This is the first of three parts on a single question: who pays for the check? Medical AI is deployed on the assumption that a clinician will review what it produces. This part asks who pays for the AI itself. Part 2 asks what the review costs in time. Part 3 asks who owns the failure when the review does not catch a problem.
I have spent more than 30 years in gastroenterology, now as a hospitalist.
The claim for this part is narrow. Clearance by the FDA is a license to sell. It is not a payment. Few cleared AI devices have a way of being paid that is their own, and most are bought from hospital and practice budgets. That leaves the review step with no separate payer.
Almost all of the numbers below come from one source, an analysis by Chi-way Li published October 14, 2026 in his series Cleared to Paid. I have read it in full. Where a figure is his classification rather than a measurement, I say so.
The gap
By June 2026 the FDA had authorized 1,614 AI-enabled medical devices, a record 335 of them in 2025. Li could tie at least 79 to a billing code of their own. Those 79 come from 40 companies.
“A billing code of their own” covers several things in Li’s count: a code paid each time the AI is used, a Medicare lab code, a temporary add-on to Medicare’s payment for a hospital stay, and, for three prosthetic-arm controls, an equipment code paid once per limb. It is a generous definition, and it still reaches only 79 of 1,614.
One caution on the count. It is Li’s own, and an update he posted on October 16 says it covers the FDA’s list only. Some AI-based lab tests sit outside that list. I come back to them below.
Li’s summary is that clearance is a license to sell, not a business model. I think that is the right frame. The FDA asks whether a device is safe and effective enough to market.
Who pays instead
The rest are not unpaid. Someone else pays. Li sorts the 1,614 devices by who appears to bear the cost:
About 76% appear to be paid from a buyer’s budget, mostly software that hospitals and imaging groups buy themselves.
About 15% come inside the price of hardware, from scanners to smartwatches.
About 5% are paid per use.
Treat the first number with care. The 76% is a classification, and Li says it rests partly on inference. It is not measured spending. I use it as a map of where the money probably comes from, not as an audit.
The practical meaning is plain. For most cleared AI, the customer is a hospital or a practice, and the money comes out of an operating budget. In most cases the buyer, not a payer, decides whether the tool is worth its price. No payer has decided that it is worth paying for.
Where insurer money does reach AI
Insurer money does reach AI. It mostly arrives by another route.
Li counts 29 devices that ride on services payers already cover, such as heart-rhythm monitoring. He also counts 29 mammography AI devices that sit inside the screening exam’s fee. In both cases the insurer pays for the test, not for the AI. And tests can pay well. Li reports that Medicare alone paid iRhythm, whose AI first reads its heart-monitor patches, about $179M in 2025 for its monitoring service.
Payments on AI’s own codes are smaller and more concentrated. Li reports that Medicare paid physicians $10.3M on AI billing codes in 2024, and 96% of it went to one category, coronary CT analysis. That figure is secondary to CMS data, and I have not re-derived it.
These are different payment streams, and they are different years. The $179M is 2025 payment to a company for a monitoring service. The $10.3M is 2024 payment to physicians on AI’s own codes. Do not subtract one from the other or add them together. The point of putting them side by side is the contrast in kind: the one large sum I found went to a service that contains AI, and the sums on AI’s own codes are small.
Why codes are rare
Why does so little AI have a code of its own? Radiology is the field with the most cleared AI, so the best explanation I found comes from radiology.
Radiology Business reported in December 2025 (Fornell) that Rubin describes the coding system as having two tracks. In the American Medical Association’s CPT system, Category 1 codes carry payment. Category 3 codes are tracking codes and are not tied to payment, so a Category 3 code lets a use be counted without anyone being obliged to pay for it. Rubin describes a higher evidence bar for Category 1, which means a developer must show more to reach the track that pays.
There is also a double-payment rule. A task that is already inherent to a paid exam is not paid twice. If an AI does something a radiologist already does as part of a paid read, there is little left to bill for. The practical result is a narrow door. A tool that mostly assists work inside a paid exam has little to bill, and a tool that does something new must first clear the higher evidence bar for the track that pays.
On how many permanent Category 1 codes exist for AI, I will not print a number. The count is in the low single digits, and sources differ. Rubin’s account and a January 2026 STAT report carried by the Connecticut Hospital Association do not agree on it. STAT also described payment for coronary plaque analysis as “very spotty.”
Li adds a second explanation, and says plainly that it is his working explanation and not a finding. Payers pay extra for an answer the AI produces itself, such as blood flow computed from a heart scan. They do not pay extra for help with work they already pay for, such as a faster read. Help is left to providers to buy, and there it meets what he calls a verification tax.
Here is the idea as I read it. Until AI is reliable enough on hard cases that nobody needs to check it, a clinician reviews each case. So the time AI saves, or the miss it catches, is hard to see and harder to bank. The time saved goes to whoever bills the exam, not to the payer. Li labels this a working explanation, not a finding, and says a later article tests it.
He also names a counterexample against his own idea. The autonomous eye exam needs no clinician’s check and has had a code since 2021, yet Medicare paid physicians only $138K for it in 2024. A code is not the same as a market.
The policy debate got ahead of the data
In 2022, Parikh and Helmchen wrote in npj Digital Medicine about how to pay for AI in medicine. Their worry was that paying for AI per use could cause overuse. They also laid out five alternative payment models.
The 2024 numbers look like the opposite problem. Medicare paid $10.3M on AI’s own codes, and almost all of it went to one category of use. Whatever else one says about that, it does not look like a system straining under runaway per-use billing.
This is my inference, not a finding. The policy conversation focused on overuse, a risk the data cannot yet test, because most AI never gets a per-use payment. Overuse remains a real risk of per-use payment; payment for AI is still too thin to say more.
What changes the picture
Two things are worth watching.
First, CMS has proposed a payment class for software-based services in hospital outpatient care, to start in 2027. Li notes it as a possible change. It is a proposal, and I have not seen it finalized. If it is adopted, software could be paid in its own right in that setting. Whether the review step is priced inside it is a separate question that I found nothing on.
Second, Li’s October 16 update points to AI that the FDA list misses. At least six machine-learning lab tests outside the list, such as Veracyte’s Decipher and Guardant’s Shield, have their own Medicare price and coverage. They follow the same pattern as everything else in this piece: insurers pay for the test’s result. An answer gets paid. Help with a workflow is the part that does not.
The strongest objection
The strongest objection is that payment inside an existing service is still payment. If an insurer pays for the mammogram, and the mammogram includes an AI read, then the AI is paid for. Calling that “unpaid” would be wrong.
I concede that completely. Li’s 29 mammography devices and his 29 devices on covered services are paid for through covered services, and I do not dispute it.
My claim is narrower, and it is my argument. The AI itself rarely has payment of its own. And none of the payment streams I found prices the review step. When a clinician checks an AI read, note or alert, the check is not a line on any fee schedule I found. Payment inside a service covers the service. It does not cover the clinician’s additional time to verify what the AI contributed.
What this means for the check
Put the pieces together. Most cleared AI is paid for by the buyer, usually a hospital budget. Where insurers pay, they pay for a test or service, not for the AI and not for the review of it. The few AI codes that exist pay small sums to a small number of uses.
Put plainly, a tool can be cleared, bought and used daily without any payer ever having judged that the review it demands is worth funding.
So the buyer carries the cost of the tool, and the buyer has to decide how to fund the clinician’s time to check it. The question is what is left of any promised time saving after the check.
Li’s own questions for hospital leaders are the practical version of this. Ask vendors how long the clinician’s check takes, and what is left of the promised saving after it. Ask which budget paid for the last AI tool you rolled out, and whether the time it promised to save showed up in staffing or throughput.
That is where Part 2, “The Unpriced Check,” picks up. For scribes and inbox drafts, the best-measured cases, it asks how much time the review takes, what the studies show, and why I found no study that totals it.
Brian K. Hudes, MD, is a practicing gastroenterologist and writes The Renaissance Physician. Disclosure: he is the founder and CEO of AiMOS Systems Corporation and has a commercial interest in health care technology and clinical artificial intelligence.
Selected sources
Chi-way Li, Most cleared AI is bought, not billed, Cleared to Paid, October 14, 2026 (update October 16, 2026).
Fornell, D., Radiology dominates FDA-cleared AI, reimbursement lags far behind, Radiology Business, December 17, 2025 (Rubin).
Palmer, STAT, January 2, 2026, via Connecticut Hospital Association.
Parikh and Helmchen, Paying for artificial intelligence in medicine, npj Digital Medicine, 2022.