Technical Guide | Healthcare

    Where to Place AI Policy Enforcement When the EHR Is Vendor-Hosted

    For AI interactions the health system can route itself (its own agents, third-party model calls, coding assistants, tool and MCP access), enforce policy in the health system's own AI execution path. That is the only placement that can allow, redact, block, rate limit or escalate an individual interaction before it completes, and that produces evidence the health system holds independently. Vendor-native controls remain the only option inside the vendor's runtime, so they need contractual and evidentiary terms. A network or gateway boundary adds coverage but sees less interaction context.

    Map the AI touchpoints and who controls each

    1. 1

      Embedded vendor AI features

      Execute inside the vendor's runtime. The health system typically has configuration options, contract terms and whatever evidence the vendor exposes, but cannot place its own enforcement in that path.

    2. 2

      Health-system-built agents

      The health system controls the code and the calls the agents make, so model calls and tool calls can be enforced in its own path.

    3. 3

      Third-party model services

      The vendor controls the model. The health system controls the outbound call, including what PHI leaves and what comes back.

    4. 4

      Coding assistants

      Run in developer environments the health system manages, so requests and responses can be brought under policy.

    5. 5

      MCP servers and tool integrations

      Access to tools and data sources is mediated by systems the health system can route through an enforcement point.

    The shared-responsibility boundary decides placement

    A health system running a vendor-hosted EHR cannot change the vendor's runtime. It still answers for what AI does with PHI, whether the AI is a feature embedded in the EHR or a model, agent or coding assistant operating around it. Enforcement location is therefore an architectural decision, not a configuration detail.

    The preamble of the Security Rule proposed rule published in January 2025 gives a useful frame. It says that where multiple persons control a technology asset, each should include that asset in its own technology asset inventory and network map. Its example is a covered entity contracting with a cloud-based EHR vendor: both the covered entity and the vendor have control over the ePHI in the EHR. This is proposed language, not a current requirement, but the design pattern holds regardless. The health system should record the EHR and its AI-related data flows in its own inventory rather than rely on the vendor's records.

    The current technical safeguards in 45 CFR 164.312 apply to both covered entities and business associates, and the covered entity is responsible for complying with Part 164 and for its policies and procedures. Delegating enforcement to a vendor does not appear to remove that responsibility. Contract terms, such as prohibiting secondary use of PHI for model training, are necessary, but a contract clause is not a technical control and does not by itself produce evidence.

    Evaluating the three placements

    The comparison below is architectural analysis, not a benchmark. Evaluate each placement against five criteria: PHI handling (can it redact or block before data leaves), scope of coverage, evidence quality (who generates it and who holds it), data residency, and latency added to the interaction.

    • Vendor-native controls: Coverage: only features inside the vendor runtime. PHI handling: depends on vendor options. Evidence: generated and held by the vendor, so harder to evidence independently. Residency: set by vendor hosting. Latency: none added by you. Cannot cover AI used around the EHR.
    • Network or gateway boundary: Coverage: traffic crossing the boundary you operate. PHI handling: coarse allow or deny unless payloads are inspected. Evidence: connection-level records, with limited visibility into individual tool calls and agent actions. Residency: under your control. Latency: one added hop.
    • AI execution path: Coverage: model calls, agent tool calls and MCP access routed through it. PHI handling: redact, block or escalate per interaction. Evidence: generated at the point of enforcement and held by the health system. Residency: depends on deployment location. Latency: one added evaluation step per call.

    What enforcement in the execution path adds

    The distinction that matters is timing and granularity. A vendor setting or a periodic review describes what should happen or what did happen. Enforcement at runtime acts on the individual interaction or tool call as it occurs: allow it, redact PHI from it, block it, rate limit it, or hold it for human approval. A disallowed interaction is stopped, not discovered later.

    Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with those five actions applied to individual interactions and tool calls. It can be deployed self-managed in the customer's cloud, with configurable retention and data residency, so evidence and any PHI-bearing payloads stay in an environment the health system controls. Trussed runs in production with healthcare and pharma organizations.

    The limits are worth stating plainly. In-path enforcement does not reach inside the vendor's runtime, so embedded EHR AI features still rely on vendor controls and contract terms. It also adds an evaluation step to each call, so latency should be measured against the interactive and batch workloads you actually run. A common pattern is to combine placements: in-path enforcement for what the health system routes, a network boundary to find and constrain traffic that bypasses it, and vendor-native settings for embedded features.

    Define the placement decision and the evidence

    The evidence question should drive the decision. For each AI touchpoint, you should be able to show a regulator or auditor what the AI was allowed to do and what it did. For HIPAA 164.312(b) audit controls, decide in writing which party produces the evidence for each interaction, where it is stored, and for how long. Evidence produced at the point of enforcement and stored in your own environment is easier to defend than a report you request from a vendor.

    Enforcement-generated evidence for an interaction should record the actor, the resource or tool requested, the policy applied, the action taken (allow, redact, block, rate limit or escalate), the time, and the approver where a human decision was required. Trussed generates continuous audit-ready evidence from enforcement, mapped to HIPAA among other frameworks.

    • Inventory every AI touchpoint that can handle PHI and add each to your asset inventory and network map, noting which party controls it.
    • Classify each touchpoint as vendor-controlled, health-system-controlled or third-party, and mark where policy can run.
    • Choose a placement per touchpoint using the five criteria: PHI handling, coverage, evidence quality, residency and latency.
    • Define AI policy content: acceptable use, approval and review, documentation standards for AI-generated content, and vendor data-sharing limits.
    • Review BAAs and vendor contracts for restrictions on secondary use and model training with PHI, and for who retains AI interaction records.
    • Assign evidence ownership, storage location and retention per interaction type, and write it into the vendor and internal agreements.

    Three candidate enforcement placements

    Vendor-native

    Settings inside the EHR vendor's runtime, which the health system cannot modify.

    Network or gateway boundary

    Controls on traffic leaving or entering environments the health system operates.

    AI execution path

    Policy enforced at model calls, agent tool calls and MCP server access.

    Map your enforcement points before you choose a placement

    Review your AI touchpoints with an expert and see how enforcement in the execution path produces audit-ready evidence in your own cloud.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

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