In June the AMA sharpened CPT Appendix S, the taxonomy that sorts clinical AI into assistive, augmentative and autonomous. Three codes describe the same photograph of the same retina — separated only by who reads it. Here is twelve years of what Medicare actually paid for each. One dot = 50 screens.
The taxonomy is the pricing theory. Assistive means the software helps and you do the work — it gets folded into somebody else’s RVU. Augmentative means the software does a real chunk and you own the output. Autonomous means the software does the whole task, and only autonomous gets to carry its own money.
Retinal imaging is the cleanest place to watch that theory meet a claims file, because all three tiers exist as sibling codes for one procedure. Press play.
92229 does not exist in this file before 2021. Four years in, it has gone from 430 screens to 5,113 — real growth, 3.5× in two years. And it is still 38% of the volume of 92228, the code that requires a physician to look at the image.
The rate ladder holds up exactly as the taxonomy predicts. In 2024, Medicare’s service-weighted allowed amount was $19.36 for staff-reviewed imaging, $23.49 when a physician reads it, and $40.58 when the software reports on its own. Autonomy is worth 1.7× a physician read, per screen.
Per screen. That is the whole catch. Multiply each rate by its 2024 volume and the ranking inverts:
92228 (physician reads it): $23.49 × 13,366 = $313,967
92229 (autonomous): $40.58 × 5,113 = $207,486
92227 (staff reads it): $19.36 × 1,615 = $31,266
The tier sets your price. Volume sets your revenue, and volume is where the tier gives you nothing. The highest-paying code in the family earns two-thirds of what the middle one does.
All three AI-taxonomy codes together were billed 20,094 times in Medicare fee-for-service in 2024. The ordinary human alternative — CPT 92250, retinal photography with interpretation and report — was billed 3,485,020 times by 45,355 providers.
And the premium is eroding. 92229’s weighted average allowed amount was $45.62 in 2022 and $40.58 in 2024 — down 11% in nominal dollars while volume more than tripled.
The 2021 step is partly a coding change. Watch 92228 jump from 776 services in 2020 to 5,301 in 2021, and 92227 fall by half in the same year. That is the shape of a family of descriptors being restructured — 92229 is created in the same year — not of thousands of clinicians simultaneously changing behavior. Any “adoption curve” drawn across 2020–2021 in this family is measuring the code set, not the market.
This is fee-for-service only. No Medicare Advantage, no commercial, no Medicaid, no cash-pay. About half of Medicare beneficiaries are in MA plans whose claims never appear in this file — and diabetic eye exams are a closed HEDIS gap measure MA plans are motivated to buy. A vendor could have a genuine business that is almost entirely invisible here.
The tier labels are mine, not the AMA’s. 92229 is the one unambiguously autonomous code of the three; I am reading 92228 and 92227 onto the augmentative and assistive rungs because of who reviews the image, which is the same axis Appendix S uses. The AMA does not publish a code-by-code tier list, so treat the ladder as an illustration of the taxonomy rather than a citation of it.
Classifying your tool as autonomous is worth doing — a code of your own beats being bundled into someone else’s RVU, and 43 taxonomically-classified CPT codes now exist across code years 2021 through 2028, up from one in 2021. But the code is a permission slip, not a market. The number that decides whether the business works is how many screens one clinic does in a year, and in this dataset the answer, four years in, is still measured in dozens.
Go look at that distribution: every clinician who billed the autonomous code in 2024, as one chart →