Reporting CMS-0057-F Prior Authorization Metrics from Agent Runtime Logs
CMS-0057-F prior authorization metrics reporting is an evidence-lineage problem. As of November 2026, impacted payers publicly report percentage and time-based metrics for non-drug items and services each March 31. Agent runtime logs can support those metrics with request identifiers, timestamps, policy decisions and human escalation outcomes. Final determinations, appeal results and request classification must still come from the payer system of record, and every published figure should trace back to per-request records.
Gap assessment for existing runtime logs
- Does every event carry a stable request identifier that joins to the determination and appeal records?
- Is there a received timestamp that can be compared with the system-of-record submission time?
- Is the standard or expedited designation captured at intake, and are changes recorded?
- Are extension events recorded with timestamps?
- Are human escalation outcomes recorded, with actor type separating agent actions from human decisions?
- Can a final determination be linked to the runtime events for the same request?
What the rule requires you to report
CMS-0057-F requires impacted payers to publicly report certain metrics about their prior authorization processes. Reporting is annual, due by March 31 for the previous calendar year, and begins in 2026. Drug data is excluded. The codified requirements sit at 42 CFR 422.122(c) for Medicare Advantage, 438.210(f) for Medicaid managed care and 457.732(c) for CHIP FFS and State programs. The reporting level differs by payer type: contract level for MA, plan level for Medicaid managed care, and State level for States.
The State text at 457.732(c) shows the metric structure. It lists all items and services requiring prior authorization. It then gives the percentage of standard requests approved, denied and approved after appeal, the percentage of extended-review requests that were approved, the percentage of expedited requests approved and denied, and the average and median time from submission to determination for standard and expedited requests. All of these are aggregated across all items and services. Confirm the exact metric list for your payer type against the full regulation text.
Two points shape the design. The metrics apply regardless of how the payer received the request, so agent-handled, human-handled and API-submitted requests all fall in the same populations. Reported figures cannot be limited to agent-touched requests. Also, current text specifies percentages, not counts. Numeric counts appear only in an April 2026 proposed rule.
Mapping each metric to runtime evidence and system-of-record data
Start from the metric definition and work back to the source. The table below shows, for each metric group, what agent runtime logs can contribute and what the payer system of record must supply. The runtime fields are design choices, not regulatory data elements.
- Standard requests: approved, denied, approved after appeal: Runtime: request identifier, intake and handling events, actor type. System of record: final determination and appeal outcome.
- Extended-review requests that were approved: Runtime: extension-related events, if captured. System of record: extension status and final outcome.
- Expedited requests: approved and denied: Runtime: expedited designation as seen at intake. System of record: confirmed classification and outcome.
- Average and median time to determination: Runtime: intake and handling timestamps. System of record: authoritative submission and determination times.
- Drug exclusion: Runtime: none. System of record: coverage or benefit data that classifies the item or service.
- List of items and services requiring prior authorization: Runtime: none. Source: payer coverage and policy records.
A three-layer reporting data model
Separate the data into three layers and document the transformation between each. The first layer is event-level runtime evidence: intake, document retrieval, tool calls, policy decisions and human escalation outcomes, each stamped with a request identifier, timestamp and actor type. Actor type is what keeps agent-assisted actions distinguishable from human decisions.
The second layer is request-level facts. Here each request carries its standard or expedited designation, item or service versus drug classification, extension status, determination, determination time and appeal outcome. Most of these values come from the system of record. The runtime layer joins to them through the shared request identifier and shows how the request was handled along the way.
The third layer is period-level aggregates: the percentages and average and median times for the reporting year, at the contract, plan or State level. Because aggregates cover every request regardless of channel, the request-level layer must be complete and deduplicated. Before publication, reconcile log-derived counts with system-of-record counts and record any differences and how they were resolved.
Policy enforcement as a source of audit-ready evidence
Evidence is more reliable when it is produced at the moment an action is taken than when it is reconstructed from application logs afterward. Trussed enforces policy in the execution path for individual interactions and tool calls: allow, redact, block, or escalate for human approval. Each enforcement decision produces audit-ready evidence, such as which tool call was attempted, which policy applied, what the outcome was and whether a human approved it. For a prior authorization agent, that supports the second layer of the model by showing what an agent was permitted to do at intake or document retrieval, and where a person took over.
The scope is specific. This evidence covers individual interactions and tool calls. It does not replace the system of record, and it is not a reporting tool. It does not establish a determination, an appeal result or a request classification. Auditors should see this boundary documented, so each reported metric is clearly attributed to runtime evidence or to system-of-record data.
Integrity and retention need explicit decisions. Retention periods for runtime evidence should be set by payer policy and counsel, and should be long enough to reproduce reported figures. The one-year accessibility period in the CHIP FFS rule applies to API data availability, not to log retention. Trussed supports self-managed deployment in the customer cloud with configurable retention and data residency, which lets the payer set these terms.
Ownership, interpretation and proposed changes
Assign an owner to each metric definition and record every interpretation, such as how extended timeframes or appeals are counted. Where captured, fields such as AI tool used or persons involved should be labeled internal governance evidence, not CMS-mandated data elements.
Track the April 2026 proposed rule, but do not build to it as if it were final. As of November 2026 it is proposed only. It would add numeric counts alongside percentages for certain existing metrics, add drug prior authorization metrics, and add Provider Access, Payer-to-Payer and Prior Authorization API usage metrics, with proposed compliance in 2028 for 2027 data. A model that already separates event, request and period layers can add fields at the request layer without redesigning the aggregates. Numeric counts would also raise the importance of complete, deduplicated request populations.
Evidence lineage for a reported metric
Event-level evidence
Runtime events and tool calls tied to a request identifier.
Request-level facts
Classification, outcome and timestamps reconciled with the system of record.
Period-level aggregates
Percentages and average and median times for the calendar year.
Published metric
Reproducible from retained records, with documented transformations.
Review the evidence behind your prior authorization metrics
Talk through how policy enforcement in the execution path produces audit-ready evidence for agent interactions and tool calls, and where your systems of record take over.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.