Patient Question Triage Agent
Sorts inbox and voicemail questions, drafts pharmacist-reviewable answers — never sends clinical advice unreviewed.
A busy independent pharmacy fields a steady stream of patient questions through voicemail and portal messages — "is my refill ready," "does this interact with ibuprofen." This agent reads each incoming question, sorts it into routine/logistical vs. clinical, and drafts a ready-to-send answer for the routine ones (pickup timing, refill status, hours) while immediately routing anything with a whiff of clinical content — dosing, interactions, side effects, "is this normal" — straight to the pharmacist queue with zero drafted answer, because that line is never the agent's to answer. Every drafted response, routine or not, still gets a pharmacist or trained tech's eyes before it goes out.
What this agent takes off your plate
- Routine logistical questions (pickup timing, refill status, hours) get a fast, drafted answer instead of sitting in a callback queue
- Any clinical question — dosing, interactions, side effects — routes straight to the pharmacist with no drafted answer attempted
- A sorted, prioritized queue instead of a flat pile of voicemails and portal messages
- Pharmacist reviews every outgoing message, routine or clinical, before it sends
- A documented pattern of common question types you can use to fix root causes (e.g. a confusing portal flow generating repeat questions)
The stack
Real costs & ROI
- Claude or ChatGPT with a signed BAA (Enterprise plan or API build)Enterprise / API pricing — confirmAs of writing, neither vendor includes a BAA on Pro/Team/Business seats — a BAA is available on Enterprise plans or via an API-based build. Expect materially higher cost than the consumer tiers and get the BAA in writing before any real data is entered.
- Voicemail transcription tool$0/momost modern phone/PMS systems already include basic transcription at no extra cost
- Zapier or Make (paid tier)$0–10/mofree tier may suffice for lower message volume; busier pharmacies will want the paid tier
The math
Time it before you build it: for one week, have whoever clears voicemail and portal messages log how long they spend sorting what came in and drafting the routine replies. If that comes to 5–8 hours a week, it's worth $150–280/week at a blended $30–35/hour rate. Automating the sort and drafting the routine-question responses is aimed at cutting that meaningfully — measure it by timing your triage for a week before and a week after go-live, and compare against your actual BAA-covered AI pricing rather than a consumer-seat price. The bigger, harder-to-price benefit is response speed on the clinical side: because routine noise is filtered out automatically, genuinely clinical questions reach the pharmacist faster instead of waiting behind a pile of "is my refill ready" messages.
Build steps
Build your topic classification rules
Define exactly what counts as "routine" vs. "clinical" before any automation touches a real question — this is the single most important judgment call in the whole build, and it's worth getting right on paper first.
Data you need
Gather a sample of 15–20 recent question types your pharmacy commonly gets (topics only, not real patient content) — e.g. "is my refill ready," "can you switch my pickup day," "does this interact with my other medication."
I run an independent pharmacy and want to build classification rules for sorting patient questions into "routine" (safe for a drafted first-response) vs. "clinical" (must go straight to the pharmacist, no drafted answer). Common question types we get (topics only): [LIST — e.g. "refill status," "pickup timing," "insurance on file," "does X interact with Y," "is this side effect normal," "can I take this with alcohol"] Please build: 1. **A routine category list** — question types safe for a drafted, pharmacist-reviewed first-response (logistics only: refill status, pickup timing, hours, general insurance-on-file questions that don't involve clinical judgment) 2. **A clinical category list** — question types that must route straight to the pharmacist with zero drafted answer (dosing, interactions, side effects, "is this normal," anything involving another medication or condition, anything a reasonable person would call medical advice) 3. **An ambiguous-case rule** — when a question could plausibly be either, default to clinical/escalate every time 4. **A one-page reference table** — question type, category, action (draft vs. escalate) This becomes the classification rule set the triage agent uses on every incoming question.
Confirm your BAA and connect the voicemail and portal intake feed
Confirm a signed BAA is active with your AI vendor before any real patient question content is processed. Then build the intake. Trigger → AI action → delivery for the voicemail side: Trigger: if your phone system is RingCentral, use the Zapier trigger "RingCentral — New Voicemail" (it includes the transcription text). For Vonage, Weave, or any system that emails you a transcript, use "Gmail — New Email Matching Search" with a search like from:voicemail@yourphonesystem.com (Gmail must be a BAA-covered Google Workspace account; for Microsoft 365 use "Microsoft Outlook — New Email" with a folder rule). Systems that only offer audio and no transcript: "Google Drive — New File in Folder" on the voicemail-audio folder → "OpenAI — Create Transcription" (Whisper) is possible only if that API account is under your BAA; otherwise this is a manual step — a tech types a two-line summary into the intake sheet. Portal messages: most PMS patient portals have no Zapier integration, so this is a manual step — the tech pastes each message into the same intake sheet, or forwards it to the same BAA-covered Gmail label so the same trigger picks it up. AI action: "Formatter by Zapier — Text → Extract" is optional to pull the caller's callback number out of the transcript, then "Anthropic (Claude) — Send Message" or "OpenAI (ChatGPT) — Conversation" (API key from the BAA-covered account) with the system prompt below and the transcript as the user message. Delivery: handled in Step 4 with two paths. Make equivalents: "Gmail — Watch Emails" → "Anthropic Claude — Create a Message" → "Router".
Data you need
Loop in your pharmacist-in-charge to confirm and approve BAA coverage before go-live. You'll need admin access to your phone system's voicemail settings, a BAA-covered Google Workspace or Microsoft 365 mailbox, and your Zapier/Make account. Check the automation platform's BAA status too — as of writing neither Zapier nor Make offers one on standard plans; if you can't get one in writing, the intake step is a tech pasting transcripts into the BAA-covered AI tool by hand.
Agent Pack
Unlock the Pharmacy Agent Pack.
One payment unlocks all 3 blueprints in this pack — build steps, the complete system prompt, guardrails, a test checklist, and the maintenance routine — for $29.
one-time, unlocks all 3 blueprints in this pack on this browser. Secure checkout via Stripe.
More in this pack
More for Pharmacy
Refill Reminder Agent
Refill-due outreach and adherence nudges, drafted from fill dates only — never diagnosis.
Prior-Auth Paperwork Agent
Drafts the clinical-justification narrative and tracks follow-ups around your CoverMyMeds / Surescripts ePA, inside a BAA-covered tool only.