CMS-0057-F Prior Authorization API Requirements: Where Agent Controls Fit
CMS-0057-F defines the Prior Authorization API, its data exchange, denial reasons and timing. It does not mandate controls on AI agents. Agent-level runtime policy enforcement is an architectural and risk-management choice that applies to individual model calls, tool calls, MCP server access and PHI exposure around the API. The payer's API obligations remain separate from that layer.
Interaction points where runtime enforcement applies
- 1
Model calls
Prompts may carry PHI, clinical documentation or payer identifiers. Policy can allow the call, redact fields before it leaves the boundary, or block it.
- 2
Tool calls to FHIR endpoints
Reads against documentation requirement lookups and writes that submit requests or retrieve status and decisions. Policy can differ per endpoint and per read or write.
- 3
MCP server access
Which servers an agent may reach, and which tools on them it may invoke, can be enforced at runtime rather than assumed from configuration.
- 4
PHI in arguments and responses
Tool arguments and returned payloads can be inspected, redacted or blocked, with escalation for human approval on higher-risk actions.
Evidence an enforcement layer should produce
- The policy in force at the time of each interaction
- The decision made: allow, redact, block or escalate for human approval
- The action actually taken, including what was redacted and what reached the model or endpoint
- The identity of the human approver and the outcome for escalated actions
- Retention and data residency you can configure to your own requirements
- Mapping of evidence to the frameworks your auditors use, such as HIPAA
What CMS-0057-F requires of the Prior Authorization API
The prior authorization provisions of CMS-0057-F cover payer API implementation, responses to requests within set timeframes, a specific reason for denials, and public reporting on approvals, denials and appeals. The requirements are codified across 42 CFR 422.122, 431.80, 457.732 and 45 CFR 156.223.
The API must be populated with the payer's list of covered items and services that require prior authorization (excluding drugs) and must be able to identify the documentation the payer requires for approval. It must support a HIPAA-compliant prior authorization request and response as described in 45 CFR part 162, and it must communicate approval, with the date or circumstance under which the authorization ends, or denial with a specific reason. For Medicare Advantage, the API must also conform to the cited standards in 45 CFR 170.215. The eCFR lists the HL7 FHIR Da Vinci Prior Authorization Support implementation guide, version 2.2.1-STU 2.2, as the specification for prior authorization submission. HIPAA transaction conformance therefore sits alongside the FHIR API rather than being replaced by it.
As of November 2026, the dates are as follows. A specific denial reason is required from January 1, 2026 for MA organizations, excluding drugs, and the response must follow the timeframes in 42 CFR 422.568(b)(1) and 422.572(a)(1). The API is due by January 1, 2027 for MA organizations and state Medicaid and CHIP fee-for-service programs. For Medicaid managed care plans and CHIP managed care entities it is due by the rating period beginning on or after January 1, 2027. For QHP issuers on the federally-facilitated exchanges it is due for plan years beginning on or after January 1, 2027. Prior authorization information must be shared while the authorization is active and for one year after the last status change.
What the rule leaves to the implementing organization
The rule specifies API content, standards and timing. It does not specify how an automated tool or AI agent is constrained or evidenced while it works inside those workflows. That is the gap architects are being asked to fill.
CMS does say providers can use the API to learn whether a payer requires prior authorization for an item or service and to query documentation requirements from the provider's own system, and that this could facilitate automated compilation of the information needed for a request. That is the clearest point at which automated tooling, including an agent, could read from the API. Beyond that, any connection to agent runtime control is an architectural inference, not a regulatory requirement.
The scope split also depends on organizational role. The API obligations bind payers. A provider-side or intermediary agent that calls a payer API is not itself subject to those obligations on this evidence. A payer's agent that drafts or processes status responses and denial reasons operates inside a regulated process, but accountability for the decision and the reason content stays with the payer. Whether to escalate to a human is an organizational choice, not a rule requirement.
Designing policy around the API, not inside it
Start by identifying your role (MA organization, state agency, Medicaid or CHIP managed care entity, QHP issuer or provider) and the matching 2026 and 2027 dates. Then inventory every endpoint an agent can reach: the Prior Authorization API, the documentation requirement lookup, and status and decision responses. Record which are read-only and which write.
Define policy per tool and per endpoint. A documentation lookup is largely a read of payer rules and may need little more than allow. A submission carries patient data and warrants field-level redaction rules and, where your risk appetite requires it, human approval before the call executes. Redaction and blocking are different outcomes. An agent that needs PHI to assemble a request will fail if the data is blocked, so decide in advance which fields a step genuinely needs and which can be masked.
Treat endpoint conformance as a separate workstream. Payload content, standards, denial reasons and response timing are properties of the payer's API implementation. An enforcement layer can validate that a structured output such as a denial reason is present before an agent passes it on, but the rule does not require that check. Keep drug-related prior authorization out of current API scope. A proposed rule published April 14, 2026 would add drugs to the API and require a specific denial reason for drugs from October 1, 2027, but it is a proposal and changes scope only if finalized.
Limits of interaction-level enforcement
Enforcement in the execution path applies to individual interactions and tool calls. It does not provide transaction-level governance across tasks and agents. A prior authorization often spans several steps: documentation lookup, assembly, submission, status polling and handling of a decision. Nothing at the interaction level ties those steps into one governed business transaction or says which agent owned which outcome. That needs separate controls and a named owner.
Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with decisions to allow, redact, block, rate limit or escalate for human approval. It generates audit-ready evidence from that enforcement, mapped to HIPAA among other frameworks, and can be deployed self-managed in the customer's cloud. Trussed makes no claim of CMS-0057-F certification or conformance. The API obligations remain the payer's to meet.
Where agent controls meet the Prior Authorization API
Rule defines
API content, standards, data exchange, denial reasons and timing
Organization decides
How agents are constrained and evidenced at runtime
Enforcement applies to
Model calls, tool calls, MCP server access and PHI in transit
Evidence shows
What the agent was permitted to do and what it did
Map your agent touchpoints before they reach production
Review how runtime enforcement and audit-ready evidence apply to the model calls, tool calls and MCP access in your prior authorization workflows.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.