In the US, most clinical AI has no dedicated payment. Hospitals buy it from their own budgets and recover the cost only indirectly. A small number of tools have CPT codes or Medicare add-on payments, including autonomous diabetic eye screening, CT-based coronary analysis and stroke triage software, and those tools are the ones clinicians actually use at scale.
There are now well over a thousand AI-enabled devices on the FDA’s list. Ask how many of them a hospital can bill for, and the answer is a small fraction, with only a handful showing meaningful volume. That gap is the single most important fact about the economics of clinical AI, and it gets less attention than any accuracy benchmark.
The argument of this piece is blunt: without a payment code, a clinical AI tool is a cost centre, and cost centres get cut. The tools with codes do not just earn money. They get used, studied, improved and trusted, because a code turns a software expense into a clinical service with a price, a documentation requirement and a place in the workflow. Regulation decides whether an AI can be sold. Reimbursement decides whether it gets used.
How is clinical AI reimbursed today?
Through a patchwork, and mostly not at all.
The default model is that the hospital or practice pays the vendor, usually a subscription or per-study fee, out of its operating budget. No payer sends a cheque for the algorithm. If the AI helps a radiologist read faster or flags a patient earlier, the organisation hopes to recover the cost through efficiency, better coding or fewer adverse events. The same logic we applied to diagnostics in a test without a treatment is just bad news holds here: a result nobody is paid to act on tends to be ignored.
There are four routes out of that default in the US Medicare system.
The first is a CPT code, maintained by the American Medical Association. Category I codes describe established services and are the basis for regular payment. Category III codes are temporary tracking codes for emerging technology; payers are not obliged to pay for them, and many do not. In 2022 the AMA added a taxonomy to CPT, Appendix S, that classifies AI-related work as assistive, augmentative or autonomous depending on how much of the clinical interpretation the machine performs. That taxonomy matters because it shapes which work gets valued, and how.
The second is the hospital outpatient payment system, where Medicare can assign a new technology payment classification to a service while it gathers cost data.
The third is the inpatient new technology add-on payment, or NTAP, which temporarily pays hospitals extra for technologies that show substantial clinical improvement.
The fourth is coverage decisions by Medicare administrative contractors, the regional bodies that decide whether a coded service is reasonable and necessary in their jurisdiction.
Each route is slow and each requires evidence. Together they have produced a very short list of paid AI.
Which AI tools have CPT codes, and what happened when they got them?
The clearest evidence comes from a study published in NEJM AI in 2023, which searched a national database of roughly 11 billion CPT claims from 2018 to mid-2023 for AI-related billing. Only two categories of AI device had accumulated more than 10,000 claims: coronary artery disease assessment, with 67,306 claims, and diabetic retinopathy screening, with 15,097. Everything else was a rounding error.
Both leaders had something in common: a code.
Autonomous diabetic retinopathy screening received CPT code 92229, effective in 2021, for retinal imaging with automated point-of-care analysis. It was the first code built for a device that makes a diagnostic decision without a clinician interpreting the image. The code let primary care clinics, which have no ophthalmologist on site, screen patients during a routine diabetes visit and bill for it. For the first time the economic case for AI screening did not depend on a hospital absorbing the cost.
Coronary assessment is the other example. CT-derived fractional flow reserve, which uses computational modelling of a coronary CT angiogram to estimate whether a narrowing limits blood flow, was tracked under Category III codes for years and accumulated most of those 67,306 claims before graduating to permanent coding. The same pattern is now repeating with AI plaque analysis. The AMA confirmed a new Category I code, 75577, for AI-driven quantification and characterisation of coronary plaque on CT angiography, effective January 1, 2026, covering products from several vendors. Medicare contractors had already issued local coverage policies for the service before that.
Stroke triage took the inpatient route. In September 2020, Medicare granted the first NTAP for AI software to Viz.ai’s large-vessel occlusion detection tool, allowing hospitals to receive up to $1,040 per eligible patient on top of the standard payment. The software automatically flags suspected large-vessel occlusions on CT angiography and alerts the stroke team. The add-on payment is temporary by design, but it signalled that Medicare would pay for software whose clinical benefit came from speeding up a process rather than making a diagnosis.
Why does a code change adoption so much?
Because a code does four things at once, and only one of them is about money.
It creates revenue, obviously. A clinic that can bill for a screen will offer the screen.
It creates a defined service. A code comes with a description of what is being done, by whom, with what documentation. That forces vendors, clinicians and payers to agree on what the AI actually does, which is more than most AI procurement contracts manage.
It creates data. Every claim is a record. The NEJM AI study was only possible because billing leaves a trace. Uncoded AI is invisible to researchers and regulators, which makes post-market surveillance nearly impossible.
It creates legitimacy. Clinicians are more comfortable with a tool that has passed through the CPT process, which involves specialty societies, and payers are more willing to cover it.
The flip side is that a code is a business model, and it can distort. A per-use payment rewards volume. If an AI analysis is paid per scan, there is an incentive to scan more, or to apply the analysis to patients for whom it adds little. Coronary CT and plaque analysis, in particular, will need watching, because the population who could be scanned is enormous and the evidence that plaque characterisation changes outcomes is still developing.
A worked example: two tools, one budget
Consider a hypothetical community health system choosing between two AI purchases, each costing about $150,000 a year.
Tool A is an autonomous diabetic eye screening camera for its primary care clinics. It has a CPT code. Suppose the clinics screen 2,000 patients a year who would otherwise have gone unscreened, and each screen generates a modest payment. The revenue may or may not fully cover the cost, but it offsets a large share, and the clinics can show improved diabetes quality measures, which matter in value-based contracts. The finance committee sees a line item that partly pays for itself.
Tool B is a well-validated inpatient deterioration model. No code exists. Every dollar comes from the operating budget. Its benefit, if it works, is fewer unplanned ICU transfers and possibly fewer deaths, but in a fee-for-service system, fewer ICU days can mean less revenue.
Tool B might do more good. Tool A is far easier to fund. This is the structural bias reimbursement creates, and it is not the fault of anyone in the room.
What would a fair payment system for AI look like?
The strongest counterargument to “we need more codes” is that coding every algorithm would be a disaster. It would fragment payment into thousands of small fees, invite billing for trivial software, and inflate spending on tools whose benefit is marginal. Many health economists argue that most AI should be paid for the way a stethoscope or a lab analyser is: as part of the cost of providing a service, bundled into existing payments, with hospitals choosing what to buy.
That argument is mostly right for assistive AI, where software helps a clinician do a task they would do anyway. It is weaker for autonomous and genuinely new services, where AI performs work that did not exist before, like screening in a clinic that had no way to screen. A sensible system would bundle the first and pay separately for the second, with evidence requirements that scale with the claim.
Legislators have noticed. The bipartisan Health Tech Investment Act, introduced in the US Senate in April 2025 by Senators Mike Rounds and Martin Heinrich, would create a Medicare payment pathway for what it calls algorithm-based healthcare services delivered through FDA-authorised AI devices, including a new-technology payment classification lasting at least five years. As of this writing the bill has not passed. Its existence is a sign that the current routes are seen as too slow.
The other lever is value-based payment. In capitated or shared-savings arrangements, a deterioration model that avoids ICU days is worth money to whoever holds the risk. This is why integrated systems and large accountable care organisations have been early buyers of uncoded clinical AI: they capture the savings. Our piece on digital therapeutics made a related point: the problem is often not the absence of a code but the absence of a buyer who benefits.
What this means for founders, hospitals and payers
For founders, the reimbursement strategy is part of the product, not an afterthought. Decide early whether your tool is a new service that deserves its own code, a feature that improves an existing coded service, or an efficiency play sold to a risk-bearing buyer. Each implies a different evidence plan. A Category III code is a start, not a business model.
For hospitals, look at AI purchases through the payment lens before the clinical one. If a tool has a code, check whether local payers actually cover it and at what rate. If it has none, identify which contract, if any, lets you capture the benefit, and budget accordingly.
For payers and policymakers, the task is to avoid two failures at once: paying separately for software that should be bundled, and leaving genuinely new AI-enabled services with no route to payment at all. The distinction in the AMA’s own taxonomy, between assistive and autonomous work, is a reasonable place to draw the line.
For clinicians, the uncomfortable point is that coded AI will be pushed harder than uncoded AI regardless of relative benefit. Clinical leaders are the ones who can insist that purchases follow evidence, not billing.
Frequently asked questions
Is there a CPT code for artificial intelligence?
There is no single code for “AI.” Specific AI-enabled services have codes, such as 92229 for autonomous retinal imaging analysis and 75577 for AI-driven coronary plaque analysis on CT, effective 2026. The AMA’s Appendix S classifies AI work as assistive, augmentative or autonomous, which guides how new codes are described.
Does Medicare pay for AI software?
Only in limited cases. Medicare pays for some AI-enabled services through CPT codes, outpatient new-technology classifications, or temporary inpatient add-on payments such as the one granted to stroke detection software in 2020. Most AI tools used in hospitals are not paid for separately and are funded from hospital budgets.
What is the difference between Category I and Category III CPT codes for AI?
Category I codes describe established, widely performed services and are the normal basis for payment. Category III codes are temporary codes for emerging technologies, used to track adoption and collect evidence. Payers are not required to cover Category III services, so many AI tools with only Category III codes still struggle to get paid.
Why do AI tools with reimbursement get adopted faster?
A payment code turns an expense into revenue, defines the service, generates claims data and signals legitimacy to clinicians and payers. A national claims analysis in NEJM AI found that the only AI devices with substantial billing volume were those with dedicated codes for coronary assessment and diabetic retinopathy screening.
What is the Health Tech Investment Act?
A bipartisan US Senate bill introduced in April 2025 that would create a Medicare payment pathway for algorithm-based healthcare services delivered through FDA-authorised AI devices, including a new-technology payment classification of at least five years. It had not been enacted at the time of writing.
A clinical AI tool without a payment route is a gift from the hospital to the patient, and hospitals are running out of money for gifts. Until payment catches up with regulation, the AI that gets used will be the AI that gets billed.
Should most clinical AI be paid for separately, or bundled into the price of care? We’d like to hear from clinicians and payers.
