Requiring Clinician Sign-Off Before an AI-Drafted Patient Message Is Sent
Clinician sign-off is a mandatory control only when delivery is technically impossible without a recorded approval. Place an approval gate in the execution path between draft generation and message delivery. Classify each draft, evaluate policy, redact PHI where required, escalate for human approval, block delivery until a uniquely identified clinician approves, and record the decision as evidence. If any send path skips the gate, sign-off is a recommendation, not a control.
- 1
Classifier
Assigns a message category and detects PHI in the draft before any routing decision.
- 2
Policy engine
Evaluates category, content, and context against written policy and returns allow, redact, block, or escalate.
- 3
Redaction step
Removes or masks PHI that policy does not permit in the draft before a human sees it or it moves further.
- 4
Approval queue
Routes escalated drafts to an authenticated clinician with the exact content to be sent.
- 5
Delivery enforcement
Refuses to send anything without a matching, unexpired approval for that exact content.
- 6
Evidence store
Retains each decision as a durable record outside the application that produced the message.
- 7
1. Classify
Determine message category and whether the draft contains PHI or clinical content.
- 8
2. Evaluate policy
Match the classification and context to a written rule and select the outcome.
- 9
3. Redact where required
Mask PHI that is not permitted, and keep both the original draft and the redacted version in the record.
- 10
4. Escalate
Send the draft to a named clinician queue. Block delivery while the decision is pending.
- 11
5. Approve, edit, or reject
An approval binds to the exact content. Any edit creates new content that needs a new decision.
- 12
6. Deliver and record
Send only against a valid approval and write the full decision to the evidence store.
Capture an approval record an auditor can use
- Approver unique identifier and authenticated session
- Timestamp of the decision
- Meaning of the action: reviewed, approved, edited, or rejected
- AI draft, redacted version, and final approved content
- Policy and rule version applied, and the outcome selected
- Redactions applied and the reason for each
Decide which messages require approval
Start with an inventory of every AI-drafted message type in use: portal replies, test result explanations, administrative notices, and any others your deployments produce. The ONC SAFER Clinician Communication guide covers patient portal communication and now addresses AI-enabled systems, which makes it a useful self-assessment reference for this inventory. It is guidance, not a binding rule on clinician sign-off.
Routing criteria are an organizational decision made by clinical governance and compliance. A common starting position is that anything containing clinical advice, result interpretation, or PHI requires approval, while a narrow set of templated administrative notices may be allowed without it. Write the criteria as machine-evaluable conditions: message category, presence of PHI, patient context, and the action the draft would take. A criterion that exists only in a policy document cannot be enforced at runtime.
Default to approval when classification is uncertain. The cost of an unnecessary review is a delay. The cost of an unreviewed clinical message is much higher.
Close the bypass paths and failure modes
Most sign-off failures are not model failures. They are paths that nobody inventoried. Review each of the following as a design check and test it directly.
- Alternative send paths: Direct API calls, EHR settings, and administrative tools that can deliver a message without passing the gate.
- Auto-send fallback: Timeouts or queue failures that release an unreviewed draft. Fail closed so a pending draft stays undelivered.
- Approver identity gaps: Shared logins, service accounts, or delegated approvals that cannot be tied to one clinician.
- Edits after approval: Any change to approved content must invalidate the approval and trigger re-review.
- PHI in the draft: Apply redaction before the draft is displayed or stored in places the policy does not cover.
- Policy drift: Record the rule version with each decision so changes to routing criteria remain traceable.
Test for enforcement, then decide how to build it
The test is simple. Attempt to send an unapproved draft through every path you identified and confirm that delivery is blocked each time. Then edit an approved message and confirm the approval no longer applies. Finally, confirm the audit record shows who approved which content under which policy. If any of these fail, sign-off is being recommended, not enforced.
Trussed enforces policy in the execution path for individual interactions and tool calls, with allow, redact, block, rate limit, and escalate for human approval as outcomes. It generates audit-ready evidence from that enforcement, mapped to HIPAA, and can be self-managed in the customer cloud with configurable retention and data residency. It works at the interaction level. It does not provide transaction-level governance or end-to-end tracing of agentic workflows, so your own design should define how a message-level approval relates to the wider clinical workflow.
Whether you build the gate or adopt one, the evaluation criteria are the same: no path to delivery without a decision, individual approver identity, content bound to approval, and evidence that survives an audit.
Approval gate between draft and delivery
Draft
AI generates a portal reply or result explanation.
Policy decision
Classify the message and evaluate policy: allow, redact, block, or escalate.
Clinician approval
A uniquely identified clinician approves the exact content.
Delivery
The message is sent only against a recorded approval.
Evidence
Each decision is recorded for audit review.
Check whether sign-off is enforced or only expected
Walk through your delivery paths and approval records with a Trussed expert.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.