Implementation Guide | Healthcare

    Requiring Clinician Approval Before an AI Agent Edits a Shared Care Plan in SharePoint

    Give the agent identity read-only access to care plan content, intercept every write tool call in the execution path, escalate care plan writes to a named clinician, and bind that approval to a hash of the exact proposed edit with an expiry. Retain the request, policy decision, approver decision and outcome as audit evidence. Native SharePoint permissions and versioning remain necessary, but they do not hold an agent's write until a clinician approves it.

    1. 1

      Target location

      The site, library or list is a care plan location.

    2. 2

      Document or item type

      The item is a care plan, not a template or scratch file.

    3. 3

      Change type

      Create, edit, delete, move, rename or permission change.

    4. 4

      PHI present

      The proposed payload contains or alters PHI.

    5. 5

      Agent and context

      Which agent identity made the call and on whose behalf.

    6. 6

      Default

      Any write the policy does not recognise is blocked, not allowed.

    7. 7

      Hold the call

      The write is paused and the agent receives a pending status, not a result.

    8. 8

      Compute the binding

      Hash the full proposed payload together with the target item and its current version identifier.

    9. 9

      Present a diff

      Show the clinician the change against the current content, with the agent identity and stated rationale.

    10. 10

      Record the decision

      Capture approver identity, decision, timestamp and an expiry on the approval.

    11. 11

      Verify at execution

      Recompute the hash and check expiry and item version. Reject on any mismatch.

    12. 12

      Execute once

      Run the write a single time and record the outcome and resulting version.

    Why native controls are not the gate

    An agent that works with shared care plans in SharePoint usually reaches them through Microsoft Graph or a SharePoint connector exposed as a tool. SharePoint permissions determine whether the agent identity may edit an item. Versioning records what changed and who changed it. Approval workflows route content for review. These are all useful, and none of them is designed to sit between an agent's tool call and the write, evaluate the call against clinical policy, and hold it until a specific clinician approves.

    The practical result is that if the agent identity holds a write grant, a bad edit lands first and is reviewed afterward. Version history then lets you restore the prior state, but a care plan may have been read by other staff in the meantime. For a clinical document, review after the fact is the weaker control. The design goal is to enforce approval at the point of the tool call, so that no unapproved write executes.

    Step 1: Separate read and write for the agent identity

    Start with the identity. Give the agent a dedicated identity, not a shared service account or a clinician's delegated session. Scope it to the specific sites, libraries or lists it needs, and grant read access only on care plan content. Whether you use delegated or application-level access, the aim is the same: the agent's own credentials should not be able to modify a care plan directly.

    Writes then happen through a separate, narrowly scoped execution path that is invoked only after approval. This keeps a compromised or misbehaving agent from bypassing the gate, because the credential that can write is not the credential the agent holds. Keeping read access unobstructed also matters operationally. ONC states that practices interfering with access, exchange or use of EHI by automation technologies, including agentic AI, could implicate the information blocking regulations as of November 2026. Gating only the actions that need clinician review limits that exposure.

    • Dedicated agent identity: One identity per agent, named in the AI use-case inventory.
    • Read-only grant on care plans: Scoped to the required sites, libraries or lists.
    • Separate write path: Credentials that can write are held by the enforcement layer, not the agent.
    • Reads stay unobstructed: Only write actions that need review are held.

    Evidence for audit and where native controls suffice

    Each gated edit should leave a complete record: the request and agent identity, the target item, the payload hash and diff, the policy rule that matched, the decision, the approver's identity and decision time, the expiry, and the outcome with the resulting SharePoint version reference. Rejected and expired requests need the same record as approved ones. Together these show an auditor what the agent was allowed to do and what it did, which supports the audit controls expectation in HIPAA 164.312(b). Decide retention period and storage location in advance and make the records tamper-evident.

    Pair the gate with a way to detect and revert an applied change. FDA's guidance on predetermined change control plans for ML-enabled devices describes mechanisms to detect and revert or stop a change that fails specified criteria. That guidance concerns devices, not SharePoint, but the principle carries over: keep SharePoint version history as the rollback path. Also document the clinical-safety rationale for holding agent writes and confirm with counsel how information blocking applies. CMS guidance calls for human oversight of AI before business decisions or external sharing. It is written for CMS users, but it is a useful example of federal practice for documenting mitigation plans and ongoing monitoring.

    Native controls are sufficient for access scoping, version history and human-only editing workflows. Runtime enforcement is needed when an agent can call a write tool and you require that call to be held, evaluated and approved before it executes. Trussed enforces policy in the execution path on individual interactions and tool calls, including escalation for human approval, and generates audit-ready evidence from that enforcement, mapped to HIPAA. It operates per interaction and tool call, so the approval-to-edit binding, tracing across a multi-step workflow and the care plan content policy remain design decisions your team owns.

    Approval gate for an agent write to a care plan

    Agent tool call

    The agent proposes an edit to a care plan document or list item.

    Policy evaluation

    The call is classified as read or write and checked against policy.

    Clinician approval

    A named clinician reviews the exact proposed change.

    Execution and evidence

    The approved edit runs once, and the full record is retained.

    Put the approval gate in the execution path

    See how policy enforcement on agent tool calls, escalation for human approval and audit-ready evidence apply to your care plan write path.

    See How Evidence Is Produced
    Next step

    How does your AI governance program compare?

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