HIPAA audit controls for AI

    Producing Department-Level Evidence of Which AI Models Accessed PHI

    To show OCR or an auditor which AI models accessed PHI by department, keep one record per model interaction. Each record should identify the model, the user or service identity, the department, the PHI handling outcome (allowed, redacted or blocked), the policy applied and the time. HIPAA 164.312(b) requires mechanisms that record and examine activity in systems that contain or use ePHI. Records produced when the interaction happens are more credible than records reconstructed after a request arrives.

    Can your current approach answer an audit request?

    • You can list every AI model, tool and assistant that could access ePHI, including unmanaged tools.
    • For a defined period, you can produce the models that accessed PHI, grouped by department.
    • Each record names an authenticated user or service identity, not a shared account.
    • Each record shows the policy applied and the PHI outcome: allowed, redacted or blocked.
    • Records are created when the interaction happens, not assembled after a request arrives.
    • Your risk analysis documents why these fields are reasonable and appropriate, and it addresses AI tools specifically.

    What 164.312(b) requires and what it leaves to you

    The audit controls standard at 45 CFR 164.312(b) requires a covered entity or business associate to implement hardware, software, and/or procedural mechanisms that record and examine activity in information systems that contain or use ePHI. It is listed as Required. The text is technology-neutral, so it applies to an AI assistant or model that processes PHI in the same way it applies to any other system that uses ePHI.

    HHS Security Series guidance states that the Security Rule does not identify the data that audit controls must gather or how often audit reports must be reviewed. Instead, the entity weighs its risk analysis and organizational factors, including technical infrastructure and the security capabilities of its hardware and software, to decide what is reasonable and appropriate. For AI, this means the field set in an interaction record is a design decision. You need to be able to explain why each field was chosen.

    Two related points shape that design. First, 164.312(a)(1) refers to access by persons or software programs that have been granted access rights, and Unique User Identification is a Required specification. Models, assistants and service accounts are reasonably treated as access subjects whose activity must be attributable to an identity. Second, the preamble to the January 2025 NPRM says ePHI used in AI training data, prediction models and algorithm data is protected by the HIPAA Rules. It also says a risk analysis should consider the type and amount of ePHI an AI tool accesses, to whom the data is disclosed and to whom the output is provided. That is a proposal, and the NPRM also proposes more specificity for audit controls. As of November 2026, 164.312(b) as currently written remains the standard to design against.

    In practice, the question reaches you as a data request. OCR investigations collect evidence through requests for data from the entity and review of relevant documents. A request such as "which AI models accessed PHI, in which departments, under what policy" is answerable only if the records already exist in a form that can be produced.

    The minimum fields for an AI interaction record

    No source prescribes a field list, so the following is an evidence-design proposal. Each field answers a specific part of the audit question, and each should be justified in your risk analysis. The scope is the individual interaction and tool call. These records do not establish model lineage, dataset provenance or end-to-end tracing of multi-step agentic workflows, and the page should not be read as claiming them.

    Beyond these six fields, the NPRM's proposed risk factors suggest two candidates worth considering: the category or volume of PHI involved, and the destination of the output. Add them if your risk analysis shows you need them.

    • Model identity: The specific model, assistant or tool that handled the interaction, including the version where it is known.
    • User or service identity: The authenticated person or service account that initiated the request, so every interaction is attributable.
    • Department: The organizational unit, such as a clinical, revenue cycle or administrative function, resolved at the time of the request.
    • PHI handling action: The outcome applied to PHI in the interaction: allowed, redacted or blocked.
    • Policy applied: The identifier and version of the policy that produced the decision.
    • Timestamp: When the interaction and decision occurred, from a consistent clock source.

    Attributing interactions to departments, and where attribution fails

    Department attribution is usually an identity problem. The reliable approach is to authenticate every request against your identity provider, map the identity to a department through directory groups or role assignments, and write the resolved department into the record at the moment of the request. Storing the value in the record, rather than looking it up later, matters because people change roles and groups are reorganized. A record that resolves the department at query time can describe a past interaction with a present-day org chart.

    Unattributed PHI access tends to come from a small set of causes. Shared service accounts collapse many teams into one identity. Applications that call a model with a single credential hide the department of the end user. Employees using unmanaged AI tools never produce a record at all. Users who belong to several groups can be assigned to the wrong department, and stale group mappings carry the same risk. Finally, some records capture the request but not the policy decision, which leaves the handling outcome unknown.

    Two practices reduce the exposure. Make the unattributed category visible in reporting, with a count of interactions that have no department, instead of dropping them. Treat each source of unattributed access as a finding in the risk analysis, with an owner and a remediation. An honest count of gaps is more defensible in an investigation than a clean report that omits them.

    Runtime evidence compared with after-the-fact reconstruction

    Reconstruction usually means assembling application logs, vendor records and staff interviews after a request arrives. These sources use different schemas and identifiers, and they rarely record which policy was in force or what happened to PHI. The policy decision then has to be inferred, which weakens the evidence. Department attribution has to be rebuilt from whatever identity data each system happened to keep.

    Enforcement at runtime reverses the order. The policy decision is made in the execution path, before the model receives or returns PHI, and the record is a direct result of that decision. The model, identity, department, outcome and policy are written together by the same component, so the record shows what was enforced rather than what someone later believes was enforced. The tradeoff is that the enforcement point must cover every route to a model. Interactions that bypass it fall back to reconstruction and appear as gaps.

    Trussed enforces policy in the execution path across models, agents, tools and MCP servers. It applies allow, redact or block decisions to individual interactions and tool calls, enforced at runtime, not reviewed after the fact. That enforcement generates audit-ready evidence mapped to HIPAA. Trussed runs in production with healthcare organizations, and it can be deployed self-managed in the customer's cloud with configurable retention and data residency. Retention length is a decision for your own risk analysis and policies.

    The audit question as an evidence chain

    Which models

    Model identity recorded on every interaction that touches PHI.

    Which departments

    Department attributed from the authenticated user or service identity.

    What happened to PHI

    Handling outcome recorded as allowed, redacted or blocked.

    Under what policy

    The policy applied is stored with the record, not inferred later.

    Test your evidence against a real audit question

    Walk through how enforcement in the AI execution path produces department-attributed records of PHI interactions, mapped to HIPAA 164.312(b).

    Request a Demo