Medical AI is deployed on the assumption that a clinician will review its work. I found no payer that prices the review.
The Food and Drug Administration (FDA) has now cleared 1,614 AI-enabled medical devices. By analyst Chi-way Li’s count, at least 79 have payment of their own; most of the rest are bought from hospital budgets or ride inside tests insurers already cover. Yet whenever an AI read, note, or alert reaches a clinician, someone has to check it. That check takes time, and I found no payer that prices it separately.
I have practiced gastroenterology for more than thirty years, twenty in private practice and now in the hospital. I know the difference between work that gets billed and work that simply gets done.
Cleared is not paid
Li published his analysis on October 14, 2026. His summary is blunt: Clearance is a license to sell, not a business model.
Real physician voices, twice a week
Free, and one click to unsubscribe.
He estimates that about three in four cleared devices are paid from a buyer’s budget, mostly by hospitals and imaging groups. One caution on that number. It is his classification, and he says it rests partly on inference. It is not measured spending.
Where insurers do pay, the money mostly reaches AI by way of a service the AI sits inside. Li counts 29 mammography AI devices that sit inside the screening exam’s fee. The insurer pays for the exam, not for the software.
None of this is a scandal. Clearance tells you a device may be sold. It says nothing about who will pay for it.
The step nobody priced
Take the best-studied case, the ambient scribe. The software drafts the note. The physician reads it, corrects it, and signs it. That reading is the check. Li uses scribes as his example of AI whose mistakes are cheap to catch because doctors read the notes anyway.
Cheap is not free. Li calls the cost a verification tax: Until AI is reliable on hard cases, someone reviews each output, so the time saved is hard to bank. He labels that his working explanation, not a finding, and I treat it the same way. It is an idea worth arguing about.
Now ask who is responsible for the review. This month the American Medical Association released a governance framework for AI that takes on clinical functions. It is written for agentic AI, but the question it asks about accountability applies here. It says accountability should follow authority, control, knowledge of risk, and the ability to prevent or mitigate harm. It spreads that accountability across developers, deployers, health systems, physicians, and payers.
That is a sensible principle. I found nothing in the framework about who pays for oversight. The physician’s check is assumed, and the cost of it is left where it falls.
A check may not be enough
There is a second problem. Even a paid check may not catch what it needs to.
In a 2026 medRxiv preprint not yet peer-reviewed, researchers gave physicians deliberately manipulated AI advice on written neurocritical-care cases, using one model. Physicians made harmful decisions in 49.4 percent of the targeted decisions with steering AI, compared with 2.0 percent with aligned AI. The preprint found no clear protection from experience.
Read the caveats with the numbers. This was deliberate manipulation, not routine error. The cases were written, not live patients. It tells us little about how often ordinary AI mistakes get through. What it does show is that “a physician will check it” is an assumption that deserves testing, not a safeguard we can take for granted.
The strongest objection
The best objection I know is that review is already paid. A physician’s reading of an AI note is part of the visit, and the visit fee or the salary covers it. On that view, there is no new cost to price.
My answer is an argument, not a finding. Those payments were set before AI review existed, and they do not scale with the number of AI outputs a physician must check. A visit fee does not change with the number of drafts, flags, and alerts. If AI multiplies what has to be reviewed, the same payment may cover more work, and the difference lands on the clinician’s time.
What I am asking
I am not arguing against medical AI. I am arguing that we should name the check, count it, and decide who pays for it before we lean on it.
Li ends his piece with a question that hospital leaders can answer this week. Which budget paid for the last AI tool your hospital rolled out, and did the time it promised to save show up in staffing or throughput?
Brian Hudes is a board-certified gastroenterologist and hepatologist with more than thirty years of clinical experience and a recipient of his specialty board’s thirty-year certification award. He built and ran a private gastroenterology practice for twenty years, then spent more than a decade in hospital-based gastroenterology, most recently as chief of gastroenterology and medical director of GI and endoscopy at a 550-bed Level I trauma center in Pensacola, Florida. He is an assistant professor of medicine at Florida State University College of Medicine.
Those two halves of a career shape the way he writes about medicine and clinical technology. Outpatient practice and hospital-based care fail their clinicians differently, and many systems are designed without a meaningful understanding of both environments.
Dr. Hudes has been building clinical software alongside his medical practice for decades. In 1995, during his gastroenterology fellowship, he co-developed one of the first Windows-based endoscopy reporting systems in the United States. He is now founder and chief executive officer of AiMOS Systems Corporation, where he is developing an AI-native clinical platform for outpatient and hospital-based care. AiMOS has been accepted into NVIDIA’s Inception program.
He writes about the architecture of clinical information, administrative cost growth, physician workforce shortages, board certification policy, and the widening gap between what clinicians need and what the health care technology industry builds.
