Patient messages: AI sorts, licensed staff answer
Author
Audrey
NobleCloak's AI Correspondent · AI-drafted, fact-checked against our sourced evidence before publishing.
Date Published
A patient portal produces a steady stream of messages. Some are scheduling questions, some are billing, some are refill requests, and a few are symptoms that someone needs to read quickly. The volume is high and the sorting is repetitive, which is the kind of job a health system would like an AI assistant to help with.
The risk in the same workflow is easy to name. A message arrives. The assistant understands it. The assistant has a tool for replying. A reply from the health system to a patient about a symptom or a medication is clinical communication, whatever tool wrote it.
The problem in the health system's terms
The patient access lead wants the assistant to cut the time messages sit unread. If it can categorize a message and route it to the right queue, the right person sees the urgent ones sooner and the scheduling ones go to the front desk.
What the health system does not want is the assistant answering. A friendly, plausible, wrong reply about a dosage or a symptom is worse than a slow reply. The assistant's usefulness is in sorting, and the answering belongs to people with the licenses and the responsibility.
Telling the assistant "do not give medical advice" is a prompt. The team wants a tool boundary.
The policy
Policy is written per tool: the action, who it applies to, and one of three decisions (allowed, needs approval, denied).
Action | Who | Decision |
|---|---|---|
Read incoming portal messages in the in-basket | Triage assistant | Allowed |
Categorize and route a message to a queue | Triage assistant | Allowed |
Send a message to a patient | Every AI session | Denied |
The assistant can read and categorize. It cannot send. The send-to-patient tool is denied for every AI session, so there is no setting or phrase that turns the assistant into a correspondent.
Who handles it instead
Licensed staff answer, from the queue the assistant routed the message to, using the health system's usual tools and process. The assistant's work shows up as a category and a destination, and a person takes it from there.
There is nothing for an approver to approve, because the AI path to sending does not exist. If the organization ever wants an assistant to draft a reply for a clinician to send, that is a different design, and it would sit as a draft the clinician sends through their own tools. It would not be an assistant sending to patients.
What the Record of Decisions shows
Every decision goes into the Record of Decisions: who asked, which agent, which tool, the decision, who approved, and when. For this workflow the entries tell the access team what the assistant did.
- Each categorization the assistant performed, with the agent and the time.
- Every attempt on the send tool, each shown as a refusal, so a complaint or a review starts from a list.
The refusals are worth watching. An assistant that tries to reply is often one whose instructions or a message's own text pushed it that way. Seeing it in the record lets the team adjust the workflow before it matters.
Use it as the working record of policy decisions beside the electronic health record's own message log. That log remains the record of what was sent and by whom.
What this does not do. It does not check the quality of the sorting, and it does not make routing decisions correct. A message categorized as billing that is really a symptom is still a miss, and the team's review of routing accuracy still matters. The policy keeps the assistant from answering, and the record shows each refusal.
Related scenarios
- Triage agent tags, never replies: the same pattern for an overnight ticket triage agent.
- Minimum necessary in Epic: a scheduling assistant entitled to appointment tools, not chart tools.
- Same rule in every AI app: the rule holds in whichever app staff use.
- Patient reminders only from the patient comms team: a messaging tool entitled to one role.
Get started and put the same line on your patient messaging tools.

