Medicare Advantage prior auth sits inside the National Coverage Determinations (NCDs), Local Coverage Determinations (LCDs) and the plan's own utilisation management (UM) criteria, with CMS marketing and operational rules layered on top. The decision space is narrower than commercial UM and more procedure-specific, so the practitioner read is closer to a CMS coverage guide than to a generic claims AI pattern. The reviewer is operating against a hierarchy of authority, not a single rulebook, and the copilot has to model that hierarchy explicitly.
CMS-0057-F changes the operating model. The final rule was published in the Federal Register on 8 February 2024. From 1 January 2026, MA organisations must meet 7-calendar-day standard and 72-hour expedited decision timeframes, provide a specific reason for any denial, and publicly report aggregate prior authorisation metrics. From 1 January 2027, the four FHIR-based APIs (Patient Access, Provider Access, Payer-to-Payer, Prior Authorization) must be operational. The published rule on federalregister.gov and the CMS press release on cms.gov are the primary references.
The third reason MA is its own problem is fairness. Section 1557 of the Affordable Care Act, as the HHS-OCR 2024 final rule made explicit, extends non-discrimination expectations to algorithms and AI tools used by covered entities in patient care. An MA plan using an AI assistive tool for medical-necessity review now operates inside HHS-OCR's lens as well as CMS's, which means demographic-stratified evaluation and ongoing monitoring are not optional. The supervising authority is hhs.gov/ocr.
PA requests arrive via portal (Availity, plan portal), EHR fax and eFax, payer-to-EHR submission, and increasingly via FHIR ePA aligned to the HL7 Da Vinci PARDD implementation guide. The first AI value is structured extraction of CPT and HCPCS codes, ICD-10 diagnoses, supporting clinical documentation and dates of service from mixed inputs into a single normalised record.
Plan benefits, network status, coverage limits, accumulator status and any prior approvals already on file. This is deterministic lookup, not AI-driven decisioning. The output is a clean, auditable coverage statement that the necessity reviewer can rely on without going back to a second source, and that the audit trail can reproduce later.
Where assistive AI most matters. The model retrieves the applicable NCD or LCD and the plan UM criteria, attributes each criterion to the specific clinical evidence in the chart, and produces a draft recommendation with citations. The clinician retains the determination. The model retains the audit trail, with every claim anchored to a source the reviewer can open.
CMS-0057-F requires the specific reason for any non-approval and adherence to the 7-calendar-day standard and 72-hour expedited windows from 2026. The notification must be machine-readable through the Prior Authorization API from 2027. The document state the plan notifies on must match exactly the evidence the reviewer used in the decision.
Appeals are clinician-led and de novo. AI may surface the original criteria, the missing evidence and any new evidence supplied in the appeal, but cannot perform the appeals decision itself. Annual public reporting of PA metrics under CMS-0057-F will include appeal overturn rates from 2026, so the appeals lane is now an externally visible quality signal.
CMS-0057-F requires MA payers to publicly post aggregate PA metrics from 2026: approvals, denials, appeals, average response time. The reporting cadence makes the AI's contribution measurable in public and shifts the conversation from anecdote to numbers. The copilot's evaluation harness must reconcile to the same numbers the plan publishes.
“The MA prior auth conversation has been about turnaround time for a decade. From 2026 it will be about turnaround time, public metrics and fairness, at the same time, and AI that cannot defend its reasoning under all three lenses is a liability, not a tool.”
No. The rule applies to Medicare Advantage, Medicaid FFS and managed care, CHIP managed care, and Federally-facilitated Exchange Qualified Health Plans. ERISA-governed commercial group plans are out of scope.
The 7-calendar-day standard and 72-hour expedited PA decision windows take effect, specific reasons for any denial must be provided to the requester, and aggregate PA metrics (approvals, denials, appeals, average response time) must be publicly reported.
The four FHIR-based APIs (Patient Access, Provider Access, Payer-to-Payer, Prior Authorization) must be operational. The HL7 Da Vinci PARDD implementation guide is the practical reference for the Prior Authorization API.
No. CMS expects a qualified clinician to issue any non-approval. The AI assists with criterion retrieval, clinical-evidence attribution and draft reasoning, not with the determination itself, and the audit trail must reflect that.
The 2024 HHS-OCR final rule extends non-discrimination duties to algorithms and AI tools used in patient care, which means MA plans using AI in necessity review must run pre-deployment and ongoing fairness evaluation, with documented methodology.
8 to 16 weeks to a production-ready PA copilot with criterion attribution, an evaluation harness tuned to MA scenarios, fairness evaluation aligned to Section 1557, FHIR readiness for 2027, and a full audit pack.
CMS-0057-F, finalised on 17 January 2024 and published in the Federal Register on 8 February 2024, resets Medicare Advantage prior authorisation. From 1 January 2026, MA payers must meet 7-calendar-day standard and 72-hour expedited decision windows, give specific denial reasons, and publicly report aggregate PA metrics. From 1 January 2027, four FHIR-based APIs must be operational. AI assists with criterion retrieval and evidence attribution, never with the determination, and operates under Section 1557 fairness expectations across an EDI 278 and FHIR transition. Moweb delivers a production copilot in 8 to 16 weeks on a fixed fee.