# An AI tool can be FDA-cleared, clinically useful, and still never reach a patient. The barrier is who pays. > Clearance was supposed to be the hard part for medical AI. It isn't. The gap between a cleared algorithm and a bedside is a billing code, and most tools lack one. Reimbursement, not accuracy, is the bottleneck. - **Pillar:** Healthcare - **Author:** Aditya Marin Gasga (Founding Editor) - **Published:** 2026-06-15T17:25:00.000Z - **Tags:** healthcare, reimbursement, cpt-codes, fda ## TL;DR An AI tool can be FDA-cleared, clinically useful, and still never reach a patient, because the real barrier is a billing code, not accuracy. Most clinical AI lacks a CPT code with a payment rate, so reimbursement, not clearance, decides adoption. import Figure from '~/components/article/Figure.astro'; [FDA clearance](/cleared-is-not-proven) was supposed to be the hard part for medical AI. It is not anymore. The real barrier between a cleared algorithm and a patient's bedside is a billing code, and most AI tools do not have one. A device can be authorized, clinically useful, and still sit unused, because the hospital that would run it has no way to get paid for running it. Reimbursement, not accuracy, is the bottleneck almost no one demos. ## The gap nobody clears The structural problem is that Medicare has [no clear benefit category for AI software](https://bipartisanpolicy.org/issue-brief/paying-for-ai-in-u-s-health-care/). Software as a service, the form most clinical AI takes, is not a recognized Medicare benefit category on its own, so each tool has to be squeezed into an existing one or go unpaid. As of January 2026 there are only [26 CPT codes for clinical AI solutions](https://bipartisanpolicy.org/issue-brief/paying-for-ai-in-u-s-health-care/), and many new-technology codes are Category III, which are temporary, used to collect data, and [often carry no payment rate and no guarantee of reimbursement](https://bipartisanpolicy.org/issue-brief/paying-for-ai-in-u-s-health-care/). In practice that means even cleared, capable tools go unbilled. In point-of-care ultrasound, for instance, there is [no pathway to bill AI-only interpretation without physician review, and the AI component itself is usually not separately reimbursed](https://www.acep.org/emultrasound/newsroom/february-2026/ai-in-pocus-an-overview-of-the-current-billing-and-reimbursement). The algorithm works. The invoice does not exist. ## Why hospitals can't just absorb it A hospital cannot run unreimbursed software at scale out of goodwill, because the margins are not there. By the American Hospital Association's own analysis, Medicare paid hospitals [just 83 cents for every dollar they spent in 2023, over $100 billion in underpayments](https://www.aha.org/lettercomment/2026-02-23-aha-response-hhs-rfi-ai-health-care). A system already absorbing that gap has no appetite to add a category of tools it cannot bill for. So the CFO, not the clinician, becomes the real gatekeeper of medical AI, and the question that decides adoption is not "does it work" but "can we get paid."

## The door is starting to open The other side of the ledger is real, and it is where the action quietly is. The [2026 Hospital Outpatient rule established national reimbursement for AI-assisted cardiac analysis](https://www.cms.gov/newsroom/fact-sheets/calendar-year-2026-hospital-outpatient-prospective-payment-system-opps-ambulatory-surgical-center), and the AMA has created Category I codes, the kind with established payment, for several AI-enabled services, part of a broader move by the FDA, CMS, and HHS toward [recognizing AI as a continuous service rather than a one-time product](https://bipartisanpolicy.org/issue-brief/paying-for-ai-in-u-s-health-care/). Medicare's new ACCESS model is [the first traditional-Medicare pathway to pay for technology-supported chronic care delivered between visits](https://www.cms.gov/priorities/innovation/innovation-models/access), the kind of continuous, remote, AI-assisted care that never fit a procedure code. And the [Health Tech Investment Act would create a predictable Medicare pathway for algorithm-based tools, end case-by-case approvals, and guarantee a five-year payment window](https://www.acep.org/emultrasound/newsroom/february-2026/ai-in-pocus-an-overview-of-the-current-billing-and-reimbursement) if it passes. The capital has noticed the asymmetry. US digital-health startups [raised $14.2 billion in 2025, with AI-focused companies taking roughly 54 percent of it](https://rockhealth.com/insights/2025-year-end-digital-health-funding-overview-a-tale-of-two-markets/), precisely because a clear revenue model, not just clinical evidence, is what makes a company fundable. ## The altitude shift Trace a validated model from the lab to the ward. It clears the FDA, posts strong accuracy, and lands in front of a hospital that would genuinely benefit from it. Then it meets the chargemaster, finds no code that pays for it, and stops. The clinical evidence was the easy mile. The reimbursement code is the hard one, and without it the algorithm becomes a slide in a pitch deck rather than a tool at the bedside. The breakthrough that puts medical AI to work is not a better model. It is a CPT code with a payment rate attached, written in a federal register notice that no one live-streams. ## The question to ask first If you evaluate clinical AI, for a health system or as an investor, reorder your diligence. Accuracy is table stakes and the vendors all clear it. Ask instead whether there is a reimbursement pathway today, which code it bills under, whether that code carries an actual payment rate, and who absorbs the cost if it does not. A tool with 99 percent accuracy and no billing code [will not get deployed](/agent-roi-pilot-production). A merely good tool with a Category I code and a payment rate will. In American healthcare, the model that gets used is the one that gets paid. ## FAQ ### Why aren't FDA-cleared AI tools widely used in hospitals? Because clearance does not guarantee payment. Medicare has no distinct benefit category for AI software, and many AI tools lack a billing code with an established payment rate, so hospitals often cannot get reimbursed for using them, which blocks adoption regardless of clinical performance. ### How many ways can clinical AI be billed? As of January 2026 there are 26 CPT codes for clinical AI solutions, but many are Category III temporary codes that carry no payment rate and do not guarantee reimbursement. Only a subset of AI tools has a code with established payment. ### Is reimbursement for medical AI improving? Yes, gradually. The 2026 Hospital Outpatient rule added national reimbursement for AI-assisted cardiac analysis, the AMA has created some Category I codes, Medicare's ACCESS model supports continuous between-visit care, and the pending Health Tech Investment Act would establish a more predictable Medicare route for algorithm-based tools. ### What should buyers of clinical AI evaluate first? The reimbursement pathway, not just accuracy. Confirm which code the tool bills under, whether that code has an actual payment rate, and who bears the cost if it is unreimbursed, because a tool that cannot be billed is unlikely to be adopted.