Can You Bill for AI? Medicare Already Cut Your Pay for the Efficiency.

There is a question that comes up in almost every AI purchasing conversation, usually about twenty minutes in, usually from whoever signs the checks: can we bill for this?

In 2026 the honest answer is almost always no. The FDA has authorized well over a thousand AI-enabled medical devices. Medicare pays separately for a small handful of them. Your ambient scribe has no code and never will. Your coding assistant has no code. Your prior authorization automation has no code.

Meanwhile, CMS finalized a 2.5% cut to the work RVUs of nearly every non-time-based service you perform, on the stated theory that efficiencies from technology and workflow accumulate over time and were never captured in the original valuations.

That is the deal on the table right now. The efficiency has been priced in. The tool has not been paid for. This post explains exactly which AI services do generate payment, why the rest don't, the taxonomy that decides which bucket a product lands in, and what CMS is proposing to change — including a comment window that closes on September 14, 2026.

The one-paragraph version

Medicare pays separately for AI only when the software does work a clinician would otherwise have had to do, and the CPT code says so. That is a very short list: autonomous diabetic retinopathy screening (92229), and as of January 1, 2026, AI-derived coronary plaque quantification (75577), which became the first Category I code with assigned RVUs to recognize augmentative software analysis as a distinct billable service. Everything else is either a Category III code that pays nothing by default, or a tool with no code at all. Ambient documentation falls in the last group permanently — it is a practice cost, not a service, and its return has to come from time and completeness rather than a new line on the claim. At the same time, the CY 2026 fee schedule applied a 2.5% "efficiency adjustment" to work RVUs and intraservice time for most non-time-based services, meaning CMS has already assumed the productivity gain whether or not you bought anything. The CY 2027 proposed rule signals where this is heading: a rename of software-based services, a large set of requests for information about paying for AI-enabled care, and tighter rules on remote monitoring that will hit vendor-supported models hard.

What Medicare actually pays for AI today

92229 — the autonomous one

CPT 92229: "Imaging of retina for detection or monitoring of disease; point-of-care autonomous analysis and report, unilateral or bilateral."

This is the landmark. It describes a diabetic retinopathy screening in which the software produces the diagnostic result at the point of care without a physician interpreting the image. CMS established a national payment rate for it in the CY 2022 fee schedule — the first time Medicare nationally priced a service where the interpretive work is performed by an algorithm rather than a person. The descriptor was revised effective January 1, 2023 to replace "automated" with "autonomous," which sounds cosmetic and isn't: it aligned the code with CPT's formal AI taxonomy.

The rate has drifted down since. National payment was roughly $47 in 2022, about $46 in 2023, and roughly $40 in 2024. Check your own Medicare Administrative Contractor rather than assuming the national figure — carrier-level pricing for this code has varied enough that ophthalmology advocacy groups have publicly pushed back on individual carriers setting it below peers.

Utilization has grown fast. By 2023 claims data, autonomous retinal screening had become one of the most widely used AI-aided procedures in the country.

75577 — the 2026 arrival

CPT 75577: quantification and characterization of coronary atherosclerotic plaque to assess severity of coronary disease, derived from augmentative software analysis of a coronary CT angiography data set, with interpretation and report by a physician or other qualified health care professional.

This is genuinely new and worth understanding even if you're not a cardiology practice, because it establishes the template. Effective January 1, 2026, Category III codes 0623T–0626T were deleted and replaced by this Category I code. CMS accepted a RUC-recommended work RVU and set a crosswalk for the practice expense component, so unlike its Category III predecessors it carries real assigned value.

Two practical notes: it is reported once per coronary CTA, and it does not include interpretation and reporting of the coronary CTA itself (75574), so it is billed alongside the base study rather than instead of it. And if your charge capture still submits 0623T after January 1, 2026, expect a non-coverage denial — the old codes are gone.

The significance is the precedent. 75577 is the first Category I code with established RVUs recognizing AI-augmented diagnostic analysis as a separately billable service. For medical imaging and diagnostics AI, that is the pathway every vendor is now trying to walk.

0992T and 0993T — the ones that show the limits

Also effective January 1, 2026, two Category III codes cover noninvasive cardiac risk assessment derived from augmentative software analysis of perivascular fat: 0992T without a concurrent cardiac CT, 0993T with one. Both include physician interpretation and report.

They are real codes. They are also Category III, which in practice means they are carrier-priced, frequently paid at nothing, and require documentation establishing medical necessity before a payer will even consider them. Category III codes exist to track emerging technology and build a utilization record that can eventually support conversion to Category I. They are a waiting room, not a payment.

The gap, in numbers

• The FDA's list of authorized AI-enabled medical devices passed 1,300 entries in late 2025 and counts of the list in early 2026 run past 1,400. The FDA itself notes the roster "is not a comprehensive resource," so the real figure is higher.
• The CPT Editorial Panel has accepted roughly 43 AI-classified codes.
• Of those, as of January 2026 only a couple are Category I codes carrying assigned RVUs.
• Estimates of how many FDA-authorized AI devices have a dedicated, permanent Medicare payment pathway run to about ten.

So: over a thousand cleared products, dozens of codes, a handful of payments. If a vendor tells you reimbursement is "coming," ask them which of those three numbers they're planning to move.

This is why the economics differ so sharply by category. Diagnostic imaging AI — Overjet and Pearl in dental radiography, the broader imaging and diagnostics category in radiology — at least has a plausible route to a code, because the software produces a finding. We looked at where that stands for dental specifically in what's FDA-cleared in dental AI. Clinical decision support tools like Glass Health generally do not, because the clinician still makes and documents the decision — and the FDA line between a CDS tool and a regulated device moved again this year, which we covered in the FDA's 2026 CDS guidance. Autonomous coding platforms such as CodaMetrix and Nym Health sit furthest from a code of all: they operate on the billing and RCM side, where the return is measured in coder hours and denial rates rather than claims.

Appendix S: the three words that decide whether you get paid

CPT introduced Appendix S in 2021 to classify AI-enabled services. It is not a coding rule — it is the vocabulary the Editorial Panel uses when deciding what a code should say, which turns out to determine everything downstream about payment.

• Assistive — the software detects and displays clinically relevant data without deriving a parameter or generating an interpretation. A physician still reads it and writes the report. The software improves accuracy or speed, but the billable service is unchanged. No separate code.
• Augmentative — the software produces a quantitative or categorical output that is qualitatively different from its input: a risk score, a clinical scale, a classification with demonstrated validity. A physician still interprets and reports. This can support a code, as 75577 and the perivascular fat codes show, but the physician work often remains captured by existing codes.
• Autonomous — the software independently derives parameters and generates clinically meaningful interpretations without concurrent physician involvement. Appendix S defines three levels, from recommendations a physician can accept or reject, up to systems that initiate management with oversight. This is where standalone payment lives, because the machine is doing work a person otherwise would.

At its May 2026 meeting the CPT Editorial Panel accepted revisions to Appendix S, sharpening the boundaries between the three categories after four years of real-world application — in particular tightening what counts as a "clinically meaningful output."

The rule of thumb this gives you: the more the software replaces clinician interpretive work, the more likely it is to be paid for; the more it merely helps a clinician work faster, the more likely it is to be a cost you absorb. Most of what a small practice buys is the second kind.

Your ambient scribe will never have a CPT code

This is worth stating plainly, because vendor ROI models sometimes imply otherwise.

An AI scribe is not a service rendered to a patient. It is a documentation tool used by a clinician — assistive, in Appendix S terms, and arguably not even a clinical service at all. There is no billable event to attach a code to, and there is no serious effort underway to create one. CMS's position on ambient-generated notes is that they are billable in the ordinary way as long as the clinician reviews and signs the note; the note is what's billed, not the software.

That means the business case for a scribe is time, capacity, retention, and documentation completeness. It is not new revenue from a new code. Our rollout guide covers what the realistic time savings actually look like.

There is a tempting adjacent argument — that a scribe pays for itself by capturing higher-level E/M codes you were previously under-documenting. Some of that is legitimate. It is also exactly the pattern payers have built downcoding algorithms to catch, which we covered in detail in what happens when your AI scribe changes how you bill. Building a purchase justification on projected code-level shift is a bad idea for both financial and compliance reasons.

The part nobody circled: the 2.5% efficiency adjustment

Here is the piece that changes how you should think about all of the above.

In the CY 2026 Medicare Physician Fee Schedule final rule, CMS finalized an efficiency adjustment that reduces work RVUs and corresponding intraservice time by 2.5% for most non-time-based services. Time-based codes are excluded. Codes newly created for 2026 were not subject to it in their first year.

The methodology: CMS summed the last five years of the Medicare Economic Index productivity adjustment, arriving at 2.5%. The rationale, in CMS's own framing, is that efficiencies in performing medical services accrue over time through workflow improvements and technology adoption, and are not captured in the normal process of developing time and work RVUs. CMS finalized recalculating the adjustment every three years, with the next recalculation not scheduled until CY 2029, and the CY 2027 proposed rule continues to apply it to non-time-based services.

Read that again in the context of an AI purchase. CMS has made a blanket assumption that technology has made physician work more efficient, and has reduced the work component of nearly every procedural service accordingly — across the board, for everyone, whether or not the individual practice has adopted anything.

Specialty societies have called it a stealth cut, and structurally that is what it is. It is also the clearest signal available about how CMS thinks about clinical AI: as something that reduces the resource cost of care, and therefore the payment for it, rather than as something that warrants a new payment of its own.

For context on the surrounding arithmetic, the CY 2026 conversion factors were $33.5675 for qualifying APM participants and $33.4009 for non-qualifying participants.

What CY 2027 proposes — and why September 14 matters

CMS released the CY 2027 Medicare Physician Fee Schedule proposed rule (CMS-1848-P) on July 14, 2026. Comments are open through September 14, 2026. If you have ever wanted a say in how AI gets paid for, this is the least theoretical opportunity you will get this year.

Payment baseline. The proposed conversion factors drop to $33.1693 for qualifying APM participants (a 1.19% reduction) and $32.8409 for non-qualifying participants (a 1.68% reduction), largely reflecting expiration of the temporary 2.5% statutory increase. CMS also proposes a practice expense "stabilization adjustment" capping annual PE RVU changes at ±5% for most existing services.

"SaaS" becomes "SaMS." CMS proposes retiring software-as-a-service terminology in the Medicare context in favor of Software as a Medical Service, describing software-based technologies that support clinical decision-making through algorithmic analysis. Naming conventions in a fee schedule are rarely accidental — this is CMS building a category it intends to pay for, or price, on its own terms.

Ten algorithmic laboratory analysis codes would move from the Clinical Laboratory Fee Schedule to contractor pricing under the Physician Fee Schedule. That substitutes MAC discretion for CLFS methodology, introduces beneficiary cost-sharing, and brings the codes under budget neutrality.

The requests for information are the real story. CMS is asking, without proposing anything yet:

• What requirements prevent technology-enabled organizations, including AI companies, from delivering Annual Wellness Visits
• Whether CMS should create dedicated technology-enabled care management codes
• Whether outcomes-based payment is appropriate for these services
• How ambient documentation, clinical decision support, and other AI tools should affect physician payment
• Whether primary care payment should move toward prospective or capitated models
• Whether alternatives to the AMA's CPT and RUC process should be considered, including using ICD-10-PCS procedure codes instead of CPT

That last one is a genuinely significant shot across the bow at the entire code-creation process AI vendors currently depend on.

It is worth noticing the asymmetry here. CMS is still asking whether to pay clinicians for using AI, while it has already deployed AI on the other side of the claim — through the WISeR model's algorithmic prior authorization for Medicare fee-for-service, and through the interoperability requirements in CMS-0057-F. The payer-side automation arrived first and did not wait for a code.

Remote monitoring gets much tighter. Following OIG fraud findings, CMS proposes limiting RTM to established patients, requiring a separately reportable face-to-face initiating visit for both RPM and RTM, and — the big one — paying only when services are furnished by clinical staff employed by the practice rather than outsourced to a third-party company, effective January 1, 2027. A large share of digital health revenue currently runs through vendor-supplied monitoring staff. If finalized, those arrangements stop being billable.

Telehealth transparency. New modifiers BB and BC, effective January 1, 2027, would require disclosure when a practitioner contracts with a platform-owning entity, making those affiliations visible in Medicare claims data for the first time. The telehealth originating site facility fee (Q3014) would rise to $32.65.

To comment, search for CMS-1848-P on regulations.gov before September 14, 2026. Specialty societies file detailed letters, but individual practice comments describing actual operational experience carry real weight in the RFI sections specifically.

What this means if you're buying AI this year

• Build the ROI on time, capacity, and denial reduction — not on new revenue. For the overwhelming majority of tools a practice buys, there is no code. Model it as an operating expense with a productivity return.
• If a vendor's ROI model depends on a Category III code, value that line at zero. Then ask them to show you paid claims, by payer and by MAC, at a specific dollar amount. "There's a code for it" and "we get paid for it" are different sentences.
• If you're in cardiology or imaging, update your charge capture now. 0623T–0626T are deleted. 75577 is the code. Documentation must explicitly state that AI-driven analysis was used to support the physician's work — that requirement is new for many coding teams.
• Check your MAC, not the national rate. For carrier-priced and contractor-priced services, the national average may not describe anything you'll actually be paid.
• Re-examine any vendor-staffed remote monitoring arrangement against the CY 2027 proposal before you sign a multi-year contract.
• Assume the efficiency adjustment applies to you regardless. Adopting AI does not trigger it and declining to adopt does not exempt you. It is already in the rate.

Questions worth asking a vendor

• Is your product classified as assistive, augmentative, or autonomous under CPT Appendix S — and who made that determination?
• Is there a CPT or HCPCS code that describes your service? Category I or Category III?
• Has CMS assigned RVUs or a national payment rate, or is it contractor-priced?
• Can you name three MACs or commercial payers currently paying claims for it, and at what rate?
• What documentation does the physician need to include for the claim to survive review?
• If there is no code: what specifically is the return you're claiming, and how would we measure it in our own data within 90 days?

The bottom line

Clinical AI in 2026 sits in an odd gap. The technology is regulated as a medical device, sold like enterprise software, and paid for like office furniture. Two codes generate meaningful Medicare payment for AI-derived clinical work. A few dozen more exist mostly to accumulate evidence. Everything else is a cost.

At the same time, CMS has already reduced the work component of most services on the assumption that technology is making care more efficient — and is now openly asking whether ambient documentation and clinical decision support should change physician payment further.

If you are evaluating tools, the practical takeaway is unglamorous: pick them for what they do to your day, your throughput, and your denial rate. If a reimbursement pathway shows up later, treat it as upside. Just don't underwrite the purchase with it.

Browse the directory by use case or specialty to see what's actually available, and compare options for family medicine practices or small clinics if you're starting from scratch.

---

This article is informational only and is not legal, medical, coding, or billing advice. Fee schedule amounts, code descriptors, and coverage policies change, and contractor pricing varies by region. Verify all codes and payment rates against the current CPT code set, your Medicare Administrative Contractor, and your individual payer contracts before submitting claims.