This is a placeholder post written in the voice and structure we'll use for real case studies. Replace bracketed sections with real client details (or anonymize). See /blog/_template.html for the empty template.
The setup
[Client] is a [size]-staff medical clinic doing roughly [N] patient interactions a month. Front desk was drowning — same questions on repeat, ten times a day. "What insurance do you take?" "When are you open Saturdays?" "How do I cancel?" "Where do I find the form for X?"
They'd tried two chatbot platforms before us. Both failed for the same reason: generic templates don't know what makes their clinic different. The bots confidently gave wrong answers, patients got annoyed, and the front desk had to fix the mess on top of the original questions.
The problem we actually solved
The chatbot question was a symptom. The real problem: the clinic's operational knowledge lived in three places — the front desk team's heads, a partial FAQ document last updated 14 months ago, and a Google Doc nobody read. No agent could be useful until that knowledge was captured.
The training week
Before writing a line of code, we shipped the front desk team five small voice recorders. The instruction was simple: narrate what you're doing as you do it. When a patient asks something, before you answer, say out loud who's asking, what they want, and how you decide what to tell them.
We collected roughly 14 hours of audio over five days. Transcribed it. Tagged it. What emerged wasn't just FAQs — it was the judgment layer: how the team decided to route, escalate, or defer questions. That layer is what generic chatbots can't reproduce. It's what we trained the agent on.
The founder also spent a half-day on-site. Watched intake. Sat through three patient calls. Noticed two workflows that the team didn't even mention — they were too routine to think to describe.
What we built
- Inbound chat agent on the WordPress site, trained on the transcribed knowledge plus the official FAQ.
- Escalation logic: anything medical, anything billing-specific, anything emotional — straight to a human within 30 seconds, with full conversation context attached.
- SMS bridge via Twilio so the same agent handles questions that come in as text.
- n8n flow behind it routing tickets, logging confidence scores, and surfacing weekly "agent didn't know" reports for the founder to review and fold back into training.
Results after 6 weeks
- 40% of inbound questions resolved before a human saw them. Target was 30%.
- Patient satisfaction on agent-handled conversations: 4.6/5. Higher than human-handled, which surprised us.
- Front desk reclaimed roughly 12 hours/week — redirected to outbound recall calls that actually book appointments.
What we'd do differently
The first week the agent went live, confidence thresholds were too generous. It tried to answer questions it shouldn't have — including one about a medication interaction. We caught it in monitoring within 4 hours, no patient harm, but it was a real reminder: escalation defaults should be aggressive on day one and only loosen with evidence.
We'd also have run the training period one more week. The judgment-layer audio kept revealing useful patterns up to the last day.
Want an agent trained on how your business actually runs — not a generic template? See our AI Automation approach, or tell us what you're trying to ship.