Medicare Just Asked Whether AI Should Change What a Doctor's Visit Is Worth
I spent last week reading 1,592 pages of a Medicare payment rule, which is not a thing I recommend doing on purpose.
Most of the coverage focused on the rate cut. Doctors get paid less in 2027, the temporary raise from last year expired, everybody wrote that story and moved on.
But somewhere in the middle of it, past the tables and the code sets, Medicare stops proposing things and starts asking questions instead. One of those questions is going to matter a lot more to your revenue cycle than the rate cut will.
Here it is, in my words: if technology does part of the work of a primary care visit, how much should the visit be worth?
Nobody has ever asked that in a payment rule before.
What a Request for Information actually is
Quick translation, because the acronym does this section no favors.
A Request for Information is a section where the agency is not proposing anything. It is asking the public what it should think. Nothing changes this year because of it. No code gets revalued, no payment moves.
It is where the next three years get decided.
If you have ever wondered how a rule change seems to arrive fully formed and everybody in the room acts like it was obvious, this is where that starts. Somebody asked a question eighteen months earlier and a small number of people answered it.
What they actually asked
Medicare put four things on the table in this section, and I want to lay them out plainly because the language in the rule does not make them easy to see.
Is primary care undervalued compared to procedures? Medicare says it is increasingly concerned that it is. That is not a new complaint, but it is a new place to see it in writing.
Should an office visit be split into different types? Right now Medicare pays for an office visit without much regard for why the patient is there. The rule floats separating them into three kinds: a visit that is part of an ongoing relationship, a visit for a one-time problem, and a visit that happens because somebody referred the patient. Each would be worth a different amount.
Should there be two tracks for care management, one for technology-enabled care and one for traditional care? That is the one that stopped me. Medicare is asking whether the same service should be worth a different amount depending on whether a machine did part of it.
Should primary care be paid a flat monthly amount per patient instead of per visit? Medicare has piloted versions of this for a decade. It is now asking how to make it permanent.
And in the same breath, in the same section, Medicare flags AI-enabled primary care as a place where fraud is more likely.
So the position is: this technology may change what the work is worth, and also we are watching it.
Why this matters more than the rate cut
The rate cut is a number. Numbers move every year. You model it, you absorb it, you argue about it, you move on.
This is a definition.
Payment rules do not just decide what things cost. They decide what counts as work. And once the definition changes, everything downstream reorganizes around it: what gets documented, what gets coded, what your vendors build, what your staff spends their day doing, what your denials look like.
I have watched that happen a few times now. The definition changes, and then eighteen months later everybody is rebuilding workflows and nobody can quite remember when the ground moved.
The ground is moving right now, and the comment window closes September 14.
Three more places the same thing showed up in one week
Here is what convinced me this is a direction and not a one-off. Four separate stories landed the same week, and they are all the same story.
AI is about to get a billing code. Nurses still do not have one.
The American Medical Association is working on a new category of billing codes for AI that does clinical analysis without a physician doing work at that moment. Reading an image, interpreting a lab, that kind of thing. If it goes through, hospitals could bill for it directly.
Nursing leaders noticed something about that. Registered nurses have never had a billing code of their own. Nursing gets bundled into the room charge. It shows up on your income statement as a cost line and it has never once shown up as a revenue line.
Rebecca Love, who runs the Commission for Nurse Reimbursement, said this: nursing will remain a cost, AI will become a revenue stream.
I have spent thirteen years around revenue cycle and I do not have a good counterargument to that sentence. Comments on the AI coding framework close August 10, three weeks ahead of the Medicare deadline.
We now have real data on what the machine is bad at
Researchers at Stanford and Harvard tested 24 AI systems against 1,100 real cases drawn from actual physician-to-specialist consultations, graded by 29 board-certified doctors.
The finding worth writing down: more than 80 percent of the severe errors were errors of omission. The AI did not usually recommend something harmful. It left something out.
That is a different failure mode than most people are guarding against, and it is a harder one to catch. A wrong answer announces itself. A missing one does not.
There was a second part to that study that I keep thinking about. Doctors using AI performed better than doctors without it, but still scored below the AI systems working alone. The reason given was that the physicians skipped good recommendations the tool had already surfaced.
So the machine leaves things out, and then the human leaves out some of what the machine found. If you are building any kind of oversight process on top of an AI tool, that is the thing to design around.
Your data silos are turning into a denial trigger
Same rule, different section. Medicare is asking how to stop paying for the same test twice.
The reasoning is worth reading closely, because it is about you. Medicare says imaging and lab results get stuck inside whichever system ordered them, providers often do not know a test already exists, and so it gets repeated.
The options on the table are automatic claim edits that reject duplicates, clawing back payments after the fact, and hard frequency limits on certain tests.
Read that again. Your interoperability problem is being converted into your billing problem. Not as a penalty for doing something wrong, just as the practical consequence of data that does not move.
I have written before that denials would eventually get prevented at order entry instead of appealed months later. This is the same idea arriving from the other direction, and the other direction has an enforcement budget.
What I got right in the book, and what I would change
I wrote in RCM 2030 that by the end of this decade the industry would converge on platforms where the record, the billing, and the payer data actually talk to each other, and where AI is not a widget you bolt on but the thing running underneath. I still believe that.
I also wrote that the back end of the revenue cycle would shrink to a thin layer that handles exceptions, because errors would get caught before they became denials. I predicted mature organizations would push denial rates below two percent.
Here is where I would revise it.
I assumed denials would fall because the causes of denials would get fixed. Cleaner data, better front-end checks, real-time eligibility. That part still tracks.
What I did not account for is that the same interoperability that prevents error-based denials also makes an entirely new category of denial possible. If the payer can see that a test was already performed somewhere else, that is not a preventable error. That is a policy decision, and it will get enforced automatically.
So my revision is this. Denials from mistakes are going to keep falling. Denials from policy are going to rise to meet them. Your total denial rate might land right about where I said it would, and the reason will be completely different from what I expected.
The skill that matters in 2030 is not going to be catching your own errors. It will be knowing the rule before the edit fires.
What I would actually do about this
Four things, and none of them require a project plan.
Read the RFI section and send a comment. It is a handful of pages and it is written in plainer language than the rest of the rule. Comments are how these questions get answered, and right now the people answering are mostly trade associations. If you run a primary care practice or employ primary care physicians, you know things about how that work actually happens that nobody at CMS does. Deadline is September 14.
Find out which of your services already have technology doing part of the work. Not to report it anywhere. Just so you know what your exposure looks like if Medicare decides those services should be valued differently. You would rather have that list now than build it in a hurry.
Ask your denial team a question. When they model next year, are they assuming denials come from errors or from policy? If the answer is errors, that model is going to age badly.
Watch the nursing reimbursement conversation. Not because a nursing billing code is coming soon. Because the argument being made is the clearest statement anybody has offered about how this decade is going to sort out who gets paid for what. If an algorithm's clinical work can be billed and a nurse's cannot, that is a choice somebody is making right now, and it is being made in a comment period almost nobody is participating in.
The part that keeps me up
I have been forecasting a hard decade for revenue cycle for a while, and I always assumed the pressure would come from rates. Less money for the same work.
That is happening, and it is not the interesting part.
The interesting part is that we are in the middle of rewriting what the work is. Which encounters count. Which labor is billable. Which data your system is expected to already have. All of it is being decided in documents that run over a thousand pages and get read by a few hundred people.
The rate cut is the headline because it is easy to cover. The questions in the back are what you will be living inside in 2030.
Sources
Centers for Medicare & Medicaid Services. Medicare and Medicaid Programs; CY 2027 Payment Policies under the Physician Fee Schedule. Proposed rule, CMS-1848-P. federalregister.gov/d/2026-14327
Condon, Alan. "New AI billing codes spark concern across nursing." Becker's Hospital Review.Link
Twenter, Paige. "Nurses' AI pushback expands to new front." Becker's Hospital Review.Link
Bruce, Giles. "24 AI tools ranked for patient safety: Stanford, Harvard study." Becker's Hospital Review.Link
NOHARM: Numerous Options Harm Assessment for Risk in Medicine. Stanford University School of Medicine and Harvard Medical School. arxiv.org/abs/2512.01241
Landi, Heather. "CMS proposes major Medicare reforms to shift physician pay, phase out MIPS and expand ACO participation." Fierce Healthcare.Link
Wilson, April E. RCM 2030: Strategy and Survival for Revenue Cycle Leaders. Chapters 1 and 8.
