How to Govern Epic AI Features: Governance Checklist for Health Systems
Use this practitioner checklist to assess whether your health system has the identity, permission, and audit controls needed to govern Epic’s embedded AI features, before and after deployment.
Why embedded AI features create a governance gap
Epic’s embedded AI features sit inside clinical workflows and can invoke models, APIs, and third-party tools. Native EHR configuration is typically built around human users and application service accounts. That model often leaves AI-specific controls incomplete: identity may not be separable per feature, data scopes may be inherited from broad clinical roles, and audit logs may record only the final chart change rather than the full AI action path.
A practical governance program maps control points first, then closes gaps with policy, process, and (where needed) a supplementary runtime governance layer.
Five control points to map before you govern Epic AI
Confirm each of the following before treating an Epic AI feature as production-ready under your risk and compliance framework.
-
Agent identity
Determine whether each AI feature has its own machine identity, distinct from the EHR application’s service account, so individual actions can be attributed and revoked without disabling the entire application.
-
Tool-call inventory
Map every point at which an AI feature can invoke an external model, API, or third-party tool, and document which data fields are exposed at each call.
-
Policy enforcement timing
Identify whether allow, deny, or redact decisions occur before a tool-call executes, at execution, or only during retrospective audit review.
-
Audit lifecycle coverage
Check whether logs capture the full sequence of an AI-driven action (invocation, data accessed, output generated, and downstream use) or only the final record change.
-
Session identity propagation
Confirm whether the acting user’s identity carries through to AI subprocesses. Loss of identity context at the tool-call layer weakens least-privilege enforcement even when the front-end session is properly authenticated.
Governance outcomes these controls support
When the five control points are in place, health systems can show coherent accountability across identity, permission, enforcement, audit, and ownership.
Agent identity
Distinct, revocable identity per AI feature.
Least privilege
Data scopes defined for AI use, not reused roles.
Tool-call enforcement
Policy checks before execution, not only after.
Audit completeness
Full lifecycle logging, not just record changes.
Team ownership
Clear IT, informatics, and compliance accountability.
Epic AI governance checklist
Work through these items for each Epic AI feature in production or advancing toward go-live. Mark gaps where native configuration is insufficient and note who owns remediation.
- Each AI feature has a distinct machine identity, separate from the shared EHR service account, with a defined revoke path.
- Least-privilege data scopes are defined for AI use rather than inherited wholesale from clinical roles.
- Every external model, API, and third-party tool-call is inventoried, including data fields exposed at each call.
- Allow, deny, and redact decisions are enforced at or before tool invocation, not only in retrospective review.
- Audit logs cover the full AI action lifecycle: invocation, data accessed, output generated, and downstream use.
- The acting user’s identity propagates into AI subprocesses so least privilege remains enforceable at the tool-call layer.
- Ownership is assigned across IT, clinical informatics, and compliance for configuration, review, and incident response.
Operational practices to close the gaps
- Inventory production and sandbox separately: Track which Epic AI features are active in production versus test environments, since governance gaps often originate from features enabled without formal review.
- Independent logging for external calls: Route any AI feature that invokes an external or third-party model through a logging and policy checkpoint independent of native EHR logging.
- Treat vendor updates as change events: Assume Epic AI feature releases can alter data access or tool-call behavior without local IT initiating the change, and require review before activation.
- Validate logs before go-live: Test audit log completeness against a simulated end-to-end AI action rather than assuming native logging is sufficient.
- Write policy for AI-initiated access explicitly: Review whether existing HIPAA-aligned access control policies address AI-initiated data access, or whether they were written only for human users.
Supplementary governance layer
Native EHR configuration rarely covers all five control points on its own. Where identity separation, pre-execution policy checks, or full-lifecycle audit are missing, plan for a runtime governance layer that enforces controls at the point of tool invocation.
Dividing governance responsibility across teams
Clear ownership prevents control points from falling between groups. IT typically owns machine identity, integration paths, and logging infrastructure. Clinical informatics owns feature selection, clinical data context, and workflow fit. Compliance owns policy alignment, access review expectations, and evidence for auditors.
For each Epic AI feature, document a single accountable owner per control point (identity, scopes, enforcement, audit, and go-live sign-off), and require joint review when a vendor update changes tool-call behavior or data access.
A working Epic AI governance checklist verifies five control points before and after deployment: a distinct machine identity for each AI feature, least-privilege data scopes defined for AI use rather than inherited from clinical roles, policy enforcement at the point of tool invocation, complete audit logging across the full AI action lifecycle, and clear ownership across IT, clinical informatics, and compliance. Native EHR configuration rarely covers all five without a supplementary governance layer.
Verify Your Epic AI Governance Posture
Use this checklist to identify where identity, permission, and audit gaps exist across your Epic AI features, and where runtime enforcement is needed to close them.
Explore Runtime Governance