in the post-ChatGPT period compared to 2022. Job-finding rates in unexposed fields held steady at about 2 percent per month while entry into exposed occupations fell by about half a percentage point.

Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen at the Stanford Digital Economy Lab documented a similar pattern using ADP payroll data, reporting a 6–16 percent employment decline for 22-to-25-year-olds in exposed occupations. Both teams attributed the effect to reduced hiring rather than increased separations. The Economic Innovation Group has argued that the Federal Reserve’s 2022–2023 tightening cycle, not AI, better explains the slowdown. For health systems that rely on entry-level pipelines for revenue cycle and HIM roles, the signal is worth monitoring regardless of its cause.

AI use remains far below its theoretical potential across industries

Across every occupational category Anthropic measured, observed coverage sits well inside the boundary of theoretical capability. The Budget Lab at Yale, tracking the same question through different methods, reached a compatible conclusion: the labor market shows stability, not disruption, at the economy-wide level. Massenkoff and McCrory released their coverage data on Hugging Face and intend to update the metric as new Economic Index data arrives.

What Anthropic’s “observed exposure” metric reveals about real-world AI use

The report’s observed exposure metric measures where automation is already occurring, not where it might. Among healthcare roles, that activity is concentrated in medical records, coding, and billing, all administrative functions staffed disproportionately by credentialed women earning above-median wages. Economy-wide disruption remains modest, but entry-level hiring into exposed occupations has already slowed, narrowing the pipeline into the roles where AI usage is highest.