The Peterson Health Technology Institute needed one example of autonomous clinical AI for its July report. It picked managing routine prescription refills [PHTI, 2026].
The same month, CMS titled the new software payment section of its outpatient rule “OPPS Payments for SaMS Diagnostic Services” [CMS, 2026].
Nothing in that section pays for a refill.
The Problem
Pharmacy has been making the same AI business case for three years. Fewer callbacks, faster verification, a rejected claim caught before it bounces, an adverse event that never happens. It is a cost-avoidance argument, and it lives exactly as long as the operating budget that funds it, which is why working pharmacy AI keeps losing the budget fight.
CMS is now changing where AI money comes from.
On July 7, 2026, CMS published the CY 2027 hospital outpatient proposed rule (CMS-1850-P, 91 FR 41734), creating a standalone Medicare payment category for clinical software called Software as a Medical Service. Comments close August 31, 2026 [CMS, 2026]. Nine days later the agency published the 716-page CY 2027 Physician Fee Schedule proposed rule (CMS-1848-P, 91 FR 43842), which adopts the same term and carries its own proposal on SaMS laboratory analyses. Those comments close September 14, 2026.
This is the first time Medicare treats clinical software as a service in its own right rather than a cost buried inside somebody else’s procedure. It matters less than the headline suggests, and the reason sits in the heading CMS chose. Diagnostic Services. The SaMS procedures the outpatient rule names as already assigned to New Technology APCs are Atherosclerosis Imaging-Quantitative Computed Tomography, LiverMultiScan, Optellum Lung Cancer Prediction, quantitative magnetic resonance for tissue composition, quantitative MR cholangiopancreatography, and Fibresolve [CMS, 2026]. Six products. Every one of them analyzes an image or a tissue sample.
The Insight
PHTI published its answer to this question the same month, and it reads like a warning aimed at exactly what CMS is drafting.
The report came out of a May 2026 workshop held under the Chatham House Rule with leaders from health systems, health plans, technology developers, investment firms, academia, and federal agencies. Its conclusion on billing codes is blunt: creating new codes for autonomous AI “would likely amplify current fee-for-service billing incentives without a link to outcomes and a low price” [PHTI, 2026]. The mechanism is not the obvious one. The problem is not that AI costs too much. It is that “if AI reduces the amount of time clinicians spend per encounter without a proportional decline in reimbursement rates, providers can see more patients with less effort while receiving the same payment per service, driving up aggregate spending” [PHTI, 2026].
Both institutions are right. Clinical software did need a payment category of its own, and starting where the FDA clearances and the outcome evidence already sit is defensible. Widening that category into per-use codes for autonomous AI would also, on PHTI’s reasoning, inflate spending. What falls out of two correct positions is a two-class system that nobody argued for.
One workshop participant put the clinical version of the problem in a sentence: “I hope we move beyond assistive AI because it just ends up dropping more work on clinicians. It is tough to manage remote data and recommendations between visits, especially when the extra time could be used for a higher-impact (and billable) patient visit” [PHTI, 2026]. That is a clinician describing how a payment rule reroutes their attention, which is what payment rules do.
“Medicare is about to split hospital AI into two classes: the kind that bills and the kind that begs for budget.”
PHTI names one payment model that genuinely fits AI-enabled care, and traditional health systems are structurally locked out of it. The CMS Innovation Center’s ACCESS model launched in July 2026 with more than 150 participating organizations on a ten-year outcome-based structure: half the payment upfront, the remainder withheld until outcomes land. Health plans covering 165 million Americans have pledged to align with an ACCESS-like design by 2028. But ACCESS participants cannot receive Medicare payments for their aligned beneficiaries outside the model, which is why PHTI expects it to “accelerate AI adoption among new market entrants operating outside the traditional delivery system” [PHTI, 2026]. TheraIntel covered the application window when it opened. The exclusion clause got less attention than the deadline did.
So the pharmacy director’s options in 2028 are a SaMS code, which requires the tool to read a scan; an outcome-based model that costs the system fee-for-service revenue on every patient it enrolls; or the operating budget.
The operating budget is shrinking in the same rulemaking cycle. CMS proposes a CY 2027 conversion factor of $33.17 for qualifying APM participants, down 1.19%, and $32.84 for everyone else, down 1.68% [CMS, 2026].
Then there is the detail that should bother anyone running clinical pharmacy. PHTI’s own exemplar of promising AI-enabled care is Mass General Brigham’s remote hypertension program: home blood pressure readings, algorithm-driven care pathways, and nonlicensed patient navigators working under pharmacist oversight. PHTI reports that scaling it proved difficult because of the investment required and “the lack of a sustainable reimbursement model” [PHTI, 2026]. The clinical design works. Fewer than one in four American adults with hypertension have their blood pressure controlled. And the payment path for the model that would move that number appears in neither proposed rule.
In Practice
| What the AI does | Proposed 2027 payment path | What that covers |
|---|---|---|
| Analyzes an image or tissue sample | SaMS category, separate outpatient payment | AI-QCT, Optellum, LiverMultiScan, Fibresolve |
| Manages a condition to a measured outcome | ACCESS-style outcome payment, no other Medicare billing for those patients | Mostly new entrants outside the delivery system |
| Refills, verification, documentation, care coordination | None proposed | Most pharmacy and nursing AI |
The Bottom Line
Both rules are proposals, and neither comment window is generous. August 31 on the outpatient rule, September 14 on the fee schedule. Both close before most health systems finish their 2027 capital plans, which means the people who will live with this framework are unlikely to have said anything about it.
The verdict is uncomfortable precisely because neither party is behaving badly. CMS built a narrow payment category anchored to services that already carry FDA clearances and established codes. PHTI documented why widening that category carelessly would raise national spending instead of lowering it. Put the two together and the largest body of working clinical AI in American hospitals, the kind that manages medications, closes care gaps, and never produces a scan, stays on the expense line, funded by a conversion factor CMS proposes to cut.
Whether it hardens that way is genuinely unsettled. Both documents are proposals, the final rules land later this year, and Medicare has widened narrow categories before. What decides it is the comment record, and the comment record on a payment rule is written mostly by the parties with a code at stake. Pharmacy does not have one.
Radiology’s AI graduates to a billing code this year. Pharmacy’s AI is still arguing for its budget, and its strongest evidence remains an invoice that never arrived.