Patient Message Reply AI Agent: Drafting In-Basket Replies, Never Sending Them

Portal messages pile up in every clinic's EHR in-basket. A patient message reply AI agent can take the slow part of that work, reading the message, pulling the chart context and writing a first draft, while a clinician keeps the one action that matters: pressing send.

This note summarises how Nestack Agent Care describes its patient-message reply-drafting agent.

What the agent does

• Takes the message as written. It reads the patient's own words, language and urgency.
• Retrieves only the context the reply needs. That means the last visit, active medications, recent results, pending orders and the care plan for that patient.
• Sorts and routes. It files each message under the clinic's own categories (clinical, refill, results) and sends it to the pool that owns it.
• Drafts from approved sources. Replies are written only from what the record and the clinic's approved answer library support, at the reading level the clinic's policy sets.
• Checks refills against protocol. It compares a refill request with the last visit and the monitoring the drug requires.

Red flags go to a person first

Before any draft exists, the message is checked against the clinic's own red-flag wording. A match leaves the queue and is surfaced to a person ahead of everything else. The agent does not assign acuity. It also holds back anything that would interpret a result, change a treatment, or answer something the record does not cover.

A human sends every reply

The draft sits in the in-basket next to the message and the chart context it used, with gaps clearly marked. The agent has no send action. A clinician or the owning pool reads, edits and sends, or reassigns or discards the draft. Every message, source read, draft, edit and sender is retained for audit.

How outcomes are evaluated

Nestack evaluates the agent on urgent messages caught, unsafe drafts, how much editing happens before sending, wrong-context retrieval and subgroup results. Typical delivery is 4 to 6 weeks on the client's own stack, followed by ongoing Agent Care.

Read the full playbook on the Nestack patient-message reply-drafting agent page, or see more healthcare AI agents from Nestack.