What Medicare incentives for AI-based devices mean for tech companies — and hospitals

STAT News – Thursday, August 13, 2026
By Katie Palmer

For hospitals, the promise of artificial intelligence is financial as much as clinical. A hospital might deploy an AI device because it promises to throw up an alert that could save a patient’s life — but the technology is far more likely to stick if it proves it can drive down costs.

Convincing hospitals and health systems that a new technology will provide that return on investment is tricky, though. So some AI startups have benefited from a temporary sweetener that helps customers get on board: Certain new technologies can apply to get add-on payments from Medicare for two or three years after they come to market. The tax-dollar-funded payments are meant as an incentive to help get new, expensive medical technologies to patients.

If a hospital thinks, “‘Oh, you’re not 100% sure you have the money, or you don’t necessarily believe 100% of the assumptions in the ROI,’ then you know what? You have this safety net, at least for the next few years,” said Tom Valent, chief business officer at Aidoc.

Aidoc is one of several AI device companies that have benefited from new technology add-on payments, or NTAPs, since 2020. The group of beneficiaries is expanding significantly in October, when five new AI-based devices will qualify for the add-on payments — the largest number ever greenlit in a single year.

“AI is just ballooning in terms of everything we do,” said Michael Chernew, a health policy professor at Harvard Medical School who has researched the impact of AI on health care costs. This year’s AI NTAPs include three software-only devices: They monitor for predicted sepsis, screen echocardiograms for cardiac amyloid, and triage a range of findings in abdominal CT scans. Two more combine hardware and software, using cameras and EEGs to monitor patients for signs of seizures and delirium.

With an NTAP, hospitals can receive extra payments worth up to 65% of the per-patient cost of the new technology when the costs of a patient’s stay exceed Medicare’s standard bundled reimbursement rate. For the new AI devices, hospitals can expect a maximum supplemental payment of anywhere from $61 to screen for sepsis to $2,275 for cardiac amyloid.

Those payouts aren’t enormous, especially compared with some of the hardware-based devices that receive NTAPs. But the add-on program and each of its new AI recipients underscore the challenges of valuing and paying for AI in medicine. Public health researchers and economists say there’s an inherent mismatch when hospitals get paid to use a subscription product on a per-patient basis, potentially incentivizing overuse. And who adopts the AI — and therefore which patients benefit — varies widely depending on who’s selling and who’s buying.

“The system is getting more complex,” said Kushal Kadakia, a resident physician at Massachusetts General Hospital who studies medical device regulation. “This simple payment arrangement intended to subsidize innovation may not be appropriately designed for the different business models we are seeing for innovation today.”

Do NTAPs incentivize tech adoption?

Medicare established new technology add-on payments for inpatient services in 2001, with the goal of incentivizing adoption of novel products that would otherwise be difficult for hospitals to justify spending on. To qualify for the extra payment, a device must meet three criteria: It has to be new, show substantial clinical improvement over alternatives available to Medicare patients, and be expensive enough to drive costs over standard reimbursement rates for inpatient services.

The add-on payments have been especially appealing for devices designated as breakthroughs by the Food and Drug Administration. Since 2020, Medicare has allowed such devices to qualify for NTAPs without meeting the newness and clinical-improvement criteria.

The potential impact of an NTAP was so meaningful for Neuro Event Labs, the maker of seizure-detecting video software Nelli, that the company applied for the payment four times before it ultimately received FDA clearance and secured a maximum add-on payment of $975 per patient. “It clearly helps with the adoption of the technology by having the NTAP,” chief operating officer Kaapo Annala said.

The majority of AI-based devices that have received NTAPs have benefited from the lowered bar for breakthrough devices, including all five of this year’s devices. They snuck in just under the wire: Starting next year, Medicare plans to end that flexibility, arguing that it should only pay for new technology with evidence of clinical benefit.

To AI developers, getting an NTAP is all upside: No matter where you’re starting from, it’s nice to be able to tell potential customers that some of their software costs can be directly reimbursed. But the program’s business impact can vary dramatically depending on the company.

To actually receive an add-on payment, a hospital needs to carefully code for the services — capturing a new procedure code, and applying only for NTAP-eligible patients — which can be a significant hurdle. “Customers are realizing that it’s much harder than they think to get the actual value from the NTAP,” said Pelu Tran, CEO of AI governance company Ferrum Health. An AI vendor might need to provide coding specialists to help — and offering that support is more difficult for young companies that are just selling to their first customers.

“Aidoc of seven years ago, when we had three customers, would it be worth it?” Valent said. Now that the company has more than 100 U.S. customers, though, “they can receive hundreds of thousands of dollars a year, if not over a million, depending on their size,” for using the company’s new NTAP-eligible abdominal CT triage solution. “That’s completely different math.”

Medical AI’s payment mismatch

Those coding challenges belie a more fundamental tension when AI is reimbursed every time it’s used on a patient.

Today, clinical AI is commonly sold to hospitals as a subscription. That’s generally a good fit for the way that Medicare pays for inpatient services, said Hannah Neprash, a health economist at the University of Minnesota School of Public Health. A hospital gets paid a set rate for a patient’s stay based on their diagnosis and the severity of their condition, and a new device’s subscription costs get folded into that bundled amount.

But for the years an NTAP is active, the Centers for Medicare and Medicaid Services essentially takes the cost of that subscription and spreads it out across appropriate cases, coming up with a maximum add-on payment on top of its standard case rate. Since its first AI NTAP, for a stroke triage algorithm, CMS has been concerned about how to accurately determine payments that are based on a manufacturer-set subscription cost. “I would expect them to price far, far above the marginal cost of actually running the algorithm and returning an answer,” Neprash said. Previous research has also shown that the costs manufacturers report to CMS in order to qualify for add-on payments often exceed those actually charged to hospitals.

There’s a standard methodology to do those calculations, but it can still be complicated. Case in point: InVision Cardiac Amyloid, an algorithm that identifies patients with suspected cardiac amyloidosis from routine echocardiograms. When the company submitted its NTAP application, the maximum add-on payment per patient was $162.50. By the time the NTAP was approved in the final rule, it had grown to $2,275 — the most expensive AI add-on payment this year, just over a $2,171 maximum payment for the delirium monitor from Ceribell.

What changed? While most devices get submitted for payment regardless of their output, “we’re likely going to bill for positives only,” said cardiologist David Ouyang, co-founder of InVision. “We’re exploring this dynamic, which is new for CMS,” he said. “No payer currently has a workflow to only pay for positives.” Its final number, he said, was based on a projection of how many positive cardiac amyloid findings — and add-on payments — it would take to cover InVision’s subscription cost.

That kind of math is how two algorithms can have such vastly different price points. The lowest add-on payments this year are for Bayesian Health’s sepsis monitoring system and Aidoc’s abdominal CT triage solution, which will bring in a maximum of $61.84 and $137.53 per patient, respectively. Both are intended to be applied broadly, at the scale of entire ICUs and inpatient abdominal CTs. With such a large patient population, “that cost of the subscription gets diluted to a much lower cost per patient,” said Jerome Avondo, Aidoc’s vice president of clinical research and reimbursement.

Those economics also mean that the value of an NTAP varies widely depending on the hospital using a technology, said Bayesian’s chief medical officer Martin Doerfler. More Medicare patients in the mix means more opportunities to claim add-on payments to offset subscription costs. “Payment per bed could be $100 for the time the patient is there,” Doerfler said, “or it could be $600 or $700.”

The consequences of clinical AI adoption

In the end, all these machinations don’t amount to much for Medicare’s long-term bottom line: New technology add-on payments are built to sunset after a device has been on the market for three years, and add-on payments for 60 qualifying devices in the coming year are estimated to hit about $1.74 billion. But health economists still have questions about how clinical AI adoption during this period could impact system-wide health care costs.

“CMS needs to be very careful in understanding all the different unintended consequences that may arise when an AI tool is adopted,” Harvard’s Chernew said.

When a hospital has already bought a subscription for an AI product and knows it can receive an add-on payment for every additional patient it uses it on, “that mismatch is a problem,” Chernew said. “That gives you a lot of incentive to run the AI algorithm, and it gives you a lot of incentive to follow up on incidental findings.”

Many AI tools are focused on picking up what might have previously been an undetected diagnosis, which could help save lives and thousands of dollars on downstream care. But highly-sensitive AI tools could also make it easier for hospitals to catch and code for more intensive diagnoses — which garner higher reimbursement rates — when they don’t make a clinical difference, Chernew said, a concept others have termed “biomarkup.” “The question is not whether or not it’s picking up sepsis or delirium,” he said, “but whether or not it’s picking up clinically significant and modifiable sepsis and delirium.”

As AI payment patterns begin to reveal themselves in actual adoption and claims data, “it reinforces some fears that how these technologies are deployed may end up costing the system more,” Kadakia said. That potential needs to be weighed against the possibility of clinical benefits — earlier detection, prevention, and treatment of disease — that could save money.

Those kinds of questions are easier to answer with more clinical evidence — which breakthrough devices, including a large number of AI-based breakthroughs, will need to provide to qualify for add-on payments starting in fiscal year 2028.

“The evidence needed for approval by the FDA and the evidence needed for payment through this pathway is diverging in a really interesting way,” Neprash said. Most AI devices come to market through the 510(k) clearance process, notes Kadakia, which is based on whether a device is substantially equivalent to a previously authorized product. “It’s hard to demonstrate the ‘substantial clinical improvement’ part when you’re coming through a pathway that’s fundamentally about substantial equivalence,” he said.

Removing the more flexible pathway for breakthrough devices to qualify for Medicare’s add-on payments reintroduces a “natural tension” between adoption and generating clinical evidence, said Aidoc’s Avondo. “Is it a shame? Yes. Is it going to hurt adoption? Yes. Is it surpassable? Of course — if your technology does have significant clinical impact.”

Access this article at its original source.