Copilot Studio Agents Processing PHI Under a Microsoft BAA: Covered Features and Gaps
A signed Microsoft BAA is a contractual baseline. It is the written assurance that the provider will safeguard PHI it creates, receives, maintains or transmits. It does not decide which Copilot Studio agents, connectors, knowledge sources or channels are appropriate for PHI. It does not enforce minimum necessary access, control what PHI appears in prompts and tool calls, or produce the audit evidence expected under 164.312(b). Those responsibilities stay with the covered entity or business associate.
Controls and evidence required before an agent processes PHI
- BAA in place, with each component in the agent recorded as in or out of scope for PHI
- Written approval list of connectors, actions and channels permitted to carry PHI, with all others blocked
- Minimum necessary rule per use case covering prompt content, retrieved content and tool-call fields
- Documented retention, residency and access decisions for prompts, outputs and logs
- Human approval defined for consequential actions, with a mechanism that pauses execution
- Activity records covering invocations, retrievals, tool calls and policy outcomes, retained and reviewable for audit
What a BAA does and does not do
HIPAA permits a covered entity to disclose PHI to a business associate once it obtains satisfactory assurances, in a contract or other written arrangement, that the business associate will appropriately safeguard the information. HHS applies the same logic to cloud services: customers must obtain a BAA with the cloud service provider. The BAA must contain the elements specified at 45 CFR 164.504(e), and it cannot authorize uses or disclosures that would violate the Privacy Rule if a covered entity made them.
That is a statement about the relationship between two regulated parties. It is not a configuration of your agent. The Security Rule's technical safeguards at 45 CFR 164.312 apply to a covered entity or business associate, so the duty to implement them cannot be assumed to sit wholly with the vendor. A business associate is also directly liable for certain provisions of the HIPAA Rules, which means a software provider and its customer each carry obligations, and neither discharges the other's.
In practice, treat the BAA as a prerequisite. It tells you that PHI may lawfully be handled by the provider. It does not tell you that every agent you can build is a sound place to put PHI.
Mapping agent components to PHI risk
Organize the review by component, because PHI enters and leaves an agent at different points and each point has a different owner. For every component below, record whether it is within the scope of your BAA and the provider's current HIPAA eligibility terms before PHI is permitted. The risks listed are structural and apply regardless of platform.
- Knowledge sources: Risk: retrieval over unstructured content can surface PHI the requesting user has no need to see, and place it in a generated answer. Customer decision: which repositories are indexed, and whether retrieval respects the user's own access rights.
- Connectors: Risk: a connector sends data to another service, and that service needs its own written assurances if it handles PHI. A BAA with one provider does not extend to it. Customer decision: which connectors are approved for PHI, and which are blocked.
- Actions and tool calls: Risk: PHI travels in call parameters and returns in responses, often outside the visible conversation. Customer decision: which tools an agent may call, with what fields, and which calls need human approval.
- Channels: Risk: external or custom channels move PHI through services and devices outside your controlled environment. Customer decision: which channels may carry PHI, and how users are authenticated on each.
- Generative orchestration: Risk: the model chooses at runtime which knowledge and tools to use, so the data flow is not fixed at design time. Customer decision: how to constrain and observe those choices on each interaction.
Runtime responsibilities that remain with you
Minimum necessary is a customer-side discipline, and agents make it harder to apply. A person typing a question can decide what to include. An agent assembles prompts from retrieved content, prior turns and tool results, so PHI can enter the model context without anyone choosing to put it there. Define what PHI may appear in prompts and responses for each use case, then decide where that definition is enforced rather than relying on instructions to the model.
Persistence deserves the same attention. CMS internal guidance on Microsoft 365 Copilot, which is not a legal requirement for other organizations and does not describe Copilot Studio, tells staff to limit PII and PHI in prompts and responses to the minimum necessary. It also notes that output may persist or be accessible beyond the user's immediate session and should be treated as unofficial until reviewed. The same questions apply to any agent: where do prompts, outputs and tool results persist, who can read them later, and how long are they kept?
Finally, human review is a design choice, not a HIPAA mandate. For actions with consequences, such as writing to a record or sending a message outside the organization, requiring approval before execution is a reasonable control, and it needs a mechanism that actually pauses the action.
Audit controls and the evidence an auditor will ask for
The Security Rule's technical safeguards include an audit controls standard at 164.312(b). Where an agent touches ePHI, an auditor or investigator will want an answer to two questions: what was the agent permitted to do with PHI, and what did it do? These are different pieces of evidence, and they are often held in different places.
The first is policy evidence: the approved components, the rules on what PHI may enter prompts and tool calls, who approved the agent for PHI use, and when. The second is activity evidence: which user or service invoked the agent, which knowledge was retrieved, which tools were called with what data, what was returned, and whether any action was blocked, altered or escalated.
Platform and tenant audit records are a useful starting point, but evaluate them against that list instead of assuming they satisfy it. Ask whether they capture tool-call parameters, whether they show what was withheld or changed before the model saw it, how long they are retained, and whether they can be tied to a specific policy decision. Gaps in any of these mean the evidence must be produced by something you operate.
A 2025 proposed rule would update the Security Rule. HHS lists it as a proposed rule as of November 2026, so current 164.312 text remains the operative standard. Design to the current rule and monitor the proposal.
The gap that remains: enforcement at runtime
Contract coverage and design-time configuration both stop short of the individual interaction. A connector can be approved and still receive a field it should never have received. A knowledge source can be correctly scoped and still return a record that a given request does not justify. Reviewing logs afterward shows that this happened, but it does not prevent it.
Trussed addresses this gap. It enforces policy in the execution path, at runtime and not after the fact, on individual interactions and tool calls across models, agents, tools and MCP servers. Each interaction can be allowed, redacted, blocked, rate limited or escalated for human approval. That makes minimum necessary rules operational at the point where PHI would otherwise enter a prompt or leave in a tool call.
Enforcement generates audit-ready evidence continuously, mapped to HIPAA, including the audit controls expectations under 164.312(b). Trussed can be deployed self-managed in the customer cloud, with configurable retention and data residency, so evidence and PHI-adjacent data stay within the organization's boundary. It runs in production with healthcare organizations, and the company holds SOC 2 Type II and HIPAA attestations among its certifications.
This complements the BAA and platform controls rather than replacing them. It does not change which components are in scope, and it does not remove the need for the approval gate above.
Where coverage ends and responsibility begins
Contract baseline
The BAA provides written assurances from the provider about PHI it handles.
Component scope decisions
You decide, per component, whether it is appropriate for PHI.
Runtime handling
Minimum necessary applies to every prompt, retrieval and tool call.
Audit evidence
You must be able to show what the agent was allowed to do and what it did.
Close the gap between contract coverage and runtime behavior
See how policy is enforced on agent interactions and tool calls, and how audit-ready evidence is produced for PHI workloads.
Explore Agent and Tool-Call Enforcement