Medicare named a payment lane for clinical AI. It hasn’t priced it yet.

The outpatient rule made news for the money — 340B repriced, the conversion factor moved, 638* more procedures walking off the inpatient-only list. That is the half everyone read.

Roughly six hundred pages in, the same rule does something it has never done before. It gives algorithm-driven care a name, a category, and a place in the payment system. It calls it Software as a Medical Service — SaMS. And then it does the thing worth your attention: it says it has not decided how to pay for it, and asks.

What CMS actually proposed

Three moves, in order:

• It named the category. SaMS is now a thing Medicare has a word for. (The parallel legislative concept, “algorithm-based healthcare services,” is what the Health Tech Investment Act calls it — but the term inside this rule is SaMS.)
• It made CY2027 a bridge year. As a first step, SaMS technologies are assigned to New Technology APCs, and things already sitting there generally keep their assignment. Nothing dramatic happens to payment next year.
• It opened the question. CMS says it is developing a comprehensive payment methodology and is soliciting comment on how algorithm-driven services generate value — explicitly flagging clinical outcomes, efficiency gains, and impact on utilization.

A bridge year is not a delay. It is the drafting window.

Read the professional-fee cut and you are reading a decision already made — the anomaly screen fired on claims data, the work RVU came down, and the comment period is about the size of a number that already exists.

SaMS is the opposite situation. The methodology does not exist yet. It gets written out of what arrives before the window closes. That is a materially different kind of comment opportunity, and it does not come around often: a payment category at the moment it is still a blank page.

The tell is somewhere else in the rule

If you want to know where this lands, do not read the SaMS section. Read the program-integrity section.

Starting January 1, 2028, under Section 6225 of the Consolidated Appropriations Act, 2026, outpatient payment for an off-campus department is conditioned on that department billing under its own NPI and the main provider filing recurring provider-based attestations. Add the expansion of outpatient prior authorization on top.

None of that is a payment change. It is plumbing — the machinery to see, audit, and price who is accountable for what. CMS is spending this rule building the ability to attach a name to a claim. Apply that same logic to software and the question stops being “is the algorithm any good” and becomes “who signed it.” Infrastructure is the tell for where policy goes next.

What the evidence says about where the machine is weak

This matters for how the methodology should be built, and it is the part the AI conversation usually gets backwards.

Stanford’s MedHELM evaluation (published in Nature Medicine, 2025) ran nine leading models across 35 medical tasks, with 29 physicians across 14 specialties designing and validating the tasks. The headline was that the best performer took a 66% win rate. The useful finding is underneath it.

Broken out by task type, the models scored highest on clinical note generation (0.73–0.85) and patient communication (0.78–0.83) — and lowest on clinical decision support (0.56–0.72) and administration and workflow (0.53–0.63).

Read that as an operator. The machine is strongest at the documentation layer, where the stakes are lowest, and weakest at judgment, where the stakes are highest. A payment methodology that prices autonomous algorithmic decisions is pricing the weakest measured capability in the stack. A methodology that pays for an attested output — a named clinician who reviewed it and stands behind it — is pricing the thing that actually carries the risk.

What an operator does in the next sixty days

• Comment, if you build or buy clinical AI. OPPS/ASC (CMS-1850-P) closes August 31, 2026. The SaMS methodology is being written from this docket. This is the cheapest leverage in the rule.
• Instrument what CMS said it will measure. Outcomes, efficiency, avoided utilization. The center already capturing PROs and complications is already generating the evidence the methodology is going to ask for — the same data that decides your ASM score.
• Make attestation a record, not a habit. Which clinician reviewed it, which model and version produced it, what it acted on. If payment ever attaches to an attested output, the attestation is the claim, and a claim you cannot evidence is a claim you cannot keep.
• Know which side of the line you are on. The professional fee is being trimmed while the facility side takes an inflation-linked update. SaMS is the first genuinely new lane in years, and it is unpriced. That is a strategic position, not a technology purchase.

None of this is a prediction that CMS pays generously for software. The bridge year could easily become another bridge year. But the category now exists, the methodology is open, and the agency is on record asking how to value it. The centers that answer with data will be the ones the answer gets built around — the same way outcomes stopped being registry hygiene and became the currency.

SaMS, New Technology APC treatment, and the Section 6225 attestation requirements are from the CMS proposed rule (CMS-1850-P, July 2026) and may change in the final rule. Figures marked * are illustrative or program-dependent. MedHELM figures are from the published Stanford evaluation (Nature Medicine, 2025). Not clinical, legal, or financial advice. MSKvalue assembles independent best-of-class products with the American Joint Replacement Registry as the outcomes backbone.