Prior Authorization AI Decision Logging: Recording Model Version, Criteria, and Inputs for Each Recommendation
For each prior authorization recommendation, write one evidence record at runtime, in the request path, containing the model identifier and version, the coverage criteria reference and version, the inputs or hashed references to them, the output, the policy action taken, and any human escalation. Capture each element where it is produced, redact or hash sensitive inputs before the record is stored, and test completeness by retrieving a single recommendation end to end.
Where each element can be captured
- Request entry: assign the recommendation and request identifiers, stamp the time, and capture the raw inputs or their references before any transformation.
- Criteria lookup: record the coverage policy identifier and version as it is resolved, so the record reflects what was applied and not what is current at audit time.
- Model invocation: record the model identifier, version, and routing decision at the call, along with the template or prompt version and the model output.
- Policy decision: record the action enforced on the output (allow, redact, block, or escalate) and the rule that triggered it.
- Reviewer handoff: record escalation to a human reviewer and the reviewer's resulting decision, linked to the same recommendation identifier.
Verifying that records are complete and reproducible
- Pick a single recommendation and retrieve model version, criteria version, inputs, output, and action from the record alone, without correlating separate systems.
- Confirm every required field is populated on a random sample, and alert on records with missing fields.
- Verify that stored hashes still match the source inputs, and that the criteria version resolves to the exact policy text in force at the time.
- Check tamper evidence: show that an altered or deleted record is detectable, and review who holds write and read access to the audit store.
- Reconstruct the evidence for a recommendation, not the model output itself. Model outputs may not repeat on a re-run, so the record must hold the output as issued.
- For each denial, confirm the specific denial reason in the record matches what the provider notice and Patient Access API response carried.
Why the decision record has to be generated at runtime
Many teams can answer audit questions about a prior authorization recommendation only by correlating application logs, a model registry, and policy documents after the fact. The schemas differ, retention differs, and the link between a specific recommendation and the model version, criteria, and inputs behind it is often inferred rather than recorded.
The alternative is to treat the record as evidence created at the moment of the interaction. The EU AI Act points in this direction: high-risk AI systems must technically allow automatic recording of events over the lifetime of the system, and the logging must enable traceability appropriate to the system's intended purpose. Whether a given prior authorization tool is classified as high-risk is a determination for each organization. The design principle holds either way: a record written in the execution path is harder to dispute than one assembled later.
A single record can carry references (identifiers, versions, hashes) to the model registry, the policy repository, and the input store instead of copying their content. That keeps the record compact and avoids a second uncontrolled copy of sensitive data. It is a design choice, not a mandated format.
A minimal field schema
Define the schema before choosing tooling. The fields below are a practical minimum for answering an auditor, regulator, or appeal inquiry about one recommendation. This is a design proposal, and you should extend it to fit your obligations.
- Record identity and timestamp: A unique recommendation identifier, request identifier, and a timestamp recorded when the recommendation is issued.
- Model identifier and version: The deployed model name and exact version. If more than one model can serve a request, also record which was selected and the routing rationale.
- Criteria reference and version: An identifier and version for the coverage policy or clinical criteria applied, not a copy of the policy text.
- Inputs: The input payload, or hashed or tokenized references to it, plus the version of any prompt or template that shaped the request.
- Output and policy action: The recommendation (approve, deny, or pend) and the action enforced on it, such as allow, redact, block, or escalate.
- Reviewer handoff and denial reason: Whether a human reviewer was engaged, and a specific denial reason field that can feed both the provider notice and the Patient Access API response.
Handling protected health information in the record
The record has to substantiate a recommendation without becoming a second copy of the clinical file. Three techniques cover most cases, and they can be combined.
Redaction removes or masks sensitive fields from what is stored, while the enforcement point still sees the full input at runtime. Tokenization or hashing replaces values with stable references, so the record can show which input was used and a reviewer with proper access can resolve it from the system of record. A metadata-only record stores processing steps and decision points but not the underlying content. The tradeoff is real: the more you strip, the more you depend on the source system retaining the original in a form that can be matched to the hash.
Treat the audit store as a sensitive system in its own right. Restrict who can read it, separate the right to read from the right to write, and prevent modification of existing records. Set retention per record class and set data residency to match where the underlying data must stay. These choices should be explicit configuration, agreed with counsel, not defaults. HIPAA's technical safeguards at 45 CFR 164.312 apply to covered entities and business associates, so map each field and each access path in the audit store to the specific standards your counsel identifies.
Mapping to requirements and where enforcement fits
Three external anchors are useful for scoping. The CMS prior authorization final rule requires a payer that denies a request to give a specific denial reason through the Patient Access API and in the notice to the requesting provider. It does not require detailed clinical criteria to be shared with patients, so the internal evidence record and the patient-facing notice are separate artifacts that should share a reason field. The NIST Cyber AI Profile, an initial preliminary draft as of November 2026, advises using logging to show requests, responses, metadata, and relevant metrics, which supports interaction-level capture. The OWASP checklist adds that audit records themselves should be confirmed secure.
Trussed enforces policy in the execution path across models, agents, tools, and MCP servers, scoped to individual interactions and tool calls: allow, redact, block, rate limit, or escalate for human approval. Because enforcement happens at runtime, not reviewed after the fact, each interaction yields audit-ready evidence, mapped to HIPAA, NAIC, NYDFS, the EU AI Act, and NIST. It can be deployed self-managed in the customer cloud, with configurable retention and data residency. The model version in a record is an identifier your registry supplies. The record does not establish model lineage or dataset provenance, and it does not trace agentic workflows end to end.
One recommendation, one evidence record
Model
Identifier and version that produced the recommendation
Criteria
Coverage policy reference and version applied
Inputs
Payload or hashed references supplied to the model
Outcome
Output, policy action taken, and reviewer escalation
Review your per-recommendation evidence record
Walk through how runtime enforcement produces audit-ready evidence for individual AI interactions in a regulated enterprise.
Talk to an Expert