Implementation Guide

    Building an audit trail for AI agents in prior authorization

    An audit trail for AI agents in prior authorization is a durable, tamper-evident record, produced at the point of policy enforcement, of what each agent was permitted to do and what it actually did. Each record shows the initiator, the policy version in force, the data accessed or redacted, the tool calls and parameters, the decision outcome and any human approval. Build it by fixing the evidence schema first, capturing records in the execution path, and storing them in write-once or hash-chained storage with explicit retention, residency and access controls.

    Fix the evidence schema first

    • Initiator and agent identity: the agent, the workload and environment it ran in, and the human or organizational authority that triggered the task.
    • Request and data sources: what was asked, and which data sources were accessed, redacted or withheld.
    • Policy decision and version: the outcome (allow, redact, block or escalate), the action classification, and the version of the policy in force.
    • Tool and MCP server calls: the tool or server invoked and the parameters passed.
    • Output and execution result: what was produced or returned, and whether the action completed.
    • Approval and correlation: the approval identifier and human approver where required, plus a correlation identifier that links related records.

    Define the evidence requirement before choosing tools

    Start with the question an auditor will ask: for this agent action, what was the agent allowed to do, and what did it do? The evidence has to answer both halves. Application logs usually capture part of the second half and almost none of the first, because the policy in force is rarely recorded next to the action.

    Scope this work to interaction-level and tool-call evidence. It does not describe end-to-end workflow tracing or transaction-level governance, which are separate design problems.

    The regulatory inputs are indirect, so state them carefully. The HIPAA Security Rule proposal published in January 2025 is a notice of proposed rulemaking, and as of November 2026 it should be treated as proposed. Its text describes maintaining audit controls such as system event logs and application audit logs, reviewing them regularly, determining which types of audit trail data are needed, and considering automated activity review. It also proposes a documented compliance audit at least every 12 months.

    NYDFS Insurance Circular Letter No. 7 (2024) is final guidance. It says the Department may audit and examine an insurer's use of AI systems and that internal audit should be engaged. Its stated scope is underwriting and pricing, so confirm with counsel how it applies to utilization review. It also reflects a NAIC Model Bulletin expectation that certain terms appear in third-party vendor contracts where appropriate. NIST and OWASP material on agent identity and audit logging is guidance rather than law, but it is the most specific description of what an agent record should contain.

    Capture policy decisions in the execution path

    Place authorization in the component that executes the action, outside the agent's own context. The agent then cannot skip, rewrite or omit the check. Every requested tool call or data access passes through a policy decision, and the decision record is written as part of that step rather than collected later.

    Four outcomes cover most prior authorization cases. Allow lets the action proceed. Redact lets it proceed with sensitive fields removed, and the record notes which categories were removed without storing the removed values. Block stops the action. Escalate pauses it until a named human approves, and the record carries the approval identifier and approver.

    Set explicit risk thresholds that trigger escalation, and log the threshold that fired. Fail closed: if policy lookup, approval validation or audit logging fails, the action does not run. This gives you a useful invariant, that no action executes without both a decision and a record.

    Propagate a correlation identifier, for example through OpenTelemetry traces, so related interactions and tool calls can be grouped. Keep that grouping at the interaction and tool-call level.

    Runtime enforcement evidence versus reconstructed logs

    When evaluating a design or a vendor, ask whether each record is generated at the point of enforcement or assembled afterward from application and observability logs. The differences show up in five places.

    • Completeness: In-path: every executed action has a decision record, because execution depends on it. Reconstructed: gaps appear wherever the trigger, decision basis or approval status was never emitted.
    • Policy linkage: In-path: the policy version is written with each decision. Reconstructed: the policy in force has to be inferred from change history.
    • Integrity: In-path: records can go straight to hash-chained or write-once storage. Reconstructed: local logs can be altered without notice.
    • Failure behavior: In-path: audit logging failure stops the action. Reconstructed: the action proceeds and the gap is found later, if at all.
    • Auditor usability: In-path: one structured record answers what was permitted and what happened. Reconstructed: an analyst must correlate traces, events and tool invocations by hand.

    Store, retain and retrieve the evidence

    Evidence records contain prompts, outputs and possibly protected health information, so the store needs the same protection as the data it describes. Use hash chaining or write-once storage so alteration is detectable. Restrict read access to defined roles, separate it from the teams that operate the agents, and redact fields wherever the audit purpose allows.

    Set the retention period explicitly, as a documented decision made with legal and compliance input that reflects the regulations, contracts and state requirements that apply to you. Retention that is merely whatever the logging platform defaults to is hard to defend. Do the same for location: decide where the store lives, and confirm that the region and account match your data residency obligations. Deploying the enforcement and storage layer in your own cloud keeps the evidence under your control.

    Then test retrieval. Run a drill in which an auditor names a recent agent action and you produce the structured record showing initiator, data sources, tool calls with parameters, policy decision and version, and human approval where required. Repeat it on a schedule, and involve internal audit so the format works for the people who will use it. Document how you decided which audit data types to capture and how they are reviewed, since the HIPAA proposal points to those procedures.

    One example of in-path enforcement

    Trussed is one example of this pattern. It enforces policy in the execution path across models, agents, tools and MCP servers, with outcomes of allow, redact, block, rate limit and escalate for human approval. Scope is individual interactions and tool calls. Enforcement is at runtime, not reviewed after the fact, and the audit-ready evidence is generated by that enforcement and mapped to HIPAA, NAIC, NYDFS, the EU AI Act and NIST.

    It deploys self-managed in the customer cloud, with configurable retention and data residency. Trussed has run in continuous production in a regulated enterprise for more than 18 months and generates more than 30 GB of audit evidence per day. Use the evaluation questions above on any option, including this one.

    Evidence path for one agent action

    Request

    Agent identity, initiator and the requested tool call or data access.

    Policy decision

    Allow, redact, block or escalate, recorded with the policy version.

    Execution

    Tool or MCP server call with parameters and the result returned.

    Evidence store

    Tamper-evident, access-controlled and retrievable on request.

    Check whether your agent evidence is produced at runtime

    Walk through your evidence schema, decision outcomes and retention design with a Trussed expert.

    Request a Demo