CMS Prior Authorization Metrics Reporting: What AI Decision Evidence Health Plans Should Keep
Health plans should keep a per-interaction record for every AI step in a prior authorization case: a case identifier, timestamp, AI function, policy applied, action taken, human approval status and data handling outcome. The CMS rule requires aggregate public metrics, not these fields. The record is what lets a plan trace each counted request to its outcome, show where AI took part, and answer a regulator or auditor. Set retention as a documented policy aligned with the HIPAA six-year documentation period.
A minimum viable evidence record for AI-assisted interactions
- Case or request identifier: the key that links the interaction to the case management outcome and to the counted metric.
- Timestamp: when the interaction occurred, in a consistent time standard across systems.
- AI function: intake, triage, clinical summarization or recommendation.
- Policy applied: which rule governed the interaction, so the plan can show what the AI was permitted to do.
- Action taken: allow, redact, block or escalate for human approval.
- Human approval status and data handling outcome: reviewer identity and role or an automated-only flag, plus whether PHI was touched, redacted or blocked.
What the CMS rule requires and what it leaves to the plan
The 2024 CMS Interoperability and Prior Authorization final rule requires impacted payers to publicly report certain prior authorization metrics, aggregated for all non-drug items and services. CMS describes an annual March 31 deadline for posting them. The rule also requires a Prior Authorization API, with compliance dates beginning January 1, 2027 for Medicare Advantage organizations and state Medicaid and CHIP fee-for-service programs. Dates for other payer types depend on rating periods or plan years beginning on or after that date.
For qualified health plan issuers on the federally-facilitated exchanges, a denial response to the provider (excluding drug requests) must include a specific reason for the denial from January 1, 2026.
None of this prescribes AI-level records. The rule does not tell a plan which fields to keep for an AI intake or summarization step. During the 2024 rulemaking a commenter suggested capturing status-change timestamps, how a decision was made, software or AI tools used, and persons involved. That is a useful design signal, not a requirement. The evidence set described here is a plan's own defensible choice, and should be documented that way.
As of November 2026, CMS has proposed amendments to the non-drug metrics, annual public reporting of drug metrics, and reporting of API usage. These remain proposals. Evidence design should leave room for them.
Work backward from the reported metrics
Start with what the plan must publish: the percentage of requests approved, denied and approved after appeal. The proposal would add numeric counts alongside percentages for certain existing metrics and a drug versus non-drug split. Each figure is an aggregate. An auditor who tests it will sample a counted request and ask how it reached its outcome.
That question has three parts. What decision was recorded and when? Which steps in the workflow involved an AI system? Was a person accountable for the outcome? The case management system answers the first. It usually cannot answer the second or third, because AI steps run in vendor tools, orchestration layers and application services that keep their own records in their own formats.
The practical design test is traceability in one direction: from any counted request, a compliance analyst should be able to retrieve every AI interaction attached to it and see what each was allowed to do. If that requires manual correlation across three systems, the evidence is not yet defensible. Assign a named owner for reconciling reported metrics against underlying decision records, since that work crosses team boundaries.
Runtime-enforced evidence compared with reconstructed logs
The difference is when and where the record is created. Evidence produced at the point of policy enforcement captures the decision as it happens. Reconstruction assembles an account afterward from records that were not designed to be read together. This is an architectural argument about defensibility, not a regulatory finding.
- Policy applied: Runtime: recorded with each interaction. Reconstructed: inferred from configuration that may have changed since.
- Action taken: Runtime: the enforced outcome is the record. Reconstructed: derived from application behavior and vendor output.
- Human approval: Runtime: escalation and approval are recorded when they occur. Reconstructed: matched from case notes and user activity.
- PHI handling: Runtime: redaction or blocking is logged as an outcome. Reconstructed: assumed from data flow diagrams.
- Completeness: Runtime: one consistent schema. Reconstructed: depends on each vendor tool retaining and exporting data.
HIPAA audit controls, retention and where evidence lives
AI interactions that process PHI in prior authorization workflows fall under the HIPAA Security Rule like any other electronic information system. Audit controls under 164.312(b) are the expectation that a plan can record and examine activity in systems that handle ePHI. The fields above map to that expectation: who or what acted, on which case, under which policy, with what result for PHI.
The January 2025 proposed Security Rule update would add 164.312(d)(2)(iii), requiring technical controls to retain records of all activity in relevant electronic information systems, as determined by the entity's information system activity review policies. It is a proposal. Its discussion also states that documentation must be retained for the later of six years from creation or the date it was last effective.
Treat AI evidence retention as a documented policy decision tied to that six-year text, and confirm the operative regulation and any final rule with counsel before fixing periods. Reference the permitted and human-approval-required AI functions in the policies your information system activity review already uses.
Three deployment questions follow. Can retention be configured to your policy rather than a vendor default? Where is the evidence stored, and can it stay in a region you specify? Can the evidence store run inside your own cloud, so PHI-adjacent records are not held in a third party's environment? Answer these before selecting tooling, because evidence that sits in a vendor tool is evidence you may not control for six years.
Where runtime enforcement produces the evidence
Trussed enforces policy in the execution path of AI interactions and tool calls. Each interaction can be allowed, redacted, blocked or escalated for human approval, and that enforcement generates audit-ready evidence mapped to HIPAA. The record is a product of the decision being enforced at runtime, not reviewed after the fact. Trussed runs in production with healthcare and pharma organizations.
For a prior authorization workflow, this supplies the policy applied, action taken, timestamp, human approval status and PHI handling fields at the interaction level. Trussed supports self-managed deployment in the customer cloud, with configurable retention and data residency, which addresses the retention and storage questions above.
The scope is individual interactions and tool calls. Trussed does not provide end-to-end tracing of an agentic workflow or model lineage. The case identifier must be carried by the calling workflow, and linking interaction records to approval, denial and appeal counts still happens against the case management system. Plan for that join explicitly in your evidence design.
From reported metric to interaction record
Public metric
Approved, denied and approved-after-appeal figures posted annually.
Case outcome
Decision held in the case management system.
AI interaction
Intake, triage, summarization or recommendation step tied to the case.
Enforced policy
What the AI was permitted to do, and the action taken.
Human review
Reviewer identity and role, or an automated-only flag.
Define your evidence record before the next reporting cycle
Review how policy enforcement in the execution path produces interaction-level evidence for AI-assisted prior authorization, and how it fits your retention and residency requirements.
See How Evidence Is Produced