Controlling Provider AI Agents That Submit Prior Authorization Requests Through Payer FHIR APIs
Treat the agent as a set of tool calls against payer FHIR endpoints, and evaluate policy on each call before it reaches the payer. Enforce allow, redact, block, rate limit and human approval decisions in the execution path, keep credentials out of the model's reach, and record each decision as evidence. This is the practical way to show what the agent was allowed to do and what it did under HIPAA audit controls.
Define the control points between agent and payer
- 1
Credential boundary
The backend services client credential sits in a secret store used by the tool executor. The model never sees it and cannot echo it into a request.
- 2
Tool registry
Each tool the agent may call is declared with its payer, endpoint, FHIR resource types and operations. Anything undeclared is not callable.
- 3
Pre-call policy evaluation
Policy runs on every call and returns allow, redact, block, rate limit or escalate before anything is transmitted.
- 4
Payload inspection
Outbound content is checked for PHI fields the request does not need, and for malformed or unsupported structure.
- 5
Approval queue
Calls matching defined conditions wait for a named human decision, which is recorded.
- 6
Evidence record
The enforcement point writes a record for every decision and outcome.
Why the agent is best modeled as tool calls
A provider AI agent that assembles clinical data and submits prior authorization requests does its consequential work through tools. One tool reads chart data. Another builds a request. Another submits it to a payer. Each tool call can disclose PHI or create a payer-facing transaction, so each is a point where policy can be evaluated.
The external context sets the shape of those calls. The CMS Interoperability and Prior Authorization final rule, published in the Federal Register on February 8, 2024, sets Prior Authorization API compliance dates in 2027. The dates are January 1, 2027 for Medicare Advantage organizations and state Medicaid and CHIP fee-for-service programs. They fall in rating periods or plan years beginning on or after January 1, 2027 for Medicaid managed care plans, CHIP managed care entities and QHP issuers on the federally facilitated exchanges. As of November 2026, CMS communications still refer to the January 1, 2027 deadline.
On the provider side, the ONC certification test method for a provider prior authorization API describes a client that submits requests under the Da Vinci Prior Authorization Support (PAS) implementation guide. It covers registration, authentication and authorization, and the prior authorization transactions. A separate 2026 CMS proposal would extend electronic prior authorization to drugs from October 1, 2027. It is a proposal, not a final rule, and should not be treated as a current obligation.
Where payer-side authorization leaves gaps for an agent
For certification, client authentication and authorization follow the Backend Services section of the HL7 SMART Application Launch implementation guide. This is a system-to-system pattern in which the credential identifies the registered client application, not a human user or a particular patient. The client must also register with the payer's authorization server.
That is appropriate for a conventional integration, but it has a consequence for an autonomous agent. A valid credential tells the payer which application is calling. It does not say whether this agent should be touching this patient, whether a given FHIR resource belongs in this request, or whether this submission needs a person to look at it. Limits at the patient, payer, resource and operation level would need to be enforced by a layer in your execution path, in addition to payer-side authorization. CMS commentary on the final rule makes a related point: FHIR does not itself define security functions, so authentication and access control must be added around it.
The HIPAA technical safeguard for access control also uses the phrase "persons or software programs". An agent is a software program, so it needs defined access rights tied to your 164.308(a)(4) procedures. Register the agent's client application with each payer, and document the registration and authentication details you use.
Specify the policy at each tool call
Write policy as conditions on the tool call, with an explicit action for each outcome. Define the conditions before the agent reaches production and keep policy versioned, so each decision can be tied to the version that made it.
Two design points deserve attention. First, pended authorization responses are handled through client-side subscriptions in the certification criteria, so the agent must handle asynchronous results and not only request and response. Second, run these policies in the test environment you share with your EHR vendor and payer partners, since CMS encourages providers to test FHIR APIs with them. Verify allow, block and escalate outcomes there before production.
- Allow: Permit calls where the payer, endpoint, resource type, operation and patient are all on the agent's approved list for the task.
- Redact: Remove or mask PHI fields the specific request does not need before transmission. Define the permitted data elements per request type.
- Block: Stop calls to unlisted payers, resources or operations, patients outside the current case, and submissions that are malformed or unsupported.
- Rate limit: Cap submissions per payer, patient and time window, and stop resubmission while a pended response is outstanding.
- Escalate for human approval: Route defined conditions to a reviewer, for example by payer, service type or failed completeness checks. Record who decided and when.
Evidence a regulated enterprise needs to produce
The audit controls standard in 45 CFR 164.312(b) requires hardware, software or procedural mechanisms that record and examine activity in information systems that contain or use ePHI. Agent-mediated calls and the enforcement layer should be treated as in scope. The evidence needs to come from the enforcement point itself, because that is the only place that knows what was decided before the call executed.
For each call, capture the agent and client registration used, the payer and endpoint, the operation and resource types, the patient reference, the policy version, the decision, the inputs and outputs as transmitted after redaction, and the approver where one applies. Retention and export need to be planned so the records can answer a HIPAA review or a payer inquiry.
Integrity controls for transmitted ePHI are an addressable specification, so document your implementation decision. The 2025 HIPAA Security Rule proposal would tighten access control language, but it is proposed only and the current regulation text governs as of November 2026.
What interaction-level and tool-call-level enforcement covers, and what it does not
Trussed enforces policy in the execution path at the level of individual interactions and tool calls. For a prior authorization agent, that means allow, redact and block decisions, rate limiting and escalation for human approval are enforced at runtime, not reviewed after the fact. Each decision produces audit-ready evidence mapped to HIPAA, and deployment is self-managed in the customer cloud with configurable retention and data residency. Trussed runs in production with healthcare organizations.
The boundary matters. Call-level enforcement does not trace an entire prior authorization case across multiple tasks or agents, and it does not provide transaction-level governance. It does not judge whether a payer's decision was right, or whether the clinical content of a submission is correct. Those need separate controls, such as clinical review and payer-side processes. Treat call-level enforcement as the layer that limits and records what the agent does at each call, and design the rest of the workflow controls around it.
Control points on a prior authorization tool call
Agent request
The agent proposes a tool call against a payer FHIR endpoint.
Policy decision
Patient, payer, resource and operation are checked before execution.
PHI handling and approval
Fields are redacted, or the call is routed to a human reviewer.
Payer FHIR API
Only permitted, scoped calls are sent.
Evidence record
Policy version, decision and outcome are recorded per call.
Review your agent's control points
Walk through your tool calls, policies and audit evidence with an expert, and see how enforcement and evidence work in the execution path.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.