Medicare’s FY2027 rule set five per-case add-on payments for clinical AI. Every one of them is exactly 65% of the cost the applicant claimed. Here is the whole price list in one chart.
A new technology add-on payment is the closest thing clinical AI has to a reimbursement identity. Last year one AI tool had one. For FY2027, five do. The amounts range across a factor of 37, and the ranking has nothing to do with how much any of them changes a decision.
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 for the case. For all five of these, the binding constraint was the first one. So the payment is:
Raise the claimed cost, raise the payment. InVision’s Precision Cardiac Amyloid was proposed at $162.50 and finalized at $2,275 — exactly 14 times higher — inside one rule cycle, because the finalized cost-per-case landed at $3,500 instead of $250.
Amol Navathe, co-author of a January Health Affairs piece on AI payment design, named the mechanism in a Penn LDI interview: 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.”
There is exactly one published look at uptake of an AI NTAP. Viz.ai’s large-vessel-occlusion detector got the first one, effective FY2021, at up to $1,040 a use. Researchers at the Neiman Health Policy Institute went and counted.
Each dot below is one of the 2,116 Medicare ischemic-stroke episodes in the study, across 1,076 facilities, October 2020 through December 2023.
Five years after the first AI add-on payment existed, 14.8% is what adoption looks like. The more interesting question is what predicted the other 85%.
Adoption followed institutional resources. Not patient need.
Comprehensive stroke center status: OR 1.5. Living in the Stroke Belt: OR 2.0. Being treated in 2022 rather than 2020: OR 6.0. Hospitals serving the most socioeconomically deprived areas were significantly less likely to bill it. Stroke severity: nothing. Age, sex, race and ethnicity: nothing.
The uptake study is a 5% sample. 2,116 episodes is not many, the confidence intervals are wide (the 2022 effect runs from 2.7 to 13.3), and the 14.8% headline comes from the institute’s press release rather than the abstract. The abstract itself reports the peak: 21% in 2022. The full text is paywalled and the facility-level data were never released.
One tool, one condition, one era. It measures billed use of LVO-detection software in ischemic stroke between 2020 and 2023. Whether sepsis flagging or delirium monitoring will diffuse the same way is an assumption, not a finding.
Billed ≠ used. NTAP requires a claim line. A hospital running the software and not coding for it looks identical to a hospital that never installed it.
The benchmark bar is a national average. $14,796 is the discharge-weighted mean Medicare payment across all 4,960,325 IPPS discharges in the CY2023 file. Sepsis (DRG 870–872) runs $15,202 and stroke (DRG 061–066) runs $10,528. Medicare Advantage, Medicaid, commercial and critical access hospitals are not in the file at all.
There is no place on the claim for what the model concluded, whether the clinician agreed, or what happened next. NTAP is a temporary payment attached to an input, and the input it is attached to is a vendor’s cost accounting.
If the add-on pays for the finding, what pays for the model that confidently rules out?