The AMA just sharpened CPT Appendix S, the taxonomy that sorts clinical AI into assistive, augmentative and autonomous. The tier is supposed to be the money. So here is the flagship autonomous code — 92229, point‑of‑care automated retinal screening with no physician read — and every single Medicare clinician who billed it in 2024, drawn as one dot.
The pitch writes itself. Assistive gets bundled into someone else’s RVU. Autonomous gets its own code and its own rate. Classify your tool as autonomous, price against the fee schedule, multiply by volume.
Multiply by what volume, though. Below is the 2024 Medicare Part B file for CPT 92229. Horizontal: how many screens that clinician billed all year (log scale). Vertical: the average Medicare allowed amount they were paid per screen. Dot area is beneficiaries. Drag the red line to your revenue assumption and watch how many real billers sit to the right of it.
At the default line — 200 screens a year, the volume that turns an autonomous code into a business — two clinicians in the United States clear it. An optometrist in Texas at 349 screens and an optometrist in Louisiana at 304. The median reportable biller did 18 screens for the entire year: about $730 of Medicare revenue, or one and a half screens a month.
And that is the flattering cut. CMS only publishes a provider×code row when the clinician served 11 or more beneficiaries. 1,143 providers billed 92229 in 2024; only 69 cleared that threshold. The other 1,074 — 94% of everyone who touched this code — billed fewer than eleven screens all year. They are not on the chart because they are, statistically, noise.
Nine optometrists account for 1,347 of the 2,633 visible screens — 51% of the volume from 13% of the dots, and both of the only two billers above 200 are among them. Set that against California, which has 238 providers billing 92229 and 714 screens between them: three screens each, for the year. This is not a market with a long tail. It is a handful of real screening programs plus a thousand clinicians who tried the code once. If you are modelling adoption, the unit is the program, not the physician.
Three CPT codes describe almost the same picture of the same retina, separated only by who reads it. They map cleanly onto the Appendix S ladder, and Medicare does pay up the ladder:
| 2024 Medicare Part B | rate | services | billers | national $ |
|---|
92229 pays 1.7× what 92228 pays per screen — and earns two-thirds as much nationally, because 92228 was billed 2.6× more often. The tier sets your price. It does not set your volume, and volume is the entire difference between a code and a company.
Worse: the autonomous rate is going the wrong way. Weighted average allowed was $45.62 in 2022 and $40.58 in 2024 — down 11% in nominal dollars while volume tripled. The premium for autonomy is being competed and re-valued away in real time.
Push the minimum-screens floor and watch the apparent story change shape. At a floor of 50 screens, ten dots survive and the field looks like a real specialty market with a clean volume–price relationship. It isn’t one: across all 69 billers the correlation between annual volume and average allowed amount is —, which is nothing. The two highest-paid dots on the chart — a nuclear medicine physician in New York at $57.49 and a family physician in California at $56.94 — are geographic wage adjustment and a handful of claims, not pricing power.
The floor is also doing something the chart can’t show you. Every dot here has at least 11 beneficiaries by construction. The left edge of this cloud is a CMS publication rule, not the bottom of the market. The real distribution has a thousand more providers stacked invisibly below x = 11.
This is Medicare fee-for-service only — no Medicare Advantage, no commercial, no Medicaid, no cash. Roughly half of Medicare beneficiaries are now in MA plans whose claims never appear here, and diabetic eye screening is a closed HEDIS gap measure that MA plans have every reason to buy. A vendor could have a real business that is almost entirely invisible in this file.
It also can’t tell you whether 5,113 is small. Set it against the conventional alternative from the same table, same year: 92250, retinal photography read by a human, ran 3,485,020 services across 45,355 providers. Autonomous AI screening is 0.15% of that. Four years after the code went live, with a rate 13% higher than the human version, the automated path has captured about one screen in every seven hundred.
And it can’t tell you the direction. 92229 volume did grow 3.5× from 2022 to 2024. That is a real slope on a very small base — the kind of curve that looks like either a beachhead or a plateau depending entirely on which two years you draw through.