The health care industry is gearing up for a battle over whether and how clinical artificial intelligence should get paid for.
As of the end of September, the Food and Drug Administration has authorized 1,357 AI-enabled medical devices. But very few of those tools are actively paid for by insurers.
Some health policy experts and clinicians don’t see that as a problem.
“With AI, so much of the conversation is about how do we get paid for the individual technology,” said Ateev Mehrotra, chair of health services, policy, and practice at the Brown University School of Public Health. “If I could wave a magic wand, I would change our paradigm to: ‘How can we use AI to improve the productivity and efficiency of clinicians so that they can care for more patients at a high quality, at a lower cost?’”
Others, especially in industry, worry that lack of payment will stymie adoption and prevent helpful AI from reaching patients.
In 2026, the debate over AI payment will intensify as more devices enter the field. So far, only three AI devices have received a permanent CPT code from the American Medical Association, a critical step in garnering payment from Medicare and, in turn, most private insurers. But there are many more waiting in the wings: The AMA has doled out more than 20 temporary category III codes to AI, many of which will eventually become permanent.
With an eye on the coming flood, the AMA is considering a potential new code category for AI, including tools that don’t require any physician input. It and other medical societies are quickly trying to make a case for physician-friendly reimbursement models.
Meanwhile, the Centers for Medicare and Medicaid Services is grappling with its own payment paradigms for AI, which have largely relied on vendors’ own valuation of their software. Senators have also proposed legislation for formalizing AI payment pathways.
Those processes could take years to be finalized. In the meantime, “health systems are looking at AI and wondering how the heck they pay for this,” said Pelu Tran, CEO of AI governance company Ferrum Health. Here are some examples of how the money will shake out on a messy, case-by-case basis in different clinical fields in 2026.
A fee-for-service signal: Coronary plaque analysis
In January, Medicare will start paying physicians a set national rate of just over $1,000 for using AI that analyzes the type and quantity of plaque in a patient’s coronary arteries. It’s one of only three AI tools that has received a Category I CPT code from the AMA, and its reimbursement patterns will serve as an important data point as the organization and CMS continue to deliberate payment strategies for AI.
Plaque’s predecessor technology, which is also paid at more than $1,000 by Medicare, uses AI to compute the flow of blood through coronary arteries from a coronary CT angiogram. “We’re just now at a place where we’re seeing close to universal payments of the procedure,” said Eric Rubin, primary CPT adviser for the American College of Radiology, in a session on AI payment at the meeting of the Radiological Society of North America in Chicago last month.
In comparison, “payment for plaque analysis has been very spotty,” said Rubin. “It is slowly becoming more uniform, but we will have to see how this progresses over time.”
Part of that will come down to physicians’ understanding of when a patient might be a good fit for the technology, and how well they document their conditions in order to qualify for payment.
“A lot of doctors don’t know to order CT plus the AI,” said Jacob Agris, vice president of product management at ConcertAI, which recently launched a product to make it easier to use and get reimbursed for certain AI, including the coronary plaque analysis. “It can actually flag them, ‘hey, if you have these indications, you should consider this if it’s appropriate for your patient.’”
That kind of workflow tool could help health systems get more AI claims reimbursed — a boon for their bottom lines, but potentially a burden for national health care spending. The first year of standardized receipts from coronary plaque will help regulators and health systems figure out where the right balance lies.
Make the patients pay: Breast imaging
Across the country, more women going in for their annual mammogram now have the option to select an AI add-on to highlight suspicious lesions. But without reimbursement from insurers, it’s usually the patients who pay up — usually around $40 to $50.
“All of us physicians felt like moving to a self-pay model was not our preferred approach,” said Greg Sorensen, chief science officer at large outpatient imaging company RadNet, which charges patients $40 for its AI-based screening program. “We would have preferred that payers would have embraced this right from the start.”
RadNet CEO Howard Berger said the AI-based screening program is generating a profit for the company. The company does about 1.6 million mammograms a year, and roughly half of women opt in to the program, bringing in approximately $30 million in revenue, he said.
This year, the self-pay trend in mammography AI will continue with a new class of algorithms. A small number of radiology centers currently offer AI add-ons looking for breast arterial calcification, for around $90. And Clairity Breast, an AI device that predicts a patient’s five-year risk of developing breast cancer, will begin a pilot program at Beth Israel Deaconess Medical Center at a price of $199, said Clairity founder Connie Lehman, who directs the Breast Imaging Research Center at the Massachusetts General Hospital. One exception is an AI-based interpretation of breast ultrasound, which does have a temporary CPT code.
In 2026, the growing out-of-pocket costs for women, who typically get mammograms every year, could put more pressure on insurers to cover some AI applied to breast imaging.
“For reimbursement, we really do need CPT codes for these AI products because otherwise we’re going to be allowing or enabling patients who can afford it to get the AI, which has improved performance,” said Sarah Friedewald, academic division chief of breast imaging at Mass General Brigham, in a session at RSNA. “And that’s really not a fair way to implement AI.”
Value-based experiments: Opportunistic screening
In the absence of reimbursement or patient payment, health systems can still choose to invest in AI. They just have to believe that the technology will improve efficiency or health care quality enough to be worth it.
“If they judge based on the published literature or their own internal experience that this is adding enough value for them, then I think that we should use that as a sign, because they’re paying out of their own pocket,” said Mehrotra. “If they’re not, then that’s also a strong signal that maybe this tool does not add enough value to our health care system.”
To that end, several health systems — typically large academic medical centers — are implementing opportunistic screening for health conditions using existing radiological images. NYU Langone is experimenting using CT scans to look for signs of osteoporosis, and Emory Healthcare is developing an algorithm to look for breast arterial calcification, a cardiovascular risk signal, in standard mammograms.
“I like to call opportunistic screening this rare triple win in the U.S. health care system, especially for AI,” said Hari Trivedi, co-director of the Health Innovation and Translational Informatics lab at Emory University. Patients get their health risks caught earlier, health systems can capture more revenue from patients referred for follow-up preventive care, and payers can save money in the long term by preventing hospitalizations from major health events.
Opportunistic screening programs are typically implemented as part of clinical research to determine whether catching risky signs in images actually improves patient outcomes — the first hurdle to decide whether a technology is worth using. But a side effect of these studies is that a health system should be able to track their financial impact, too. Health systems may not report on those readouts in the same way, but their results will be worth tracking all the same.
