Healthcare's AI reckoning starts in your back office.

When executives talk about AI in health care, they usually mean the clinic: ambient scribes, diagnostic models, the radiologist and the algorithm. The first large economic effect is elsewhere. Anthropic's Economic Index tracks what millions of people and businesses actually do with Claude, mapped to the U.S. government's O*NET task catalogue, and it points to the administrative and knowledge workforce that turns care into claims, codes, authorizations, payments and plans.

The exposure is in the paperwork

Anthropic's labor-market study measures observed exposure, the share of an occupation's tasks that show up in real, work-related AI use. Medical records specialists come in at about 67%, close behind computer programmers (75%) and customer service representatives (70%) and level with data entry keyers.[1] Clinicians are at the other end of the scale. Healthcare practitioners are the largest group of high-wage occupations that make little use of AI's most capable models, and when physician tasks do show up, 82% of the time it is patients asking about their own symptoms.[2]

The first large AI effect in health care will be a cheaper claim, well before it is a better diagnosis.

Augmentation in the chat window, automation in the pipes

In Claude.ai conversations, 52% of usage augments the person and 45% automates the task. Enterprise API traffic is 64% automated. Office and administrative tasks are 8% of consumer usage but 15% of API usage, and the fastest-growing API workloads include invoice processing and calendar management.[3] College-level tasks see about 12× speedups, compared with 9× for high-school-level tasks, and a task that takes a person 3.1 hours alone takes about 15 minutes with AI. Success rates, though, are 67% in chat and 49% through the API,[3] so faster production creates a new bottleneck in verification.

The first signal is in hiring

Anthropic finds no meaningful rise in unemployment among highly exposed workers since late 2022, yet a 14% drop in the job-finding rate for 22-to-25-year-olds in exposed occupations. Each 10 percentage points of observed coverage goes with a 0.6-point lower BLS growth projection through 2034.[1] In a survey of 9,700 users, more than a third put the chance of a junior colleague losing their job above 60%.[2] Today's entry-level roles are where tomorrow's revenue-cycle directors, utilization managers and payer-contracting leads learn the work.

Three shifts leaders should plan for

From labor arbitrage to workflow ownership. Much of health care administration is outsourced and priced per person, per hour or per claim. When the marginal cost of a coded chart or a worked denial falls toward the cost of compute, the value moves to whoever owns the workflow, the data and accountability for the outcome.

An administrative arms race. Payers will automate prior authorization and claim review; providers will automate appeals and documentation. Unless the rules change (standardized electronic prior auth, shared adjudication logic), the system could spend less per transaction and still process far more of them.

Verification becomes the core skill. AI tends to deskill at first, because it takes over the tasks that need the most education, while some managerial roles get upskilled as routine admin drops away.[3] Organizations have to decide on purpose which roles become reviewers of AI output and which become owners of exceptions and judgment.

What this does not mean

Observed exposure measures where AI is used. It does not measure whose job goes away. Health care is regulated, reimbursement-driven and slow to change contracts, and Anthropic's own productivity estimate falls from 1.8 to about 0.7–0.9 points of annual growth once success rates and bottlenecks between tasks are taken into account.[3] Expect the change to be gradual in any one year and large over five.