At 9 AM on a Tuesday, the owner of a weight loss clinic called me with a familiar complaint: "We're getting plenty of leads from Google, but half of them ghost before their first appointment."
I asked her to walk me through the intake process. She emailed me a four-page PDF. Medical history, lifestyle questions, insurance details, consent forms, dietary habits — the works. Patients were supposed to download it, print it, fill it out by hand, scan it, and email it back. In 2026.
By 1 PM that same day, we had a conversational AI intake bot live on her website. Within the first week, her intake completion rate went from roughly 45% to 89%. That's not a typo.
Here's exactly how I built it.
The Problem
Weight loss clinics live and die by new patient acquisition. This clinic was spending $4,000 a month on Google Ads, generating around 120 leads. But only about 55 of those leads actually completed the intake paperwork. The rest dropped off somewhere between "I'm interested" and "here's my medical history."
The owner had tried a few things. She shortened the form. She offered to let people fill it out in the office. She had her front desk call leads to walk them through it. Nothing moved the needle significantly. The friction wasn't just the form — it was the entire experience. Nobody wants to deal with a PDF when they're already stressed about their weight.
The Build
I used Claude as the conversational backbone and deployed it as a chat widget embedded directly on the clinic's website. Here's the stack:
- Claude API for the conversational engine — it handles natural language, asks follow-up questions, and validates responses in real time - n8n for the automation workflow — when a patient completes the conversation, n8n formats the data, pushes it to the clinic's EHR system, and triggers a confirmation email - A simple Next.js chat component embedded on the clinic's booking page — nothing fancy, just a clean chat interface that feels like texting
The bot doesn't just ask the same questions the PDF did. It has a conversation. It starts with "Hey, let's get you set up for your first visit — I'll walk you through a few questions. It takes about 5 minutes." Then it asks questions one at a time, in plain language.
Instead of "Please list all current medications and dosages," it says "Are you currently taking any medications? If so, just list what you can remember — we'll confirm the details at your visit."
That shift matters more than any technology choice. People are willing to share information when the experience feels human and low-pressure.
The bot also handles branching logic. If a patient mentions they're on a GLP-1 medication, it asks relevant follow-up questions about dosage and side effects. If they mention a previous surgery, it flags that for the provider. If they're unsure about something, it tells them it's okay to skip and that the clinical team will follow up.
Building the conversation flow took about two hours. Connecting n8n to the EHR and email took another hour. Testing and refining took one more. Four hours total from start to live deployment.
The Results
In the first 30 days:
- Intake completion rate jumped from 45% to 89% - Average time to complete intake dropped from "never" (because most people abandoned the PDF) to 6 minutes - The front desk saved roughly 15 hours per week they'd previously spent chasing people for paperwork - The clinic booked 28 more first appointments than the previous month — on the same ad spend
The financial impact was immediate. At an average patient value of $1,200 for the initial program, those 28 additional patients represented over $33,000 in new revenue. The bot cost less than $500 to build and about $60 per month to run.
The Tradeoffs
I want to be honest about what didn't go perfectly.
First, the bot occasionally misunderstood medication names. "Ozempic" was fine, but less common generics sometimes got garbled. We solved this by adding a confirmation step — the bot repeats back what it heard and asks patients to verify.
Second, some older patients were uncomfortable chatting with a bot. About 11% of visitors still preferred the phone. We kept a "call us instead" button visible at all times. You can't force a channel shift — you offer a better option and let people choose.
Third, the EHR integration was fragile at first. The clinic used an older system with a limited API. We ended up routing data through a structured email template as a fallback, which their staff could import manually. Not elegant, but reliable.
Lessons Learned
1. Conversational UX beats form UX every time. People will share more information when it feels like a chat, not a government form. 2. Start with the drop-off point. The clinic's real problem wasn't lead generation — it was the gap between interest and action. Focus your AI where patients are falling out of the funnel. 3. Don't over-engineer the first version. The initial bot had maybe 40% of the features I wanted to add. Ship the simplest thing that works, then iterate. 4. Fallbacks matter. Always have a human option. Always have an error path. The bot will misunderstand someone, and that person needs a way forward. 5. Measure from day one. We tracked completion rates, drop-off points, and patient satisfaction from the first hour. That data drove every improvement.
FAQ
Does the bot handle HIPAA compliance? The bot itself doesn't store patient data. Conversations are processed in real time through the Claude API, and the structured output goes directly to the clinic's HIPAA-compliant EHR. We use encrypted connections throughout and the n8n instance runs on the clinic's own infrastructure.
How much does it cost to run? About $60 per month in API costs at their current volume, plus the n8n hosting. Compare that to the $4,000 monthly ad spend — the bot makes the ads work harder.
Can patients go back and change their answers? Yes. The bot supports "actually, let me change that" at any point. It keeps track of the conversation state and lets patients revise any answer before final submission.
What happens if the bot gets confused? It's designed to gracefully hand off. If it can't understand a response after one clarification attempt, it says "No worries — our team will follow up on this one" and flags the item for the front desk.
How long did it take to train the bot on the clinic's specific protocols? About two hours. We fed it the existing intake form, the clinic's FAQ, and a few pages of their patient handbook. Claude handles the rest through its general medical knowledge, with guardrails to keep responses within scope.
Ready to cut your patient drop-off in half?
If your clinic is losing leads between the ad click and the first appointment, the problem probably isn't your marketing — it's your intake process. We build conversational AI bots that turn website visitors into booked patients. No PDFs, no phone tag, no dropped leads.
Book a Strategy Audit and we'll map exactly where your patients are falling off and how to fix it.