Implementation Guide | Healthcare

    How to Prove to Auditors That AI Usage Telemetry Never Left the Health System Network

    Proving containment means producing system-generated evidence, not restating a policy. Assemble four things: a data-flow map of every place AI telemetry is generated and stored, network and egress configuration showing no path to vendor analytics or third-party model endpoints, retention and residency settings exported from the deployed system, and time-bounded enforcement records showing what was allowed, redacted or blocked. Together these support the audit controls expectation in 45 CFR 164.312(b), which requires mechanisms that record and examine activity in systems that contain or use ePHI.

    Map telemetry across the execution path

    1. 1

      User and identity

      Who initiated the request, with session metadata and privilege changes recorded.

    2. 2

      Policy layer

      Where the decision to allow, redact or block is made, and where that decision is recorded.

    3. 3

      Model endpoint

      Whether inference runs inside your boundary or at a third-party provider.

    4. 4

      Tool calls

      Which systems agents and assistants can reach, and what each call records.

    5. 5

      Log and evidence stores

      Where telemetry persists, for how long, and in which region or account.

    6. 6

      Ancillary tooling

      Debug, support and monitoring systems that may copy prompts or responses.

    Assemble the audit-ready evidence package

    • Telemetry inventory and scoping decision for each category and store, including the rationale for any store treated as out of scope
    • Data-flow diagram covering the full execution path and ancillary tooling, with the date it was last verified
    • Egress and network control configuration showing no outbound route to vendor analytics, SaaS observability or unapproved model endpoints
    • Exported retention and residency settings from the deployed system, not vendor descriptions
    • Time-bounded enforcement and audit reports showing identity, action, outcome and synchronized timestamps
    • Exceptions register with business justification for any control that could not be applied

    What counts as AI usage telemetry

    For audit purposes, AI usage telemetry is the record an AI system produces about its own use. Five categories cover most deployments: prompts, responses, tool-call records, user and session metadata, and policy decisions. Prompts and responses can contain patient details directly. Tool-call records can carry parameters such as identifiers or query terms. Session metadata and policy decisions may not hold clinical content, but they can still reveal PHI-related activity, such as who queried what and when.

    Classification is your organization's call. Decide per category, and per store, whether it contains or implies PHI, and record that scoping decision. ONC's audit-log test method gives repeated searches for famous individuals as an example of activity worth tracking, which is a useful reminder that access patterns can be sensitive without any message content.

    Section 164.312(b) requires hardware, software and/or procedural mechanisms that record and examine activity in information systems that contain or use ePHI. If an AI system touches ePHI, its activity records fall within the logic of that standard, and the stores holding those records need to be accounted for.

    Architecture patterns and what each lets you prove

    Any path that carries telemetry outside the boundary is an egress path, and your evidence must show it was absent or blocked. The patterns below reflect general architectural reasoning rather than a regulatory classification.

    Self-managed deployment in the customer cloud keeps the telemetry stores under your control, which narrows what you must prove. It does not prove containment by itself. You still need network configuration, residency settings and enforcement records from the running system.

    Treat any telemetry transfer to a third party as a potential disclosure that needs review under the Privacy Rule's limits on use and disclosure of PHI. Your counsel should confirm how that applies to each flow.

    • Vendor-hosted SaaS telemetry: Prompts, responses or metadata may be stored by the vendor. Proof depends on vendor attestations and on what you can verify about their storage, which is harder to demonstrate from your side.
    • Third-party model API calls: Prompts leave the boundary by design. Proof requires showing what was sent, what was redacted before sending, and what the provider retains.
    • Self-managed in the customer cloud: Stores sit in your account or network. Proof comes from your own egress rules, residency and retention configuration, and enforcement records.

    Evidence generated by enforcement versus evidence reconstructed afterward

    A log review after the fact can show what happened, but it cannot show that data was prevented from leaving. If sensitive content has already reached an external endpoint, the log records the exposure rather than the control. Enforcement in the execution path changes the quality of the evidence: the decision to allow, redact or block is made before data crosses the boundary, and the record of that decision is produced at the same moment.

    The ONC audit-log model offers a practical template for what such a record should contain: the user, the action, the outcome, a timestamp synchronized to an NTP standard, and changes to user privileges. ONC's criteria apply to certified health IT modules and are not binding on AI telemetry, but they describe what reviewers are used to seeing. ONC also expects reports to be producible for a specific time period, so confirm your system can export one for the audit window.

    Trussed is one example of this approach. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with allow, redact, block, rate limit and escalation for human approval at the level of individual interactions and tool calls. Enforcement generates audit-ready evidence mapped to HIPAA, and it supports self-managed deployment in the customer cloud with configurable retention and data residency. It runs in production with healthcare and pharma organizations. Verify any such claim against the deployed configuration in your own environment.

    Evaluate the deployment model and keep the proof current

    When assessing a platform or deployment model, ask questions that can be answered with artifacts. Anything the vendor cannot back with documented, approved proof should be treated as outside the supported scope.

    Operationally, set a review cadence for AI logs and enforcement records. HHS notes that regularly reviewing audit logs and access reports can help detect suspicious activity, and CMS recommends periodic auditing of AI-related records. Be ready to hand artifacts over promptly. CMS's internal process allows five business days for requested artifacts, which is a CMS benchmark rather than a HIPAA rule, but it is a reasonable target. Many organizations also adopt a policy against entering PHI into publicly accessible AI platforms, as CMS does for its own users. That is a policy choice, not a HIPAA mandate.

    • Can all telemetry categories be stored entirely in our own cloud or network, and do any components send data to vendor-hosted services?
    • Which retention and residency settings can we configure, and can we export evidence of the active values?
    • Does enforcement occur in the execution path before data leaves the boundary, and are the decisions recorded as system-generated records?
    • Can the system produce time-bounded reports with user identity, action, outcome and synchronized timestamps?
    • What documented proof exists for each claimed capability, and which items fall outside shipped functionality?

    From claim to proof

    Define

    Inventory telemetry categories and decide which stores are in scope.

    Map

    Trace generation, processing and storage across the execution path.

    Control

    Enforce allow, redact and block before data leaves the boundary.

    Evidence

    Export configuration and enforcement records for the audit period.

    Turn the auditor's question into an evidence package

    See how policy enforced in the execution path produces audit-ready evidence, and how self-managed deployment with configurable retention and data residency fits your boundary.

    See How Evidence Is Produced
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.