Turn a message into a proposed record update
Keep the original message, AI draft, clinician decision and approved record change distinct.
Draft a change instead of silently replacing a fact
Give the assistant a fictional message and the relevant fictional record excerpt. Ask for the source and date, a concise patient-reported summary, what conflicts with the existing record and what requires confirmation. Show the proposed change beside the existing text.
For Maya, “I stopped the cholesterol pill” is a patient report. It should trigger reconciliation rather than silently removing a medication or recording a confirmed drug adverse effect. Preserve the original message and the uncertainty.
Keep approval at the clinical boundary
A clinician confirms the facts and decides which change belongs in the record. A staff member may prepare the draft within their role; the AI does not choose treatment, approve a medication change or send a patient reply on its own.
For a practical pilot, begin with drafts that a person manually reviews and enters in the approved clinical system. This keeps the review step explicit while you learn whether the draft helps.
An integration needs its own verification
Before enabling automated record writes, test the correct patient and encounter, access controls, duplicate messages, failed writes, retry behavior and the audit history. Confirm the saved result, not just a success message. Make a failed or overdue review visible to a named owner.
The presentation illustrates this workflow; it does not demonstrate a deployed clinical integration. Permission to read a message is separate from permission to edit a chart or send a reply.