There is a moment, in the life of nearly every medical AI company, that feels like the finish line and is actually the starting gun. The FDA clearance letter arrives. The team celebrates, the investors send their congratulations, the press release goes out. And then, often, nothing happens. No revenue. No adoption at scale. A cleared product sitting on a shelf, technically authorized to be sold and commercially inert.

This piece is about why that happens so often, and about the single reframe that separates the companies that build a business from the ones that build a very expensive demo. The reframe is this. A clearance proves the technology. It does not prove the business. And in medical AI, the most important model in the company is not the neural network. It is the model of how a dollar travels from a clinical action to your bank account.

The participation trophy

Start with the number that should be on every medical AI founder's wall. More than 1,400 AI-enabled devices have been cleared by the FDA through 2025, the large majority of them in radiology. It is a genuinely impressive body of regulatory work. It is also a graveyard. The vast majority of those cleared products never obtain a dedicated payment pathway, and a product that does not get paid for does not get used at scale, no matter how good the model behind it is.

Clearance, in other words, has become the participation trophy of the field. It is hard to get, it is necessary, and it is nowhere near sufficient. It proves you can clear a regulatory checklist. It says nothing about whether anyone will pay you for what you cleared. The two are different games with different rules, and the second one is the one that has a profit and loss statement attached.

Clearance proves the technology. It does not prove the business. The two are different games, and only one of them has a profit and loss statement attached.

Why clearance feels like the end

It is worth understanding why so many smart teams mistake the entry fee for the game, because the mistake is structural, not stupid. Regulatory clearance is legible. It has a defined pathway, a checklist, a submission, a decision. You can build a plan around it, hire for it, and know when you are done.

It rewards exactly the kind of focused technical execution that strong founding teams are good at. So teams pour themselves into it, hit the milestone, and experience the very real relief of a hard thing accomplished.

Reimbursement offers none of that comfort. There is no single checklist. The path runs through coding bodies, coverage decisions, contractors, and payers, each with their own timelines and their own evidence demands, most of them measured in years rather than months. It is illegible, slow, and political in the small-p sense. So it is easy to defer, easy to assume will sort itself out, and easy to discover, too late, that it was the actual business all along.

The most important model is the business model

Here is the line that should reorganize a medical AI roadmap. The most important model in the company is the business model. Not the architecture, not the benchmark, not the demo. The model of how value created in a clinical moment becomes revenue you can count on.

In medical AI that model has a specific shape, and it is unforgiving. A clinician takes an action. That action has to map to a billable service. That service has to have a code. That code has to be covered by payers, at a rate that supports a business, in a setting where your product is actually used. Break any link in that chain and the revenue does not flow, regardless of how good the underlying technology is. A brilliant model attached to a broken payment chain is not a business. It is a research project with a sales team.

And the corollary, the one worth saying plainly. A vision you cannot execute on payment is not a strategy. It is a daydream with a deck. The field is full of beautiful daydreams. The companies that matter are the ones that treated the payment chain as a first-class engineering problem, as rigorous and as early as the model itself.

The sleight of hand to watch for

Because reimbursement is illegible, it is also where the most artful corner-cutting happens, and founders and investors both need to learn to spot it. The classic move is the code that is not a code. A company says, on stage, that it has secured a CPT code, and the room hears revenue. What the company often has is a Category 3 code, a temporary tracking code that captures utilization and pays, in most cases, nothing. It is a step on the path, and a legitimate one. It is not payment. Presented as payment, it is a footnote dressed up in good lighting.

Real, durable payment comes from a Category 1 code with coverage behind it, and that is a multi-year campaign of evidence, utilization, and coverage decisions, not a line item you secure in a quarter. The distinction between a tracking code and a paying code is the difference between a story and a business, and anyone underwriting a medical AI company has to know how to tell them apart.

Why two years early is the whole game

The reason reimbursement has to be designed in rather than discovered comes down to a brutal mismatch of clocks. A regulatory pathway can run a year or two. A reimbursement pathway, from first engagement to a durable, well-paying code with coverage behind it, routinely runs three to five. Those clocks do not start together unless you start them together.

A company that begins its reimbursement work when its clearance lands is a company that has just signed up for three more years of runway before the business actually turns on, often more runway than it has.

The teams that win compress that gap by running the clocks in parallel. They begin the coding and coverage work, the evidence generation, the payer conversations, years before the product is cleared, so that clearance and payment arrive closer together instead of separated by a chasm.

This is the compounding head start that is almost impossible to catch from behind. The competitor who started two years earlier is not two years ahead on a single task. They are two years into a loop of evidence and utilization and coverage that feeds on itself, and the gap widens while the latecomer is still filling out their first application.

Evidence for clearance is not evidence for coverage

There is a trap inside the evidence question that catches even sophisticated teams, and it is the assumption that the data which earned the clearance will earn the coverage. It usually will not, because the two audiences are asking different questions. The regulator asks whether the product is safe and does what it claims, which is largely an analytical and clinical-performance question.

The payer asks whether covering it changes outcomes or saves money enough to justify paying, which is a clinical-utility and health-economic question. A dataset built to answer the first can be nearly silent on the second.

So the strongest companies design their evidence to answer the payer's question and the regulator's question at the same time, from the same studies, rather than discovering after clearance that they have to start the real evidence campaign over. That means building in the outcome and economic endpoints a coverage decision will eventually demand, while the trial is still being designed, years before anyone at a payer asks for them.

Evidence is the longest lead item in the whole enterprise, and the teams that treat it as a post-clearance task have already lost years they will never recover.

The door is opening, for the prepared

None of this is a counsel of despair, because the landscape is genuinely shifting. CMS has opened outpatient payment for AI-assisted cardiac analysis. The first durable, separately payable codes for AI-driven services are landing, following the precedent set by software analysis tools that fought this battle before them. The pathway exists now in a way it did not five years ago.

But it is narrow, and it rewards foresight specifically. The companies walking through that door in 2026 are the ones who designed for it years earlier. They picked clinical claims that map to a payable service. They built the evidence a coverage decision would eventually demand, before anyone demanded it. They engaged the coding and coverage process while competitors were still polishing the model. They treated reimbursement as the long-lead engineering problem it is. The door opens for the prepared, and preparation here is measured in years.

Painkiller, not vitamin

Underneath the mechanics sits a simpler truth that predicts which products get paid. Payers cover painkillers, not vitamins. A product that solves a real, expensive, painful problem, that prevents a worse and costlier outcome, has a natural path to payment because it saves the system money or prevents harm it would otherwise bear. A product that is merely nice, that adds a marginal convenience or a small efficiency, has to fight for every dollar and usually loses.

So the reimbursement question and the product question turn out to be the same question asked twice. Are you a painkiller. Does removing you from the workflow cause real pain, clinical or financial, that someone with a budget feels. If yes, the payment chain has a reason to exist and your job is to build it deliberately. If no, no amount of reimbursement strategy will save a product the system does not actually need.

The strongest reimbursement plan starts with building something the system cannot comfortably do without.

The radiology lesson

If you want to see this dynamic in its purest form, look at radiology AI, which is the field's canary and its largest graveyard. Hundreds of radiology AI products have cleared the FDA, many of them genuinely good at what they do. And a striking share of them generate little or no revenue, because the work they assist is already paid for under an existing bundle, and the system has no separate line to pay the algorithm on top. The hospital is already being reimbursed for the read. The AI makes the read better or faster, which is valuable, but value is not the same as a payment mechanism, and without a mechanism the hospital is being asked to buy the tool out of its own margin.

That is the lesson the rest of medical AI should study, because it generalizes. A cleared product that improves a service which is already bundled into someone else's payment has no natural path to its own revenue. The improvement is real and the payment is somebody else's.

The companies that escape this trap are the ones whose product maps to a service that can be paid separately, and they figured that out before they built, not after they cleared.

The three doors to payment

It helps to know that there are only a few doors into reimbursement, and each has a different lock. The first is a new, permanent code of your own, the durable prize, which requires years of evidence and utilization and a coding body's blessing, and which pays reliably once you have it. The second is fitting into an existing code or bundle, which is faster but often means the payment is not really yours, it is the procedure's, and you are competing for a slice of a fixed pie. The third is the set of temporary or novel pathways, the new-technology add-ons and breakthrough-linked routes that can provide a few years of separate payment while you build toward the permanent code.

Each door is a different strategy with different timelines and different evidence demands, and choosing among them is one of the most consequential decisions a medical AI company makes, usually far too late.

The teams that win pick their door deliberately, early, and design the product and the evidence plan around the lock they actually have to pick. The teams that lose assume a door will be open when they arrive, and find a wall.

The buyer is not the user

There is a structural feature of healthcare that trips up founders from every other industry, and it sits at the center of the payment problem. The person who uses your product is usually not the person who pays for it, and neither of them is usually the person who benefits financially from it. The clinician pulls for the tool. The facility or the payer pays for it. The system, or the patient, captures the benefit.

Those are three different parties with three different incentives, and a business model that delights the user while ignoring the payer is a business model with no revenue.

Designing for payment means designing for the payer's logic from the start. What does removing cost or risk from the payer's ledger look like, and can you prove it. The strongest medical AI businesses can draw a straight line from what the clinician does with the product to a number on the payer's books that improves because of it.

That line is the business model, and if you cannot draw it, you do not yet have one, however good the technology is.

Know your numbers, walk in with a model

The practical lesson lands on the founder as a discipline. The teams that win at reimbursement do not discover it. They model it. They know, early and cold, what service their product maps to, what that service pays, in what setting, and what evidence the payers will want before they will cover it.

They walk into the coding and coverage process with a plan, not a hope, the same way a serious operator knows their burn rate cold rather than discovering it at the bottom of the runway.

That is the whole posture in a sentence. Build the technology to earn the clearance. Build the business model to earn the payment. And never confuse the first for the second, because the first is the entry fee and the second is the game.

Clearance gets you onto the field. Reimbursement is how you score.

Sources

More than 1,400 AI-enabled devices cleared by FDA through 2025 (1,451 total; 1,104 radiology) - FDA AI-enabled device list, via The Imaging Wire, Mar 2026.

CMS outpatient payment for AI-assisted cardiac analysis (HCPCS G0680; AI-ECG/echo and coronary-plaque codes; ~26 clinical-AI CPT codes by Jan 2026) - Bipartisan Policy Center; FierceHealthcare.

Category III = temporary tracking codes (often unpaid); Category I = permanent paid codes - AMA CPT; us2.ai reimbursement overview.