Implementation Guide | Healthcare

    Evidence for every AI-drafted patient message: what to capture and keep

    For each AI-drafted patient message, keep one record that ties together the interaction, the context supplied, the policy version and decision (allow, redact, block, or escalate), the PHI handling action, the model output, the reviewer action, and the final message sent. HIPAA 164.312(b) requires mechanisms that record and examine activity in systems that use ePHI but does not list fields, so the exact record, retention and access rules come from your risk analysis. The strongest evidence is written when the policy decision is enforced, not rebuilt later.

    A minimum evidence record for each message

    1. 1

      Interaction identity

      Unique message ID, date and time, AI version in use, patient ID, and the responsible clinical user. These match the SAFER transaction log fields.

    2. 2

      Input and retrieved context

      The prompt and any context pulled from the EHR or other systems. Store a reference or a redacted copy where the full text is not needed to defend the decision.

    3. 3

      Policy decision and PHI handling

      The policy version in force at generation time, the outcome (allow, redact, block, or escalate for human approval), and what was redacted or withheld.

    4. 4

      Model output

      The draft as generated, before any edit, so a reviewer's changes can be seen against it.

    5. 5

      Reviewer action

      Reviewer identity, approve, edit or reject, and timestamp. This is the human responsibility step SAFER describes.

    6. 6

      Final message and retention metadata

      The message as sent, delivery time, record class, retention period, and the date the record may be disposed of.

    Criteria for evaluating your current architecture

    • Can the system produce one record for a given message showing the policy version, the decision, and the final message sent, without joining separate logs?
    • Is the record written when the policy decision is enforced in the execution path, or assembled later from application logs, vendor dashboards and EHR records?
    • Does the record include the AI version, date and time, patient ID, responsible clinical user, and the reviewer's approval action?
    • Is PHI redacted before storage where full text is not needed, and are encryption and role-based access documented in the risk analysis?
    • Can you choose where evidence is stored, including your own cloud, and configure retention per record type?
    • Is there a named owner who reviews the evidence on a set schedule?

    What 164.312(b) requires, and what it leaves open

    The HIPAA audit controls standard, 45 CFR 164.312(b), requires covered entities and business associates to implement hardware, software, and/or procedural mechanisms that record and examine activity in information systems that contain or use electronic protected health information. It is outcome-based. It does not name data elements, formats or retention periods.

    That leaves the compliance team to decide what a defensible record looks like for an AI-drafted portal reply, appointment follow-up or care instruction. HHS's January 2025 proposal to strengthen the Security Rule says an entity must weigh its risk analysis and organizational factors, such as technical infrastructure and the security capabilities of its hardware and software, when choosing reasonable and appropriate audit controls. The same proposal states that HHS believes more specificity would improve compliance with the audit controls standard, and it would add a standard titled Audit Trail and System Log Controls. It is a proposal, so treat it as a signal about where expectations are heading, not as a current obligation.

    The most specific AI-related guidance is ONC's SAFER guide. It says AI-generated recommendations should be reviewed and approved by a human who takes responsibility before they reach patients. It also recommends a transaction log of AI use, similar to the EHR audit log, that includes the AI version, date and time of use, patient ID and responsible clinical user. That is a documented floor. Everything beyond it is a design decision you should tie to your risk analysis and record.

    Runtime evidence versus reconstruction

    Most organizations today could assemble this record only by joining application logs, vendor dashboards and EHR entries. Each system records its own slice, and none of them records which policy applied when the draft was generated. Reconstruction after the fact can show that a message was sent. It struggles to show what the AI was allowed to do at that moment.

    The alternative is evidence written at the point of the policy decision, in the execution path. When a control allows, redacts, blocks or escalates a draft, that decision and the policy version behind it become the record, as a byproduct of enforcement rather than a separate logging project. This also reflects what HHS's proposal implies: evidence will come from several system types, so a single record per interaction reduces the stitching an auditor must trust.

    The scope matters. Interaction-level evidence shows what happened to one message or one tool call. It does not by itself establish lineage of the underlying model or a trace across a multi-step workflow, and your documentation should not imply that it does.

    Retention, residency and access without new PHI exposure

    The evidence store is itself a system that holds or may hold ePHI, so it falls within the audit controls standard and needs its own protections. Keeping evidence and minimizing PHI pull in opposite directions. The practical resolution is to decide, element by element, what must be stored in full and what can be stored redacted or by reference. Apply redaction before storage, encrypt the store, restrict access by role, and protect records from alteration.

    Where evidence resides is a design decision to document. HHS's proposal asks, for example, where audit information will reside and whether it sits on a separate server. Keeping evidence apart from the systems it audits, and inside infrastructure you control, simplifies residency and access questions. SAFER also recommends daily, off-site, complete, encrypted backups of patient data. An evidence store that holds patient identifiers may warrant equivalent treatment.

    For retention, the proposed Security Rule would require documentation to be kept for six years from creation or the date last in effect, whichever is later. Because that is only a proposal, set retention with counsel's reading of current HIPAA documentation requirements and applicable state medical record rules, and make the period configurable by record type. SAFER also calls for regular review of the AI transaction log, so assign an owner and a review cadence. Collection without review is a weak control.

    Where Trussed fits

    Trussed enforces policy in the execution path for individual interactions and tool calls, with outcomes of allow, redact, block, rate limit, or escalate for human approval. The enforcement generates continuous, audit-ready evidence mapped to HIPAA, including 164.312(b) audit controls. It deploys self-managed in the customer cloud, with configurable retention and data residency, and it runs in production with healthcare organizations.

    This fits the record described above: the policy decision and PHI handling action are produced by enforcement rather than added afterward. Trussed's scope is interaction-level and tool-call evidence. It does not provide transaction-level governance, model lineage or end-to-end tracing of agentic workflows, so those questions need other controls.

    A sensible next step, with or without Trussed, is to inventory every system in the message workflow, define the minimum record, and test whether one real message can be explained from a single record.

    One evidence record per AI-drafted message

    Interaction

    Who, when, which AI version, which patient.

    Policy decision

    Policy version and the outcome: allow, redact, block, or escalate.

    Human review

    Reviewer identity, action taken, and time.

    Final message

    What was actually sent, with retention metadata.

    See what a per-message evidence record looks like

    Walk through how enforced policy decisions become audit-ready evidence for AI-drafted patient messages, stored in your own cloud.

    Request a Demo