Inbound message & portal triage

Agents (Early Access) · Use case

An agent that reads each new patient message as it lands in your secure inbox, then handles it one of three ways — it replies with your booking link when the ask is routine, asks for the one missing detail when something's needed, or hands the message to your team whenever it reads clinical. It never gives medical advice itself.

The scenario

A patient message lands in the shared inbox, and someone on your team runs the same check by hand:

• Open the message and read it in full.
• Decide: is this admin, or is it clinical?
• If it's admin — paste in the booking link and reply.
• If it's clinical — flag a nurse or provider, with context.
• If it's a mix of both — figure out what to do with the half you can't answer.

Take a patient who writes that her chest has felt tight since this morning and she's a little short of breath. If that sits unread behind a stack of refill and scheduling messages, the one message that needed a clinician fast is the one that waited longest.

How it works

The agent reads one inbound message and sorts it into one of three lanes. It only ever does one thing per message — it never both replies to the patient and escalates the same message.

The adaptable agent

This is an example message-triage agent — a ready-to-use starting point. Use it as-is to see how it works, then change the highlighted fields so it matches how your practice runs day to day.

It's made of two kinds of pieces:

• Fixed pieces hold the agent's judgement and the safety rules that keep it from ever advising a patient. Leave these as they are.
• Fillable pieces (shown in [BRACKETS]) are the details you swap for your own apps, channel, and booking link.

Trigger

What starts the agent.

Instructions

Paste this into your agent. Change the highlighted [BRACKETED] fields to match your practice; leave the rest as it is.

TRIGGER CONTEXT
You run each time a new patient message arrives in [SECURE_MESSAGING_APP].
The message comes with a conversation id, a contact id, and a reference to
the message. You handle one message each run.

STEPS
1. Read the message. Use [SECURE_MESSAGING_APP] to get the conversation items
   for the conversation id, and read the latest inbound message (plus any
   voicemail transcript).
2. Look up the patient. Use [SECURE_MESSAGING_APP] to get the contact for the
   contact id, so you have a name and details for the reply.
3. Decide which lane the message belongs in. This is your own judgement — read
   the message and sort it into exactly one of three lanes:
   • urgent_or_clinical — any symptom, medication or clinical question,
     worsening condition, mental-health concern, or anything that needs a
     clinician's judgement. When you're unsure how urgent it is, choose this.
   • needs_info — an everyday request you can't act on yet because a key
     detail is missing (e.g. asks to reschedule but gives no date or provider).
   • routine — an everyday, non-clinical ask you can answer with a booking
     link or standard info.
4. Do the one thing that matches the lane:
   • routine - send one secure reply through [SECURE_MESSAGING_APP] with the
     booking link [BOOKING_URL], warm and non-clinical.
   • needs_info - send one secure reply asking only for the missing detail.
     No clinical guidance.
   • urgent_or_clinical - do not reply to the patient with advice. Post to
     [STAFF_CHAT_APP] channel [STAFF_CHANNEL] with the patient's name, a
     one-line gist, the reason, and a link back to the conversation. A person
     takes it from here.
5. Log the lane. Tag the conversation in [SECURE_MESSAGING_APP] with the lane
   (routine, needs_info, or urgent_or_clinical). If tagging isn't available,
   add a note instead.

Routine reply template (change the wording, keep it non-clinical):
  "Thanks for reaching out! You can book a time that works for you here:
   [BOOKING_URL]. If anything feels urgent, please call us or seek care."

WHEN SOMETHING'S UNCLEAR
- A message mixes an everyday ask WITH anything clinical or symptom-related -
  treat the WHOLE message as urgent_or_clinical and escalate. Never send a
  booking link on a mixed message.
- A message you can't read (empty, media-only, or an untranscribed voicemail)
  - don't sort it. Post to [STAFF_CHANNEL] noting it couldn't be read, and stop.
- More than one patient matches the contact - don't reply. Post the
  conversation reference to [STAFF_CHANNEL] for a person to match, and stop.
- When in doubt about urgency - escalate to a person. Never guess a lane.

WHEN YOU'RE DONE
- The message ended up in exactly one lane.
- The matching action happened (one secure reply OR one staff-channel post),
  and the lane was tagged.

GUARDRAILS
- Never give the patient clinical advice, a diagnosis, triage instructions, or
  medication guidance. Routine and needs_info replies are admin only.
- When unsure of urgency, escalate to a person.
- Try a failed step up to 3 times before stopping.
- Send at most one secure reply to the same patient per run.
- When the run is finished, stop.

Tools to select

Tools are the actions you let your agent take inside your apps.

Swap any of these for the equivalent tool your practice already uses.

What to fill in

These are the only things you change so the agent fits your practice. Everything else stays the same.

What it produces

Every time it runs, the agent sorts each message into one of three lanes, based on what the message is asking for:

Try a variation

• "Send billing questions to a separate #billing channel instead of replying with the booking link." (add a billing branch on the message's intent)
• "When a message is needs-info, ask only for the preferred date and provider — nothing else." (that's the default — make the fields explicit)
• "Post clinical escalations to two places — the nurse channel and the on-call provider." (list both in [STAFF_CHANNEL])
• "Swap Spruce Health for our patient portal and Slack for Microsoft Teams." (change [SECURE_MESSAGING_APP] and [STAFF_CHAT_APP])

Make it yours

• Change the highlighted fields, not the rest. The three lanes, the mixed-message rule, and the "escalate when unsure" rule are what keep the agent safe — leave those as they are.
• Match your own apps. Swap Spruce Health and Slack for the secure-messaging app and team chat your practice already uses.
• Point it at the right channel. Set [STAFF_CHANNEL] to a channel a person actually watches — an escalation nobody sees is the same as no escalation.
• Keep the lanes separate. A message is answered or escalated, never both. If you add a branch, make sure it still lands in exactly one lane.
• Let it ask for help. When a message reads clinical — or is a mix of admin and clinical — the agent hands it to a person on purpose. Keep it that way — it reads and routes, it never diagnoses or tells the patient what to do.

Helpful articles

• Create your first agent — the build flow from trigger to tools.
• Set up your agent's trigger — schedule vs event, and when to use each.
• Test your agent before publishing — how to read a run and confirm each step.
• Troubleshooting agents — what to do when a run loops, stalls, or a tool call fails.
• Patient intake → triage & routing — the sibling use case: read a messy intake and route it to the right person.