Medicare just set five per-case add-on payments for clinical AI, from $61.84 to $2,275. Here is what each one is actually worth, laid over 2,259 real U.S. hospitals and the Medicare cases they already bill.
The FY2027 inpatient rule approved 19 new technology add-on payments. Five are AI or algorithm-based — up from one the year before. The reimbursement identity for clinical AI now exists.
It is worth being precise about what the number means. NTAP pays the lesser of 65% of the average cost of the technology or 65% of the cost in excess of the MS-DRG payment. In practice, for all five of these, the finalized amount is exactly 65% of the cost the applicant claimed. The price is set by what the vendor says it spends, not by what the tool changes.
Every IPPS hospital with at least 60 Medicare sepsis or stroke discharges in CY2023, positioned by how many eligible cases it bills and by what the add-on is worth as a share of the Medicare payment the case already generates. Circle area is the modeled annual add-on revenue.
Volume: r = 1.00, by construction. The add-on is a fixed dollar amount times a count of uses. Nothing in the formula reads a patient, an outcome, or a decision. Hold the adoption rate constant and the map of who gets paid is a map of who is big.
Switch Who adopts to the AJNR-observed pattern and the picture tilts further. Across 2,116 Medicare ischemic-stroke episodes at 1,076 facilities — the only published look at uptake of an AI NTAP — billed use ran at 14.8% of eligible cases and was higher at comprehensive stroke centers (OR 1.5, CI 1.1–2.0) and in the Stroke Belt (OR 2.0, CI 1.3–3.0). Hospitals serving the most socioeconomically deprived areas were significantly less likely to bill it. There was no association with stroke severity, age, sex or race/ethnicity.
MODEL, NOT MEASUREMENT The AJNR pattern here is a reconstruction: we apply those two odds ratios to each hospital (metropolitan location as a stand-in for comprehensive stroke center status, home state for the Stroke Belt) and re-calibrate so the case-weighted mean still equals your slider. It shows the shape the published data implies. It is not the study’s facility-level output, which was never released.
Half the patients are missing. This is Medicare fee-for-service Part A only. Medicare Advantage, Medicaid, commercial and self-pay are not in the file, and MA now covers over half of Medicare enrollees. Critical access hospitals are excluded outright — which is most of rural America.
The case base is a stand-in, not an eligibility rule. Each NTAP has its own coverage criteria. Aidoc’s BriefCase-Triage is for abdominopelvic CT; Ceribell’s delirium monitor only counts for patients 65 and over. We hold the denominator fixed at real sepsis or stroke volume so you can compare the rates against real Medicare workload. Do not read a dot as a revenue forecast.
CMS suppresses small cells. Hospital/DRG combinations with too few discharges are absent rather than zero, so low-volume hospitals are systematically under-represented. Drag Min eligible cases upward and watch the non-metro share fall — some of that is real concentration, and some of it is the file.
2023 volume, 2027 prices. The discharge counts are the most recent CMS vintage available; the payment rates take effect October 1, 2026.
Viz.ai’s ContaCT got the first AI NTAP, effective FY2021, at up to $1,040 a use. Five years and one published uptake study later, the open question is still whether the scans it triggered changed management. NTAP is a temporary payment attached to an input. There is no field on the claim for what the model concluded, no field for whether the clinician agreed, and no field for what happened next.
If the add-on pays for the finding, what pays for the model that confidently rules out?
Health Affairs authors Sita Kottilil and Amol Navathe put the mechanism plainly in a Penn LDI interview about their January paper: NTAP’s requirement that a technology demonstrate substantial additional cost “may incentivize developers to set high prices,” and paying for AI on labor-style inputs “may lead to overspending and overuse.” InVision’s figure moved from $162.50 in the proposed rule to $2,275 in the final one — exactly 14x — inside a single rule cycle. That is what an input-priced pathway looks like when it is working as designed.