If you run a clinic prescribing GLP-1 medications — Ozempic, Wegovy, Mounjaro, Zepbound — you already know the support problem. These medications work, but patients have questions. A lot of questions.
"My dose was increased and I feel nauseous — is that normal?" "Can I drink alcohol on semaglutide?" "I've been on it for six weeks and the scale hasn't moved — should I be worried?" "I'm traveling next week — how do I store the injection?"
These questions flood your phone lines, your patient portal, and your staff's email. Most of them have straightforward answers. But each one takes 5-10 minutes of a nurse's time — and nurses have better things to do than answer "Can I take Tylenol with Ozempic?" for the fourth time today.
I built a GLP-1 coaching bot for a weight loss clinic that handles these questions 24/7, gives patients evidence-based answers, and only escalates to clinical staff when it genuinely needs to. Here's the step-by-step.
Step 1: Define the Scope (What It Does and Doesn't Do)
This is the most important step and the one most people skip. A coaching bot for GLP-1 patients is not a diagnostic tool. It does not prescribe. It does not adjust dosages. It does not replace the provider-patient relationship.
What it does:
- Answers common questions about GLP-1 medications (side effects, storage, timing, diet, exercise, interactions) - Provides encouragement and coaching between appointments - Tracks patient-reported symptoms and flags concerning patterns for clinical review - Reminds patients about injection schedules, upcoming appointments, and lab work - Educates patients about what to expect at each dosage level
What it never does:
- Diagnoses conditions - Recommends medication changes - Tells a patient to stop or change their medication - Provides specific medical advice that should come from their provider
I built a "boundary document" — a clear specification of what the bot can and cannot discuss. This document gets embedded in every prompt as a system instruction. If a patient asks something outside the scope, the bot says: "That's a great question for your provider — I'll flag it for your care team so they can follow up at your next visit."
Step 2: Build the Knowledge Base
The knowledge base for a GLP-1 coaching bot needs to be medically accurate and kept current. Here's what I included:
- Prescribing information for semaglutide and tirzepatide (from manufacturer documentation) - Common side effects by dosage tier — patients on 0.25mg semaglutide have different concerns than patients on 2.4mg - Dietary guidance specific to GLP-1 patients (smaller portions, protein-first eating, hydration) - Exercise recommendations appropriate for patients at various stages - Drug interactions — the known interactions and common OTC medications patients ask about - Storage and handling — temperature requirements, travel tips, what to do if a dose is missed - The clinic's specific protocols — their titration schedule, their dietary program, their follow-up cadence
I worked with the clinic's medical director to review every piece of content. This isn't optional. A coaching bot that gives wrong medical information is a liability, not an asset. The review took about three hours and was some of the most important time spent on the project.
Step 3: Build the Conversation Layer
I used Claude as the conversational engine. Here's why Claude specifically: it's excellent at maintaining a warm, coaching tone while staying within defined boundaries. It doesn't try to be a doctor. When prompted correctly, it's remarkably good at saying "I understand your concern — here's what I can share, and here's when you should talk to your provider."
The prompt architecture has three components:
1. System prompt — Contains the boundary document, the clinic's protocols, and the bot's persona (warm, encouraging, knowledgeable but not clinical) 2. Knowledge retrieval — When a patient asks a question, the system searches the knowledge base and injects the most relevant content into the conversation context 3. Conversation memory — The bot maintains a patient profile including their current medication, dosage, start date, reported side effects, and goals. This lets it give contextual responses: "Since you just moved up to 1.0mg two weeks ago, some increased nausea is common. Here's what usually helps..."
Step 4: Build the Symptom Tracking and Escalation System
This is where the bot becomes genuinely valuable to the clinical team — not just the patients.
When a patient reports a side effect, the bot logs it: symptom, severity (the bot asks them to rate 1-10), duration, and whether it's impacting daily activities. This data feeds into a dashboard that the clinical team reviews daily.
The escalation rules are tiered:
- Green: Common side effects within expected parameters — bot handles independently with guidance - Yellow: Side effects that are persistent, worsening, or higher severity — bot provides guidance AND flags for clinical review within 24 hours - Red: Symptoms that could indicate a serious adverse reaction (severe abdominal pain, signs of pancreatitis, severe allergic reaction) — bot immediately directs patient to call the clinic or 911, and sends an instant alert to the on-call provider
Getting the escalation thresholds right took two full rounds of review with the medical director. We erred on the side of over-escalation initially and loosened it as we gained confidence.
Step 5: Deploy and Integrate
The bot runs as a web chat widget on the clinic's patient portal and as an SMS interface via Twilio. Patients can reach it 24/7 from either channel.
The tech stack: - Claude API for conversation - Pinecone for knowledge base vector search - n8n for workflow orchestration (symptom logging, escalation routing, appointment reminders) - Twilio for SMS - Next.js for the web chat widget - Supabase for patient profiles and symptom tracking data
The full build took about five days: two days on the knowledge base and clinical review, two days on the technical build, and one day on testing with the clinical team.
The Results
After 60 days:
- Clinic call volume dropped 40%. The majority of calls had been medication questions that the bot now handles. - Patient engagement increased. Patients interacted with the bot an average of 3.2 times per week — far more than they'd ever call the clinic. - Symptom tracking gave the clinical team better data. Instead of relying on patient memory at appointments, providers had a longitudinal record of side effects and their trajectory. - Patient satisfaction scores for "support between visits" went from 3.1 to 4.5 out of 5. - No adverse events were missed. The escalation system caught every case that needed clinical attention — and caught some earlier than the old call-in system would have.
What I'd Do Differently
The biggest mistake was launching with too much content in the knowledge base. The bot would sometimes give overly detailed answers when a patient just wanted a quick yes or no. "Can I have coffee on semaglutide?" doesn't need a three-paragraph response about caffeine metabolism. We tuned the prompts to favor concise answers with an option to "Tell me more" if the patient wants depth.
I'd also invest more time in the persona from day one. The first version was helpful but bland. Patients respond better when the bot has a personality — encouraging, slightly casual, maybe a touch of humor. "The nausea is your body's way of adjusting — you're doing great. Here's what helps most patients get through the first couple weeks..." beats "Nausea is a common side effect. Consider the following strategies."
FAQ
Is a coaching bot considered a medical device? In most jurisdictions, no — as long as it provides general wellness information and educational content without diagnosing, prescribing, or providing personalized medical advice. However, regulations vary and you should consult healthcare legal counsel. The boundary document is critical for keeping the bot in the "wellness tool" category.
How do you keep the knowledge base medically accurate? We update the knowledge base whenever new prescribing information is released or the clinic changes its protocols. The medical director does a quarterly review of the bot's most common responses. We also track any corrections made by clinical staff when they review flagged conversations.
Can patients use the bot instead of going to their appointments? No, and the bot actively discourages this. It sends appointment reminders and, when answering questions, often says "This is a great topic to discuss with your provider at your next visit." The bot supplements care — it doesn't replace it.
What if the bot gives wrong information? Every response includes a disclaimer that the bot provides general information, not medical advice. The symptom tracking system ensures that concerning situations always get human review. In 60 days, we had two instances where the bot's answer was technically correct but could have been more nuanced — both were caught in weekly reviews and the prompts were updated.
How much does it cost per patient per month? Roughly $2-4 per active patient per month in API and infrastructure costs. At the clinic's scale of 200 active GLP-1 patients, that's $400-800 per month — trivial compared to the staff time it saves and the patient retention it drives.
Your GLP-1 patients need support between visits.
If you're prescribing GLP-1 medications and your phone lines are jammed with patient questions, we can build a coaching bot that gives your patients 24/7 evidence-based support while freeing your clinical team to focus on care delivery.
Book a Strategy Audit and we'll design a coaching system tailored to your protocols and patient population.