
Why Almost No AI Tool in Medicine Has a Real CPT Code
FDA clearance proves your AI is safe to sell. It says nothing about whether anyone will pay for it. Only three AI services carry a permanent CPT code, and the rest stay stuck without the outcomes evidence payers actually require.
The clearance letter arrived, the press release went out, and the board treated the FDA's signature as the hard part, now behind them. Twelve months later the product is running in a dozen hospitals and the revenue line is close enough to zero that finance stops rounding it. Somebody finally asks the question nobody on the engineering side prepared for: it cleared, it shipped, so where's the money?
Getting cleared and getting paid are two separate fights, and the FDA only referees one of them. The agency decides whether your tool is safe and effective enough to sell. It has no opinion on whether anyone will actually pay for it, and in U.S. healthcare a tool nobody can bill is a tool that stalls at a dozen hospitals instead of scaling past a hundred, however good the underlying model is.
We've argued elsewhere that the model is roughly 15% of the work and the operating layer is the other 85%. Reimbursement lives inside that 85%, and it's the most expensive part of it to learn late, because the bill for skipping it doesn't land until long after the model ships: after the clearance letter is framed on someone's wall, and after sales has already put a revenue number in front of the board.
Three CPT codes are doing all the work
Start with how lopsided the numbers are. Radiology accounts for roughly 80% of all FDA-cleared AI in medicine, and yet, as the Bipartisan Policy Center laid out this year, only three AI-based services have ever earned a permanent Category I CPT code, the designation that marks a service as established enough for routine payment. One of the three, CT-derived fractional flow reserve, accounts for nearly all of the real billing volume: about 14,000 claims and $12.7 million in Medicare spending in 2023. Every other AI code under the physician fee schedule cleared fewer than 3,600 billing instances that same year.
The 2026 code set looks like progress until you check what's actually attached to it. Roughly 26 CPT codes now touch clinical AI in some form, but a code is not a payment. In the 2026 Medicare Physician Fee Schedule, CMS flagged several of the newest AI codes as not separately payable, meaning you can put the code on the claim and Medicare still won't cut a check for it. You can bill against a code like that all year and watch the revenue line stay flat.
One distinction the trade press routinely flattens is worth keeping straight. A 510(k) clearance and a De Novo authorization are different regulatory pathways with different reimbursement histories behind them, and "FDA-cleared" gets used as shorthand for both. Which one you actually hold shapes what a payer expects to see next, so it's worth naming precisely in your own materials even when the write-ups about you don't.
Payers want proof, not a press release
This next part is the avoidable one. Most AI tools don't get paid for a reason that has nothing to do with payers misunderstanding the technology. The evidence they require was never built in the first place.
CMS pays for what is "reasonable and necessary," and private payers ask for the same thing using different words: proof of clinical utility, evidence that the tool changes a patient outcome or removes real cost from a workflow. What satisfies that is a measurable change in what happens to patients or to spending, demonstrated on real cases rather than a held-out validation set. A npj Digital Medicine analysis of radiology-AI reimbursement draws the line cleanly: coverage is strongest in stroke and vascular triage, where multicenter studies tied AI use to measurable workflow and outcome improvements, and weakest in areas like breast imaging, where tools clear the FDA on strong performance metrics while the prospective, real-world trials that would prove benefit at scale are still unfinished.
Once you see it, the mechanism is simple. Payment doesn't follow accuracy; it follows whoever can show, with real-world data, that their tool changed something a payer cares about. A radiology-reimbursement breakdown puts the reason most vendors stay stuck in a single line: they never ran the studies that demonstrate patient benefit. A clearance file proves performance to a regulator, and the people who decide whether the tool gets paid are reading for something that file was never built to hold.
The evidence has to be designed in
And evidence at this level can't be assembled after the fact. It's an engineering output, and you only get it if you designed for it from the start. Showing a payer that your tool improved outcomes means you were capturing those outcomes prospectively, in production, tied to the specific cases the model touched, in a form an outside reviewer can audit. Concretely, that means a data pipeline, an evaluation harness that measures clinical effect rather than model performance, and the continuous-validation discipline that keeps the evidence current as the patient population shifts underneath it. It's the same audit-ready trail your observability stack should already be keeping, pointed at a different reader.
Teams that try to build this after launch usually can't, because the underlying data was never collected. The ones that get paid built the capture into the deployment itself, so "did this tool change outcomes" is a question their system answers on demand with real numbers, not a study someone commissions two years too late.
Look at the handful of AI tools that have actually pried open a payment pathway and the pattern repeats. When CMS granted a New Technology Add-On Payment to Ultromics' EchoGo Heart Failure, the first AI heart-failure diagnostic to receive one, worth up to roughly $1,024 per inpatient stay, the application carried a clinical evidence base, not just a clearance letter. When CMS approved an NTAP for Annalise's CT triage tool for obstructive hydrocephalus, the submission carried hard validation data, an AUC near 98.7%. Both arrived with proof attached. Most AI vendors show up with a demo and an accuracy number instead, and CMS treats that difference exactly the way you'd expect.
The doors opening now reward whoever built evidence first
Before filing any of this under "later," notice which direction the system is actually moving. CMS finalized its Transitional Coverage for Emerging Technologies pathway in late 2024 to speed Medicare coverage for breakthrough devices, and it still runs on evidence. In December 2025, CMS announced the ACCESS Model, which explicitly tests outcome-aligned payments tied to demonstrated clinical improvement. The AMA is weighing a new code family for "clinically meaningful algorithmic analyses," and a bipartisan Health Tech Investment Act has been introduced that would route FDA-authorized algorithm services into dedicated New Technology payment categories for years at a time.
Every one of those points the same way: payment is moving toward demonstrated outcomes and away from novelty. Teams that instrumented for evidence will find these doors easier to open. Everyone else has to build the evidence base first, and that work runs on a longer clock than the policy calendar does.
Before your next payer conversation
You don't need a payer-strategy team to find out where you stand. You need to test whether your deployed system could ever make the case.
- Write the clinical-benefit claim a payer would have to believe, in one sentence: using our tool changes this outcome for this population by this much. If you can't state it crisply, that's exactly why the coverage conversation keeps stalling. You're selling accuracy, and payers are buying outcomes.
- Check whether your system is actually capturing the proof behind that claim. Is it collecting outcome data prospectively, linked to the cases the model touched, in a form you could defend to a stranger? If what you have is model metrics and no outcome data, that gap is where the architecture has to change, and no amount of documentation closes it.
- Pick your payment path before you need it. A Category I CPT code, an inpatient NTAP, a private-payer pilot, an outcome-based contract: each one wants a different kind and amount of evidence. Decide early which you're building toward, because it changes what you instrument starting now. Retrofitting evidence to fit a pathway you chose late is the expensive way to learn this.
Two rules we hold to on every regulated build apply directly here. The evaluation harness that proves the model works has to exist before you ship, because it's the same harness that eventually proves the tool is worth paying for. And the evidence question belongs in the architecture conversation, eighteen months before anybody opens a reimbursement file.
Where the coverage case actually gets built
A coverage case is something you accumulate, not something you file.It gets assembled case by case out of whatever the deployment happened to record, which makes it an architecture question long before it becomes a regulatory or commercial one. That's where a First Build starts: about a month against your real data, ending with a working system and the instrumentation that ties every model-touched case to the outcome that followed it. One pipeline, two readers: your clinical team asking whether the tool works, and a payer asking you to prove it. We have built to that standard on regulated device data, contributing V&V documentation and product automation behind Canary Medical's own FDA De Novo authorization for Persona IQ; see our recovery-intelligence work with Canary Medical.
The window for that decision is narrow and early. You take it when you choose the data pipeline, and from then on you are either accumulating the evidence or planning to reconstruct it from claims data two years later at many times the cost. If your product is cleared, or authorized, and deployed, and still not billable, that's the decision worth revisiting before the next finance review. Scope a First Build and bring the product stuck between approval and coverage.
Related Insights
Insight
Engineering Medical Devices from Silicon to Clinical Workflow
Moving telemetry from a medical device into a clinical workflow requires more than working hardware. This guide details the 6-layer architecture, from firmware and BLE gateways to FDA 524B compliance and RPM billing, needed to ship connected medtech into live clinical environments.
Read More
Insight
Your AI Can Write the Quality Document. It Can't Sign It.
Can you use AI in GxP documentation? Yes, to draft. The FDA's first detailed action on AI in quality systems did not fault the model for writing the SOP. It faulted the humans who never read it, under a rule written decades before AI agents existed. What that means for the documents your team is drafting now.
Read More
Insight
Moving Clinical AI from a Data Export to a Live EHR
Clinical AI models that look finished on a de-identified export routinely stall against a live EHR. The 2026 FHIR mandates standardize the pipe, not the pipeline: what still breaks is terminology mapping, stale reads, and provenance, and all three are engineering problems.
Read More
Insight
24 Hours to Report an Exploit Your Model Can't Log
Learn how to meet the 24-hour reporting rules for embedded AI. Master edge provenance, model reconstruction, and CSIRT notifications.
Read MoreFacing a problem like this one?
Tell us what the business needs to win and who owns it on your side. If it is a fit, a First Build starts within weeks. If it is not, we will tell you why.