How to Link a Prior Authorization Decision to the Downstream Claim in Agent Evidence
Carry the authorization reference and the claim reference as separate fields in the request context of every AI interaction and tool call. Enforce policy on those interactions at runtime, write both references into each evidence record, and add the claim reference once the claim exists. Auditors can then retrieve evidence by either key. The source systems stay the system of record for authorization and claim outcomes.
Test retrieval before an auditor does
- Retrieve all evidence by authorization reference alone.
- Retrieve all evidence by claim reference alone, including records that predate the claim.
- Confirm that interactions with only one reference present are recorded correctly, with the other field explicitly null.
- Confirm that each escalated interaction shows the human approval outcome and timestamp.
- Inspect payloads and keys for raw PHI that minimum necessary would not support.
Treat the link as an evidence-design problem
Prior authorization has two parts, a provider request and a payer decision, and CMS describes it as an interactive process rather than a single transmission. The request and the decision are distinct records. The claim is a third record, created in a different system and adjudicated later, usually under its own identifier.
An authorization decision is also not a payment decision. In the Medicare WISeR model, for example, a provisional affirmation is a preliminary finding that a future claim likely meets requirements. Payment still depends on coverage and documentation requirements being met and the claim being billed correctly. Interaction-level evidence therefore cannot prove how a claim was adjudicated.
The achievable goal is narrower and still useful. For a given authorization or claim reference, show what the AI agent was permitted to do and what it did at each interaction and tool call. Trussed enforces policy in the execution path and produces audit-ready evidence for individual interactions and tool calls. It does not provide transaction-level governance across tasks or end-to-end workflow tracing, so the connection between an authorization and a claim comes from keys your teams define and attach.
Define the correlation keys
Start by inventorying which identifiers exist at each point in the workflow and which system issues them. Typical candidates are an authorization reference, a claim identifier and a member or encounter reference. Their formats differ between systems, and the mapping has to come from your own systems rather than a general convention.
Keep the authorization reference and the claim reference as two separate fields. Do not merge them into a composite key. The authorization precedes the claim, so the claim reference does not exist when the authorization-stage interactions occur. The field must be allowed to be empty and filled in later.
Choose an authorization reference that survives status changes. Authorization status, approval or denial dates and end dates all change over time, and the key should not depend on any of them. On the claim side, a 2022 CMS proposal on attachments described a plan-assigned claim control number that links an attachment to a previously submitted claim. That was a proposal, not confirmed current law, but it shows the kind of claim-side key worth checking for in your claims system.
Decide who may attach each key and when. A key written by an arbitrary caller weakens the evidence. Prefer opaque or tokenized references over raw member identifiers, in line with HIPAA minimum necessary.
Pass the keys into the execution path and enforce policy
The references only help if policy enforcement can read them. Treat them as request context supplied by the calling application or orchestrator, not as something inferred from prompt text.
- Attach context at the entry point: The application sets the authorization reference, and the claim reference when it exists, on every model request and tool call it initiates.
- Validate before use: Reject or flag requests where a reference is malformed or issued by an unexpected system.
- Write policy against the context: Rules for interactions touching authorization or claim data can depend on the references, the tool being called and the data involved.
- Apply the action: Allow, redact, block or escalate for human approval, enforced at runtime, not reviewed after the fact.
- Associate the claim later: When the claim is created, record an association between the claim reference and the authorization reference, then carry both forward.
What each evidence record should carry
Each interaction or tool-call record should let an auditor answer three questions without opening the underlying clinical content: what was the agent working on, what was it allowed to do, and what happened. The field set below is a design recommendation, not a regulatory requirement.
At minimum, include the authorization reference, the claim reference (or an explicit null), the policy decision, the action taken, whether redaction was applied, the human approval outcome where escalation occurred, and a timestamp. Add the tool or model identifier so the record shows which capability was invoked.
The late-association step needs its own record. If the claim reference arrives after earlier interactions were recorded, do not rewrite history. Add an association entry that links the two references with a timestamp and the system that supplied it. Retrieval by claim reference can then resolve to the earlier authorization-stage records through that entry.
Trussed generates this evidence from enforcement in the execution path, so the record reflects the decision actually applied rather than a later reconstruction. Mapping to frameworks such as HIPAA, NAIC and NIST is part of how the evidence is organized for review, but the correlation fields are the part you design.
What the evidence shows, and the data handling around it
State the limits in the evidence itself. It shows policy enforcement and agent actions for individual interactions and tool calls. It does not show that an authorization was clinically correct, and it does not show a claim's adjudication result. Those remain in the authorization and claims systems of record, and the evidence only points to them through the keys.
HIPAA minimum necessary applies to PHI in evidence payloads and in the keys themselves. A covered entity or business associate must make reasonable efforts to limit PHI to what the purpose requires. In practice, redact content not needed to reconstruct the policy decision, store tokenized references instead of member identifiers, and keep clinical text out of evidence unless a specific need is documented.
Retention and data residency are decisions your organization owns. In a self-managed deployment in your own cloud, Trussed supports configurable retention and data residency, so evidence can stay in the regions and for the periods you set. Do not borrow periods from unrelated rules. As of November 2026, 42 CFR 457.730 sets access timelines for CHIP prior authorization information, including the request, decision, status and dates, beginning January 1, 2027. Those timelines govern payer-side data availability, not retention of AI agent evidence.
Keys that connect one decision to one claim
Authorization reference
Issued by the authorization system and present on every related interaction.
Claim reference
Added when the claim exists. Null before that.
Policy decision
Allow, redact, block or escalate, enforced in the execution path.
Evidence record
One record per interaction or tool call, retrievable by either reference.
Make authorization and claim evidence retrievable
See how runtime policy enforcement produces audit-ready evidence for individual interactions and tool calls, and how reference identifiers fit into it.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.