Implementation Guide | Healthcare

    Enforcing HIPAA Minimum Necessary on Amazon Bedrock Agents Action Groups

    Apply minimum necessary to action groups in two layers. Design-time controls such as IAM least privilege and narrowly scoped Lambda functions and API schemas limit what an action group can reach. A per-interaction policy in the execution path then decides, for each tool call, whether to allow, redact, block or escalate the request parameters and the returned payload. That second layer also produces the evidence of what was decided and why. HHS has issued no AI-specific statement in the material reviewed, so this mapping is an interpretation of technology-neutral rules.

    1. 1

      Request parameters

      The agent may pass patient identifiers, names or free-text context that the action does not need to complete the task.

    2. 2

      Backend query breadth

      A function that returns a full record when the task needs one field exposes more than the purpose requires.

    3. 3

      Response payload

      Returned records often carry fields beyond the task, and the agent can carry them into later reasoning steps.

    4. 4

      Session state

      Values retained across turns can bring earlier PHI into later, unrelated tool calls.

    5. 5

      Final answer

      PHI the user does not need, or is not entitled to see, can surface in the generated reply.

    6. 6

      Human handoff paths

      Where control returns to the calling application, the same payload handling questions apply there.

    7. 7

      1. Inventory action groups by purpose

      List each action, the purpose it serves and whether the call is routine and recurring. Decide use, disclosure or request status with your privacy officer.

    8. 8

      2. Define allowed data elements per purpose

      For routine calls, document standard protocols that limit PHI to the amount reasonably necessary, as 164.514(d) contemplates for routine and recurring disclosures.

    9. 9

      3. Tighten the design-time layer

      Scope IAM roles, narrow the Lambda function and API schema, and remove fields the backend does not need to return.

    10. 10

      4. Enforce on the request

      Evaluate parameters in the execution path. Allow compliant calls, redact or block parameters outside the allowed elements.

    11. 11

      5. Enforce on the response

      Redact or block fields beyond the purpose before the payload reaches the agent, and route defined high-risk calls to human approval.

    12. 12

      6. Record every decision

      Write the evidence at the point of enforcement and retain it under your own retention rules.

    Evaluation criteria for any enforcement approach

    • Does it evaluate individual tool call parameters and responses, not only access to the action group?
    • Can rules be expressed per purpose, with documented rationale for each allowed data element?
    • Can it allow, redact, block and escalate, and does escalation record the approver?
    • Do the 164.502(b)(2) exceptions and standard transaction flows have a clear representation in policy logic?
    • Is evidence created at enforcement time and retained under your control?
    • Can claims about healthcare production use and HIPAA alignment be verified independently?

    What design-time controls cover and where per-interaction enforcement starts

    Which systems can the action group reach?

    Design time. Answered by IAM scoping, resource boundaries and function design.

    Which operations does the agent offer?

    Design time. Answered by narrow API or function schemas that expose only needed actions.

    Which fields may this call send for this purpose?

    Per interaction. Requires evaluating the actual parameters at the time of the call.

    Which fields may this response return?

    Per interaction. Requires inspecting the actual payload before it reaches the agent.

    Does this call need a human decision?

    Per interaction. Requires a rule that can pause the call and record the approver.

    Can you show what was enforced?

    Evidence. Requires a record created by the enforcement point, not reconstructed afterward.

    What the standard asks of an agent tool call

    Under 45 CFR 164.502(b), a covered entity or business associate must make reasonable efforts to limit PHI to the minimum necessary to accomplish the intended purpose when using or disclosing PHI, or when requesting it from another covered entity or business associate. Section 164.514(d) supplies the implementation requirements. Paragraph (d)(1) requires meeting paragraphs (d)(2) through (d)(5) for any request for, or use or disclosure of, PHI. Separate specifications cover minimum necessary uses, in (d)(2), and disclosures, in (d)(3).

    The text is technology-neutral, and nothing in it addresses AI agents or tool calls. Applying it to action groups is therefore a design interpretation. The useful part of that interpretation is that the standard covers uses, disclosures and requests. An agent's outbound request and the payload returned to it can both be analyzed against the purpose of the interaction.

    Two cautions apply. Whether a given call is a use, a disclosure or a request depends on your role and the facts, so your privacy officer should make that determination. Also, 164.502(b)(2) lists situations where minimum necessary does not apply, and the 2000 final rule preamble notes that the standard is not required for data elements specified in the implementation guides for HIPAA standard transactions. Review the full exception list before deciding which flows need treatment. The standard is a reasonable efforts standard, so a documented rationale for each policy decision matters as much as the decision itself.

    Audit evidence and where Trussed fits

    Section 164.312 covers technical safeguards, and 164.312(b) addresses audit controls. For agent activity, the practical question is whether you can show, per invocation, what was requested, what policy applied, what was decided and who approved any exception. A sensible audit-ready record for an action group invocation contains the requested action and parameters, the policy and version applied, the decision (allow, redact, block or escalate), the fields redacted or blocked, and the approver where one was involved. Records written by the enforcement point show what was actually enforced. Records assembled later only show what was observed.

    Trussed is one way to implement the enforcement layer. It enforces policy in the execution path across models, agents and tools, at the level of individual interactions and tool calls, with actions that include allow, redact, block and escalate for human approval. Enforcement is at runtime, not reviewed after the fact, and it generates audit-ready evidence mapped to HIPAA. Trussed runs in production with healthcare organizations and can be deployed self-managed in the customer cloud, with configurable retention and data residency. It is HIPAA aligned and SOC 2 Type II certified.

    It does not replace IAM scoping, schema design or your privacy determinations. Those remain the design-time layer, and the policy criteria still come from your own documented analysis.

    Minimum necessary at the action group boundary

    Label: Agent decides

    Description: The orchestration chooses an action and its parameters at runtime.

    Label: Policy evaluates the request

    Description: Parameters are checked against the data elements allowed for the purpose.

    Label: Action group executes

    Description: The Lambda function or API returns a payload that may hold more PHI than needed.

    Label: Policy evaluates the response

    Description: Allow, redact, block, or escalate for human approval.

    Label: Evidence recorded

    Description: The decision and its basis are retained for audit.

    See per-call enforcement and evidence for your action groups

    Walk through how policy is enforced in the execution path and what an audit-ready record of each invocation contains.

    See How Evidence Is Produced