Power Platform DLP Policies vs Runtime Enforcement for PHI in Copilot Studio Agents
Power Platform DLP policies are a necessary baseline: they classify connectors into business, non-business and blocked groups at the environment or tenant level. They do not evaluate the content of individual agent interactions or tool calls, so PHI can still pass through a permitted connector, prompt or response. Runtime enforcement in the execution path evaluates each interaction and tool call and can allow, redact or block it, producing per-interaction evidence. Healthcare teams whose agents touch PHI should treat DLP alone as insufficient and evaluate a complementary runtime layer.
Criteria for deciding whether DLP alone is enough
- Placement: does it sit in the execution path of each interaction and tool call, so decisions apply before data moves?
- Actions: can it allow, redact, block, rate limit and escalate for human approval, per policy?
- PHI handling: how does it detect PHI, and can you test detection and redaction against your own data types?
- Human approval: can high-risk tool calls be routed to a named approver before they execute?
- Deployment: can it run in your own cloud, with configurable retention and data residency?
- Evidence: what is recorded per interaction, how long is it retained, and can counsel and auditors map it to 164.312(b), access control and integrity?
What connector-level DLP controls
Power Platform DLP policies work by classifying connectors into groups: business, non-business and blocked. The policy is applied at the environment or tenant level, and it decides which connectors an agent or app may use together. For a healthcare organization this is a sound first control. It narrows the set of systems an agent can reach, it is administered centrally, and it prevents a maker from wiring an agent to a connector the organization has already ruled out.
The limit is the unit of control. DLP reasons about connectors, not about what flows through them. It does not evaluate the content of an individual agent interaction or tool call. A connector that is correctly placed in the business group will carry whatever the agent sends it, including PHI that should not have been sent. Configuration correctness and content safety are separate questions, and DLP answers only the first.
What runtime enforcement adds
Runtime enforcement sits in the execution path between the agent and the models, tools and systems it calls. Each prompt, tool call and response passes through a policy decision before it proceeds. The scope is the individual interaction, not the environment.
This changes what a control can do. Instead of a static permit-or-deny decision per connector, a runtime layer can allow an interaction, redact PHI from it, block it, rate limit it, or escalate it for human approval. The decision is made at the moment of the interaction and is enforced at runtime, not reviewed after the fact. It also means the control can distinguish two calls to the same permitted connector: one that carries a minimum necessary payload and one that carries a full patient record.
Side-by-side comparison
The table compares the two approaches on the dimensions that matter for PHI. DLP descriptions reflect connector classification at environment or tenant level. Runtime descriptions reflect enforcement in the execution path.
- Enforcement point: DLP: policy applied to connector groups at environment or tenant level. Runtime: decision made in the execution path of each interaction and tool call.
- Granularity: DLP: the connector. Runtime: the individual prompt, tool call or response.
- PHI content awareness: DLP: does not evaluate the content of interactions or tool calls. Runtime: evaluates content against policy, so detection quality becomes a selection criterion.
- Available actions: DLP: classification of connectors as business, non-business or blocked. Runtime: allow, redact, block, rate limit, escalate for human approval.
- Failure it addresses: DLP: an agent reaching a connector it should not use. Runtime: PHI passing through a connector, prompt or response that is permitted.
- Evidence output: DLP: the policy as configured. Runtime: a record of what was allowed, redacted or blocked for each interaction.
PHI scenarios that a correct DLP policy does not stop
Consider an agent that summarizes patient correspondence. A clinician pastes a note containing a patient name and diagnosis into the prompt. The model is reachable, the connectors are all in the business group, and the policy is correctly configured. DLP has no basis to intervene, because it never inspects the prompt.
Now consider a tool call. An agent retrieves a record from a permitted system to answer a scheduling question, and the response includes clinical fields the task does not need. The connector is allowed, so the full payload reaches the model and may reach the user. A runtime layer can redact the fields that exceed the task or block the response.
A third case is the output path. An agent returns PHI to a user whose role does not entitle them to it, or writes PHI into a destination within an allowed connector group. Again the connector classification is correct and the exposure is real. In each scenario the gap is the same: the control that exists reasons about where data may go, and the exposure depends on what the data is. Microsoft-native controls beyond DLP may address parts of this, and each should be assessed for what it enforces at the interaction level and what it records.
What each approach gives you under HIPAA 164.312(b)
The Security Rule technical safeguards in 45 CFR 164.312 must be implemented in accordance with 164.306. They include access control, which allows access only to those persons or software programs that have been granted access rights, and integrity, which protects ePHI from improper alteration or destruction. Audit controls sit in 164.312(b). Because the access control standard names software programs, agents that touch ePHI are reasonably in view, although how the standard applies to AI agents is an interpretation to confirm with counsel.
The two approaches produce different kinds of evidence. A DLP policy yields a configuration record: it shows the rules as written and which connectors were classified where. That record supports the written policies and procedures that the rule requires organizations to maintain, and it is useful in a periodic evaluation. It does not show what any particular agent did with PHI on a given day.
Runtime enforcement yields activity records tied to individual interactions and tool calls: what was requested, which policy applied, and whether the outcome was allow, redact or block. Which kind of record satisfies an auditor for audit controls is a question for your compliance and legal teams. The practical position is to keep both: configuration records to show intent, and per-interaction records to show behavior. Introducing or changing agents is also an operational change that may warrant re-evaluation under the Security Rule's evaluation standard.
Where Trussed fits
Trussed is one example of runtime enforcement. It enforces policy in the execution path across models, agents, tools and MCP servers, with allow, redact, block, rate limit and escalate-for-human-approval actions at the level of individual interactions and tool calls. Enforcement generates audit-ready evidence mapped to HIPAA, and deployment is self-managed in the customer cloud with configurable retention and data residency.
Trussed runs in production with healthcare and pharma organizations, and has 18+ months of continuous production in a regulated enterprise. It complements DLP rather than replacing it: connector classification remains the baseline, and the runtime layer addresses what passes through the connectors that policy permits.
Two controls, two scopes
Connector-level DLP
Classifies connectors as business, non-business or blocked at environment or tenant level.
Runtime enforcement
Evaluates individual interactions and tool calls in the execution path.
Evidence
Policy configuration records versus per-interaction enforcement records.
Test your Copilot Studio agents against both layers
Walk through your PHI scenarios with an expert and see how enforcement decisions become audit-ready evidence.
Talk to an Expert