Will AI shrink the clinical workforce?

Forwarded this by a colleague? Subscribe free and get it every week.

Issue two. Five minutes, five sections, no hype. Here is what is in it.

In this issue:

Christina Jewett of the New York Times reports that HHS is building a Medicare payment category for AI software that supports care or diagnosis, and that officials have discussed paying "AI doctors" run by tech companies 60 to 80 percent of what a physician earns for the same service. More than 200 companies are already in Medicare pilots. One is Limbic, an AI therapist, whose chief executive told the Times that a single clinician would oversee thousands of its AI therapists. The FDA has no rules yet for agents that prescribe, and the AMA's CEO says the evidence for autonomous models rests largely on simulations. No rule has been proposed.

Why it matters: If Medicare adopts anything like the reported approach, the payer gains a reason to route the refill, the follow-up, and the low-acuity visit to the agent, with you signing off on the output. Watch for two things in your own practice: a payment code appearing for a task you do now, and a supervision ratio nobody has stated, along with who is liable when one of those encounters goes wrong.

Christina Jewett, The New York Times (paywalled); Naomi Diaz, Becker's Hospital Review (free summary)

Since January, the WISeR (Wasteful and Inappropriate Service Reduction) model has required prior authorization for 13 services in traditional Medicare in Arizona, New Jersey, Ohio, Oklahoma, Texas, and Washington, epidural steroid injections and vertebral augmentation among them. WISeR is a new Medicare Innovation Center (CMMI) pilot that uses AI and machine learning, combined with human clinical review, to run prior authorization and pre-payment reviews on a select set of services in traditional Medicare. The Electronic Frontier Foundation pried about 1,000 pages of CMS records loose with a FOIA lawsuit. They show requests running far past the 72-hour target, one unanswered for 83 days, and a vendor, Innovaccer, warning CMS a month before launch that its software was not fully tested. CMS launched anyway.

In Washington, the vendor Virtix denied 53% of more than 6,000 decisions through March, the Seattle Times reports; a vendor in another state denied under 20%. Vendors are paid from the spending they avert. Virtix says its approval rate is now near 70%; how many denials were overturned is not known.

Why it matters: Texas is a pilot state. If you do interventional pain or spine work, log the submission and decision date on every WISeR request, appeal the denials, and report any delay that changed care. The two rates differ sharply; whether the vendors saw comparable patients, services, or time periods is not established.

Lena Cohen, Electronic Frontier Foundation (primary records linked there); Brittany Trang, STAT (STAT+); Suzanne Blake, Newsweek; Jessica Fu, The Seattle Times

Cleveland Clinic's Applied Analytics Lab, led by spine surgeon Ghaith Habboub, built a "morning surgery briefing" that summarizes the preoperative notes and runs 11 models for kidney injury, bleeding, cardiac events, infection, readmission, reoperation, and death. Trained on about 2 million operations, tested prospectively on 60,000, live for spine and orthopedic cases since March. The source is the institution's own magazine: no discrimination or calibration figures, and nothing on whether a briefing changed a plan.

Why it matters: A briefing like this could change the plan before induction: a kidney-injury flag might alter the fluid and nephrotoxin plan, a bleeding flag might mean blood is ready and antiplatelet timing rechecked. The report does not show that any flag changed management or an outcome. It would do more if the flag reached the preoperative clinic days earlier, while there is still time to optimize anemia, glucose, or a cardiac workup. If your informatics group proposes something similar, ask for that routing, ask how each flag is reviewed, and ask whether its predicted risks matched what actually happened to their patients.

Cleveland Clinic Consult QD (staff report, no byline), featuring Ghaith Habboub, MD

In a NEJM Perspective published online September 12, Dhruv Khullar, a Weill Cornell internist and health policy researcher, takes on the prediction most of us have heard in the lounge: once AI agents can do cognitive work, there will be fewer of us. He argues the long run points the other way. It is an opinion piece with no new data, and worth ten minutes anyway.

He grants that AI might turn some clinical services into software, the cure economists have long wanted for medicine's cost disease. Even so, he gives three reasons it need not shrink the workforce.

The first is the Jevons paradox: make a resource cheaper to use and people use more of it. Cataract surgery and joint replacement got faster and safer, and volume rose. There are more radiologists in the United States than a decade ago, not fewer.

The second is the lump-of-labor fallacy, the assumption that there is a fixed amount of work to go around. Work changes as tools change, and about a quarter of Americans live in primary care shortage areas. In the near term he expects AI agents to extend clinicians, not replace them.

The third is that tasks are not skills. AI can automate a note or an ECG read, but the job runs on skills that link many steps: interpreting results, managing uncertainty, negotiating a plan, leading a team. Under what economists call O-ring theory, one failed step can sink the result, and a missed diagnosis costs more once a cure exists. Automate some tasks and the human ones left over gain value.

Here is why the price matters to you and not only the CFO. Khullar says demand rises only if AI is priced near its marginal cost; priced so the vendor captures most of the value, it stimulates far less. An AI paid 60 to 80 percent of the physician fee is not a tool that makes you faster. It is a substitute billing for the same work from the same pool.

Expect the routine follow-up and the low-acuity visit to be carved off first, productivity targets to rise because the tool supposedly freed your time, and a supervision role in which the number of encounters under your name keeps rising, while you hold the liability and see none of the Jevons upside. Priced near cost, the same tool widens access and the volume comes back to you. Pricing is one lever on the workforce question, not the whole answer, and it is being set now. And "in the long run" carries a lot of weight: none of this says what happens to staffing in your department next year.

The takeaway: AI does not have to mean fewer clinicians, but the price will decide whether it works for you or instead of you. When an AI tool arrives in your department, ask three things: which of your tasks it bills for, how many of its encounters you are expected to supervise and whether that time is paid, and whether your productivity targets change once it is live.

Dhruv Khullar, New England Journal of Medicine 2026;395:1041-1043 (Perspective, with an audio interview by Stephen Morrissey)

Before you try this: check your organization's approved-tool list, and never put patient information into a tool it hasn't approved.

On July 28 OpenEvidence launched a free education platform that turns your own searches into accredited CE and, for some boards, MOC credit. It is open to NPI-verified physicians, NPs, and PAs; AKH, Inc. is the accredited provider. (Details are as of September 2026 and will change.) The announcement does not list which boards accept the MOC credit, and I could not confirm the ABA. Check your board portal before you count on it.

Topics come from what you ask, so ask well. One worth trying: "In adults taking a GLP-1 receptor agonist before elective surgery, what do current guidelines and trial data say about holding the drug and aspiration risk? Separate guideline statements from study evidence, and list the gaps."

The stop rule: CME credit is not validation of the answer. Do not use it alone for patient-specific dosing or contraindications. Verify the primary source and your institution's protocol.

OpenEvidence announcement (Business Wire, no byline); OpenEvidence CME user guide. "Leading" and "most popular" are the company's own words. Loading Dose AI has no financial relationship with OpenEvidence or AKH.

Loading Dose AI is for educational purposes only and is not medical advice. Nothing here should guide the care of a specific patient, and reading it does not create a physician-patient relationship. AI tools help gather and draft each issue; I read every source and edit every word. Opinions are mine and not my employer's. Full disclaimer