Implementation Guide: Healthcare

    Restricting Copilot Studio knowledge sources to HIPAA minimum necessary PHI

    Treat minimum necessary as two problems. At design time, inventory each agent's knowledge sources and scope them to the agent's purpose and the user's role, using native controls such as source selection, user-authenticated access, sensitivity labels, DLP and environment policies. At runtime, enforce policy in the execution path so each response is allowed, redacted or blocked against that scope, and keep per-interaction evidence. Native scoping limits what an agent can reach. It does not by itself show what each response contained.

    Evidence an auditor can reasonably expect

    • The written purpose, role scope and approved knowledge sources for each agent, with approval dates and owners
    • The source inventory showing PHI categories per source and the scoping decisions made
    • Records of native control configuration, including DLP and environment policies and sensitivity label use
    • A per-interaction record of the policy applied and the outcome: allowed, redacted or blocked
    • Records of escalations and human approvals where a response was held for review
    • Retention settings showing where evidence is kept and for how long

    What minimum necessary asks of an agent

    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 of a use, disclosure or request. Section 164.514(d) sets out how to implement that standard. It separates minimum necessary uses, in (d)(2), from minimum necessary disclosures, in (d)(3). For routine and recurring disclosures, it calls for policies and procedures, which may be standard protocols, that limit the PHI disclosed to the amount reasonably necessary for the purpose. The standard has exceptions, so confirm with your privacy officer which uses of agent output it covers. As of November 2026 these provisions are in force.

    The standard is framed as reasonable efforts and sets no technical specification. No Copilot Studio setting is required by HIPAA, and mapping the rule to platform controls is a design judgment you will need to document. A practical reading is that each agent needs a defined purpose and a defined set of user roles, and those two things become the criteria that scope its knowledge sources.

    The standard is written around uses and disclosures of PHI, not only around access. An agent that can reach a large SharePoint site may be authorized to read it and still return more PHI in a given answer than the user's task needs. That is why the guide below separates scoping from enforcement.

    Design-time scoping, step by step

    These steps are the criteria exercise that 164.514(d) implies. Keep the outputs as written records, because the rule speaks in terms of policies, procedures and criteria.

    • 1. Define purpose and roles per agent: Write one statement of purpose and a list of permitted user roles for each agent. A standing agent with a fixed purpose is a candidate for a standard protocol, in the sense of 164.514(d)(3)(i).
    • 2. Inventory every knowledge source: List each SharePoint site, Dataverse table, uploaded file, website and connector the agent can reach, with the PHI categories each may contain.
    • 3. Compare sources to purpose: Remove or narrow any source whose PHI exceeds what the purpose needs. Prefer narrow libraries, views or tables over whole sites or databases.
    • 4. Apply native controls: Use source-level scoping, user-authenticated access so retrieval reflects the individual user's permissions where configuration allows, sensitivity labels, DLP policies and environment policies.
    • 5. Assign ownership: Name owners in security, privacy and the agent team. Require review whenever sources, roles or the agent's purpose change.
    • 6. Define output review: Decide which agent outputs need human review before use. The CMS memo that treats Copilot-assisted content as unofficial until reviewed is one example. It is agency policy, not a HIPAA requirement.

    What native controls cover and where gaps remain

    Native scoping works at the level of reach. Source selection decides what an agent can draw on. Inherited permissions and user-authenticated access decide what a given user can retrieve. Sensitivity labels, DLP policies and environment policies add boundaries around data and around where agents are built and run. Together these are the right first layer, and they reduce the volume of PHI an agent can retrieve. Verify the exact behavior of each against current vendor documentation for your tenant before relying on it in your written criteria.

    The remaining gap is at the interaction level. A user may be permitted to open a record, and the agent may be correctly connected to a permitted source, but the response can still include more PHI than the question required. Scoping by source and permission does not by itself show that each response drew only the minimum necessary PHI. Nor does it produce a per-response record of what was returned and what was withheld.

    The CMS memo applies the minimum necessary limit to Copilot prompts and responses, not only to data access. It governs CMS users and is not a rule for all covered entities, but it reflects the same response-level framing.

    Runtime enforcement in the execution path

    Runtime enforcement evaluates each interaction against the agent's documented purpose and role scope while the interaction is happening. The outcomes are simple: allow the response, redact the PHI that falls outside scope, or block the response. This is enforced at runtime, not reviewed after the fact, and it uses the criteria you wrote in the scoping steps as its reference. If those criteria are vague, enforcement will be vague too.

    Trussed enforces policy in the execution path of AI interactions and tool calls, with allow, redact, block, rate limit and escalation for human approval available as outcomes. Its scope is the individual interaction. Trussed does not provide business transaction identity, model provenance or end-to-end tracing of agentic workflows, and this guide does not rely on those. It is a complement to native scoping, not a replacement for it. Trussed runs in production with healthcare and pharma organizations, and can be deployed self-managed in the customer's cloud with configurable retention and data residency.

    Design the split deliberately. Native controls narrow the set of data an agent can reach, which keeps the runtime policy small and easier to reason about. Runtime enforcement then handles what scoping cannot express, such as PHI categories that are legitimate for one role but outside the purpose of a given question.

    Tradeoffs and evaluation criteria

    Narrow scoping reduces answer coverage. Agents with tightly limited sources will decline more questions, and that is the intended behavior for a minimum necessary design. Expect to tune scope with the agent owner rather than widening sources by default.

    Runtime enforcement adds a component to the response path, so evaluate it as you would any control there. Ask how it expresses policy by agent purpose and user role, whether it can redact rather than only block, how it behaves on failure, and whether it produces evidence per interaction that maps to HIPAA. Trussed generates continuous audit-ready evidence from enforcement, mapped to HIPAA among other frameworks, and has run 18+ months in continuous production in a regulated enterprise. Whatever you choose, test it against your own agent scopes before it governs live PHI.

    A sound next step is to pick one production agent, complete the six scoping steps, and use the result to define what an interaction-level policy and its evidence should look like.

    From minimum necessary to enforceable control

    Define

    Write down each agent's purpose and the roles that may use it.

    Scope

    Limit knowledge sources and access to what that purpose requires.

    Enforce

    Allow, redact or block PHI in each response at runtime.

    Evidence

    Retain per-interaction records that an auditor can review.

    Turn minimum necessary into enforced policy and evidence

    Talk through your Copilot Studio agents, their knowledge sources and the interaction-level evidence your auditors will ask for.

    Talk to an Expert