Paying for AI: Three Case Studies in How Reimbursement Shapes Clinical Adoption
The Real Question Is Not Whether AI Works
Clinical AI is often discussed as a contest of accuracy: Which algorithm finds disease earlier, reads images better, or predicts deterioration more reliably? Katie Palmer’s three-part STAT series, Paying for AI, shifts the question. Even an effective algorithm must fit into hospital workflows, produce a financial return, and survive a Medicare payment system designed largely around physical supplies, clinician time, and conventional procedures.
Together, the articles show that payment does not merely reward adoption after the evidence is established. It can determine which products reach patients, how broadly they are deployed, and what downstream care they generate.
##
Palmer, Katie. 2026. “AI Could Check Millions of CT Scans for Heart Risk. Who Will Pay for It?” Paying for AI, Part 1. STAT, April 15, 2026.
https://www.statnews.com/2026/04/15/coronary-artery-calcium-ai-opportunistic-screening-examined/ (STAT)Palmer, Katie. 2026. “In the Battle of Sepsis Algorithms, Performance Alone Doesn’t Predict Victory.” Paying for AI, Part 2. STAT, May 12, 2026.
https://www.statnews.com/2026/05/12/ai-sepsis-detection-startups-challenge-epic-systems/ (STAT)Palmer, Katie. 2026. “CMS Signals Intent to Revamp How It Pays for Clinical Software and AI.” Paying for AI, Part 3. STAT, July 16, 2026.
https://www.statnews.com/2026/07/16/cms-to-revamp-payments-for-clinical-software-ai/ (STAT)
Part 1: Finding More Disease—and More Spending
The first article examines AI that detects coronary calcium incidentally on chest CT scans obtained for other reasons. Roughly 19 million such scans are performed annually, and an estimated 20% to 40% of visible calcium findings go unreported. Medicare now pays about $15 for certain algorithmic analyses.
This appears to be an ideal use of AI: no additional scan, little marginal computational cost, and an opportunity to identify cardiovascular risk earlier. Yet the economics are complicated. Positive findings may lead to statins, but also stress tests, coronary CT angiography, invasive angiography, and procedures. Hospitals may gain substantial downstream revenue without proof that population-wide deployment reduces heart attacks or mortality.
The article calls this risk “biomarkup”: technology continually creates new measurable abnormalities that generate additional medical activity.
Part 2: The Best Algorithm May Not Win
The second article turns to sepsis prediction. Epic’s first model performed poorly in practice, but its successor remains formidable because Epic is already embedded in more than 40% of U.S. hospitals. More broadly, EHR companies supply an estimated 79% of predictive algorithms used by hospitals.
Competitors may demonstrate earlier or more accurate detection, but hospitals also weigh integration cost, implementation time, regulatory status, and reimbursement. Bayesian Health pursued FDA clearance and may qualify for approximately $62 in supplemental Medicare payment per analyzed patient.
The lesson is blunt: superior performance alone does not overcome an incumbent platform that is inexpensive and easy to activate.
Part 3: CMS Begins Building a Framework
The third article describes CMS’s proposed 2027 policy. CMS would adopt the term Software as a Medical Service, move 21 separately paid software services into New Technology APCs, and identify them with a new O1 payment status.
CMS is also considering whether software should be discounted when billed with another service and whether payment should reflect outcomes or downstream savings. Ten algorithm-related laboratory codes would move toward similar treatment, raising difficult questions involving contractor pricing, CLIA oversight, coding, and the boundary between laboratory testing and stand-alone software.
Pulling the Series Together
The trilogy presents payment as part of AI’s clinical architecture, not an afterthought. Separate reimbursement can accelerate useful innovation—but can also reward excessive detection and downstream utilization. Bundling controls spending but may leave hospitals unwilling to adopt worthwhile tools.
The key takeaways are that workflow can matter more than accuracy, incumbency can outweigh innovation, and CMS increasingly wants evidence that AI improves outcomes or lowers total costs—not merely that another algorithm can produce another billable result.
##
Sidebar: Five Surprises in Paying for AI
1. A $15 AI analysis could trigger tens of millions of dollars in downstream care.
The coronary-calcium algorithm itself is inexpensive. Its financial importance lies in what follows: statins, stress tests, coronary CT angiography, invasive angiography, stents, and bypass surgery. Vendors estimate that this follow-up activity could generate tens of millions of dollars for a large health system.
2. More than half of patients screened by one system appeared to have significant coronary calcium—but most did not need another intervention.
In one health system, 63,000 of 120,000 patients had moderate or severe calcium scores. Yet an estimated 85% might not merit notification because they were already receiving appropriate treatment or had more urgent illnesses. The challenge was no longer finding abnormalities; it was deciding which findings mattered.
3. A technically better sepsis algorithm may still lose to the one already inside the electronic health record.
Epic’s market position can outweigh comparative performance. Hospitals often choose its algorithm because it is inexpensive, familiar, and easy to activate—even when an outside vendor may detect sepsis earlier or more accurately. An estimated 79% of hospital predictive algorithms come from EHR vendors.
4. Some clinical-AI companies have abandoned patient-care tools for billing software.
Palmer reports that the difficulty of selling regulated clinical prediction products has led some companies to shift toward more lucrative, less-regulated tools that help hospitals document or defend reimbursement. The economics can pull innovation away from diagnosis and toward revenue-cycle management.
5. CMS may move some algorithmic laboratory services outside the traditional laboratory payment system.
CMS proposes treating certain analyses of laboratory-generated data more like imaging AI, with hospital payment through New Technology APCs and physician-office pricing by regional contractors. That creates a striking regulatory “no-man’s-land”: the algorithm may be neither clearly governed as a CLIA laboratory test nor clearly regulated as an FDA device.
