AI CPT Codes for Radiology: What Gets Reimbursed and What Does Not

The short answer: There is no general CPT code for "using AI." The AMA codes the service the software performs, not the software itself, and it sorts that work into three categories in Appendix S: assistive, augmentative and autonomous. Most radiology AI that has a code at all has a Category III code, the temporary kind with a T on the end, and Category III codes are widely treated as investigational and denied. The one durable exception in imaging so far is not a CPT code at all: Viz.ai's stroke module was granted a Medicare New Technology Add-on Payment of up to $1,040 per eligible patient, the first NTAP CMS ever awarded to AI software.

Updated August 2026.

Every radiology AI business case eventually runs into the same question from the CFO: does this thing pay for itself, or does it just cost money? Vendors tend to answer with throughput and turnaround. Finance wants to know whether a claim goes out the door. Those are different questions, and the honest answer in 2026 is that for most imaging AI, the reimbursement path is real but narrow, slow, and not something to build a purchase decision on.

What are AI CPT codes?

CPT codes describe procedures and services. When the AMA started receiving applications for software that does part of the clinical work, it needed a way to separate what the machine did from what the physician did. That is what Appendix S provides: a taxonomy that classifies AI-driven medical services into three categories based on how much clinical work the software performs and how much judgment stays with the physician.

The autonomous tier splits again into three levels. Level I produces a recommendation the physician accepts or rejects. Level II gives a "reasonable opportunity to negate the impending action prior to implementation (e.g., by means of an alert)." Level III implements the action itself with oversight spread across multiple interventions. Almost nothing in diagnostic radiology sits above Level I today, and the triage products most groups are shopping for are firmly assistive or augmentative.

This taxonomy matters commercially for one reason. Separating machine work from physician work is what creates a path to payment at all. If a code cannot describe who did what, a payer has nothing to price.

What is the difference between Category I and Category III CPT codes?

Category I codes are permanent, carry assigned relative values, and are the ones payers routinely price and pay. Category III codes are four digits followed by the letter T, and they are provisional. They exist to track emerging services, to collect utilization data, and to support the case for eventual permanent status. They do not carry assigned values, which means payment is discretionary and usually local.

The practical translation: a Category III code lets you report that you did something. It does not mean anyone will pay you for it. Most major insurers treat Category III codes as experimental or investigational and deny them by default. Consistent reimbursement is unlikely until the AMA converts the service to a Category I code, which historically takes something in the region of five years and is not guaranteed to happen at all.

The conversion does happen, though, and radiology has a clean example. Code 75580, which covers fractional flow reserve analysis derived from coronary CT angiography, is a Category I code, and its descriptor explicitly includes augmentative AI software analysis of the CTA data with physician interpretation and reporting. It replaced the Category III codes 0501T through 0504T. That is the whole path in one line: temporary code, years of utilization data, permanent code with real value attached. It is the model every imaging AI vendor is trying to follow, and very few have.

Which CPT codes exist for radiology AI?

Fewer than the marketing suggests. The set that specifically describes automated or AI-driven analysis of imaging is small, and most of it is still Category III.

0689T, 0690T and 0691T took effect July 1, 2021 and became billable January 1, 2022. Note what 0691T actually describes: opportunistic analysis of a CT somebody already ordered for another reason. That is a genuinely useful thing for a practice to do and a good illustration of why coding this work is hard, because no new scan was performed and no new patient encounter took place.

Does insurance pay for CPT Category III codes?

Usually not. Category III codes have no assigned relative value units, so there is no national payment amount attached to them, and coverage decisions are made locally at each payer's discretion. Most commercial insurers categorize them as investigational and deny them. Some payers have started reimbursing selected AI-assisted imaging services, but coverage is inconsistent enough that you cannot plan around it.

The advice that coding specialists consistently give is to submit the claims anyway. Utilization data is exactly what the AMA weighs when deciding whether to promote a service to Category I, so reporting the code even when it is denied is how the category eventually becomes payable. That is a long game played on behalf of the whole specialty, not a revenue line for this quarter. It also means somebody in your billing operation absorbs the denial work, which is worth scoping honestly before you commit, because a predictable denial still costs staff time to chase down and write off rather than quietly disappearing.

How did Viz.ai get paid for stroke AI?

Not through CPT. Viz.ai went through a different mechanism entirely, and it remains the most instructive story in this space.

CMS granted Viz LVO, the large vessel occlusion triage module, a New Technology Add-on Payment of up to $1,040 per use in patients with suspected stroke. It was the first NTAP CMS ever awarded to artificial intelligence software, and it was later renewed. To get it, the company had to demonstrate reduced time to treatment and improved clinical outcomes.

NTAP sits inside Medicare's Inpatient Prospective Payment System. It pays a hospital above the standard MS-DRG amount when a new technology raises the cost of treating an episode beyond what the DRG covers and delivers what CMS calls a "substantial clinical benefit." Three constraints follow from that, and they are the part vendors skip:

• It is inpatient only. An outpatient imaging center gets nothing from an NTAP.
• It goes to the hospital, not to the reading radiology group. If you are an independent practice reading for a hospital, the payment lands on the other side of the contract.
• It is time limited by design. NTAP is a bridge for new technology, not a permanent payment stream, and it has to be renewed.

So when a stroke AI vendor references reimbursement in a pitch, the useful follow-up is which mechanism, for which site of service, paid to whom, and for how long. Our Viz.ai alternative page covers where that product's strengths actually sit, and Aidoc vs Viz.ai vs RapidAI compares the three acute triage platforms side by side, including their regulatory positions.

What changed in the 2026 CPT code set?

The 2026 code set carries 418 changes in total: 288 new codes, 84 deletions and 46 revisions, and it expands the section covering assistive and augmentative AI services. The AMA has also refined the Appendix S taxonomy itself, sharpening the boundaries between the three categories and tightening what counts as a clinically meaningful output.

That refinement is more consequential than any single new code. The clearer the line between assistive, augmentative and autonomous, the easier it becomes to write a code that a payer can price. Progress here looks like plumbing rather than headlines, and it is the thing to watch over the next few cycles.

What this means if you are buying radiology AI this year

Build the business case on operations, not on billing. The reimbursement path exists, but for the overwhelming majority of imaging AI it currently returns nothing, and the products where it does return something are narrow, mostly inpatient, and mostly paid to the hospital rather than the reading group.

Four questions worth putting to a vendor that mentions reimbursement:

• Which specific code or payment mechanism, and is it Category I, Category III, or an add-on payment outside CPT entirely?
• Which site of service does it apply to, and does that match ours?
• Who receives the payment in our contractual arrangement?
• Can you name customers who have actually been paid, as opposed to customers who have submitted claims?

That last one separates a real revenue story from a slide. If the honest answer is that nobody is being paid yet, that is fine and it is worth hearing plainly, because it moves the conversation back to the ground where these products are genuinely justified: reading capacity, turnaround, and what happens to the urgent study sitting in the queue. Our guide to how much radiology AI costs works through the cost side of the same decision, and whether radiology AI is FDA cleared covers the regulatory claim that usually appears on the adjacent slide, where the difference between cleared and approved matters just as much as the difference between coded and paid.

Frequently asked questions

Is there a CPT code for AI?

There is no single code for AI. The AMA codes the service performed rather than the technology, so an AI-driven service gets a code describing that specific clinical work, classified under Appendix S as assistive, augmentative or autonomous. Several exist, including 0691T for automated vertebral fracture analysis on existing CT and 75580 for FFR analysis from coronary CTA.

Can you bill for AI in radiology?

You can report it where a code exists, but reporting and getting paid are different things. Most radiology AI services carry Category III codes, which have no assigned relative values and are commonly denied as investigational. Submitting them anyway builds the utilization record that supports eventual Category I status.

Does Medicare pay for AI in radiology?

In limited circumstances. The clearest case is the New Technology Add-on Payment under the inpatient system, which paid up to $1,040 per eligible patient for Viz.ai's stroke triage module. That is inpatient only, goes to the hospital, and is time limited. There is no broad Medicare payment for imaging AI in the outpatient setting.

What is the difference between a CPT code and an NTAP?

A CPT code describes a service so it can be reported and priced, typically in the physician fee schedule or outpatient setting. An NTAP is a temporary supplement to a hospital's inpatient DRG payment for a costly new technology with substantial clinical benefit. They are separate systems, and a product can have one without the other.

Will Category III AI codes become permanent?

Some will. Category III codes are typically reviewed for conversion within about five years, and 75580 shows the path completing in imaging. Conversion depends heavily on utilization data and published evidence, so codes that go unreported tend to stay temporary or expire.

If your group is weighing a reporting platform at the same time as an imaging AI purchase, keep the two decisions separate. A dictation or drafting platform is a straightforward operating cost with no billing angle at all, which our page on radiology dictation software and AI voice assistants sets out, while the AI radiology software comparison covers the detection and triage vendors where these coding questions actually arise.