↓ Skip to main content
Expert LinkedIn
AI AI in Healthcare

Clinical AI's Next Bottleneck Isn't Accuracy. It's Who Pays - and by What Standard.

10 min read
Dr. Sophia Patel
Dr. Sophia Patel AI in Healthcare Expert & Machine Learning Specialist

By mid-June, the FDA’s public list of AI-enabled medical devices had reached 1,524 entries. One month later, the White House and the Department of Health and Human Services were quietly organizing a sprint on benchmarking and evaluation for clinical AI, while CMS signaled that it wants to rethink how Medicare pays for clinical software in 2027.

Taken together, those moves tell you something the industry’s product demos still obscure. Clinical AI does not have a breakthrough problem. It has a market-design problem.

A clinical balance scale in a stark hospital-white setting overloaded with glowing radiology films, AI report printouts, and compact diagnostic devices on one side, while thin reimbursement forms and benchmark cards on the other fail to counterweight it.
Clinical AI is producing more usable tools than the health system can yet price or standardize.

The product wave is already here
#

If you only track technical progress, the story looks straightforward. On June 25, the FDA granted breakthrough designation to two generative-AI radiology tools: Cognita’s report-drafting system, now inside Radiology Partners after last year’s acquisition, and Aidoc’s First Read product for detecting and describing four life-threatening findings on chest X-rays (STAT, June 25). Two days earlier, OpenEvidence said it would add an FDA-cleared heart-disease detection tool developed by Pierre Elias, turning a physician chatbot into a more directly clinical product surface (STAT, June 23). On June 24, Casey Ross reported on a Nature study linking cardiac fibrosis to sudden cardiac death risk, showing again that AI is increasingly useful not only for administrative summarization but for finding clinically meaningful signal in data physicians already capture.

Katie Palmer’s April reporting captured the problem cleanly through coronary artery calcium screening. Every year, patients receive roughly 19 million chest CT scans in the United States that may contain visible signs of cardiac risk, yet an estimated 20% to 40% of incidental calcium goes unreported. An AI model that flags those findings sounds like an obvious win. It may even be one. But the moment such a model starts surfacing more risk, the hard questions begin: Who pays for the follow-up? Who absorbs the new cardiology workload? Which downstream tests count as medically necessary because the algorithm saw something a general scan was not ordered to assess?

Model capability can create clinical possibility. It does not, by itself, create an institution willing to reimburse consequence.

CMS just admitted the real bottleneck
#

On July 16, CMS effectively acknowledged as much. In its proposed outpatient and physician payment rules for 2027, the agency signaled that it wants a more consistent payment structure for clinical software and AI - one that reflects the technology’s effect on patient outcomes, not just its existence as software (STAT, July 16). That sounds administrative. It is actually the most important healthcare AI story of the month.

For years, the field has treated regulation as the main gate. Get FDA clearance. Get breakthrough designation. Publish the validation study. Yet CMS is confronting a harder institutional fact: American medicine has pricing grammar for physical equipment, procedure time, and consumables. It does not yet have stable grammar for software that changes detection rates, redistributes clinician labor, or triggers downstream care probabilistically rather than visibly.

That is why Palmer’s reporting matters beyond payment-policy specialists. A cotton swab, a contrast agent, and the wear on a CT scanner are all legible to Medicare accounting. An algorithm that finds coronary calcium on a scan ordered for another reason is not. The same problem shows up in sepsis detection, radiology drafting, remote monitoring, and clinician-facing chatbots that increasingly want to move from advisory layer to workflow infrastructure.

If my June analysis of The AI War Inside Your Hospital Bill described how opaque administrative AI can poison trust before a patient ever meets a clinical model, this is the mirror image of that problem. Here, the tools may be clinically promising, but the economic architecture around them is still immature.

In other words: healthcare AI is no longer stuck only at the level of proof-of-concept. It is increasingly stuck at the level of reimbursement logic.

Benchmark wars are not a payment framework
#

The second institutional lag is standards.

On July 23, Mario Aguilar reported that the White House, HHS, FDA, and the Office of the National Coordinator for Health Information Technology would convene experts for a one-month sprint on the benchmarking and evaluation of clinical AI, aiming for a consensus set of principles (STAT, July 23). That is a telling development. Governments do not organize sprints like this when a market already knows how to compare products credibly. They do it when evidence is arriving faster than institutions can normalize it.

The timing was not accidental. Throughout July, clinical AI vendors and commentators were still arguing over what recent benchmark studies did or did not prove. Brittany Trang’s July 29 AI Prognosis newsletter laid out why benchmarking clinical LLMs from OpenEvidence and Doximity is complicated. The trigger was a mid-June Nature Medicine paper comparing OpenEvidence and UpToDate Expert AI with general-purpose LLMs, followed by a July 9 counter-study backed by OpenEvidence that claimed stronger performance from its own system (STAT, July 9).

This is a payment problem. Payers, hospital committees, and procurement teams cannot build a durable reimbursement or deployment framework on the back of dueling headlines about who “beat” whom on a benchmark. A model may win on a narrow clinical question set and still fail as a practical tool if the benchmark does not travel across institutions or workflows.

Earlier this month, in Healthcare AI Got Its First Patient-Facing LLM Approved. The Question Nobody Will Answer., I argued that approval without a clear accountability chain creates strategic ambiguity. Benchmarking without a shared evaluation standard creates the same ambiguity one institutional layer higher. It lets everyone say their tool is validated while avoiding the harder question: validated for what exact use, in which environment, with what downstream burden, and against which clinically meaningful comparator?

Until those questions have a common answer, the benchmark wars will generate attention faster than they generate trust.

Performance alone still does not predict adoption
#

There is a third lesson hiding underneath the payment and standards debate: in hospitals, performance rarely wins on performance alone.

Palmer’s May 12 reporting on sepsis algorithms should be required reading for anyone who still believes a superior model naturally takes the market. Her conclusion was explicit: performance alone does not predict victory. Startups with promising models still have to overcome Epic’s installed workflow, hospital procurement inertia, alert-fatigue history, and the practical question of how clinical benefit converts into economic legitimacy. The old Epic sepsis algorithm failed in the real world despite looking persuasive on paper. Half a decade later, new entrants such as Bayesian Health, along with LLM-assisted note-mining approaches, are still fighting not just for evidence but for a place inside hospital operations.

That matters because clinical AI is not sold into a blank slate. It enters a deeply integrated environment where distribution, workflow position, and reimbursement shape whether a tool feels like infrastructure or like extra labor. Radiology Partners buying Cognita and OpenEvidence layering in a cleared heart-disease product both make sense in that light. Even Mario Aguilar’s June 16 reporting on a biotech that turned a failed trial into an AI model points in the same direction: the strategic value is not only the model itself, but the ability to repackage evidence into something operationally legible (STAT, June 16).

The company that wins the next phase of clinical AI will not necessarily be the one with the prettiest ROC curve or the most impressive demo. It will be the one that can make its evidence, workflow role, payment logic, and monitoring requirements intelligible to the same skeptical institution at the same time.

What a mature clinical AI market would actually ask
#

If the market were more mature than its marketing, every serious evaluation of a clinical AI product would begin with three questions.

What outcome is being purchased? Not “What can the model do?” but what exactly the health system is buying: fewer missed findings, fewer readmissions, faster report turnaround, earlier intervention, less clinician time, or lower downstream cost. These are not interchangeable. A model that finds more disease may be clinically impressive and financially destabilizing if the payment system was not designed around the new follow-up load.

Who absorbs the second-order work? Opportunistic screening tools, sepsis alerts, and risk stratification systems do not stop at prediction. They create queues, callbacks, consultations, repeat imaging, appeals, and documentation. Medicine is full of technologies that looked efficient until their spillover work surfaced somewhere else in the system. Clinical AI will repeat that history unless payment and staffing models account for the burden honestly.

Which benchmark travels? A result that holds in one health system, with one data pipeline, one note style, or one patient mix, may not survive translation. That is why the HHS benchmarking sprint matters so much. The healthcare system does not need one more absolute leaderboard. It needs evaluation principles strong enough to compare tools across settings without pretending that all settings are the same.

Those questions decide whether a tool remains a pilot, becomes a code, or disappears after conference-season hype.

The uncomfortable truth
#

The FDA’s list is already in the four digits. Breakthrough designations are spreading into generative radiology. Chatbots are moving closer to clinical action. AI systems are finding overlooked cardiac risk, fighting over sepsis detection, and trying to turn failed trial data into product-grade insight. The supply of technical possibility is no longer the scarcest thing in the room.

The scarcest thing is institutional agreement.

Agreement on how to benchmark clinical AI without confusing publicity for evidence. Agreement on how to pay for software whose value shows up probabilistically and downstream. Agreement on who carries the operational and legal burden when the model changes what clinicians notice, document, or act on.

That is why this moment matters more than another “AI beats doctors” headline. Healthcare is quietly moving from a world where the big question was whether these models could do useful work to a world where the harder question is whether the system can describe, price, and govern that work honestly enough to let it scale.

Clinical AI will not become routine care when the demos get more dazzling. It will become routine care when the evidence, payment code, and accountability chain finally point at the same thing.

Working on clinical AI adoption, evaluation, or reimbursement and seeing the gap between product claims and hospital reality? I would like to hear about it.

Email me at sophia.patel@tlnw.uk

References
#

AI Content Notice

This article was created using artificial intelligence technology. Whenever possible, we include references and sources to support the information presented. Readers are encouraged to consult these sources for further information. While we strive for accuracy and provide valuable insights, readers should independently verify information and use their own judgment when making business decisions. The content may not reflect real-time market conditions or personal circumstances.

Related Articles

Tags

#clinical ai #healthcare reimbursement #cms #fda #clinical benchmarking