A symptom-triage assistant that routes patient enquiries to the right specialist before they even book — built with hard guardrails and an offline review trail.
Brief
A four-branch clinic chain was losing patients to early friction: people described their symptoms, got routed to the wrong consultant, and bounced. They wanted an AI front door — but one that *never* gave clinical advice and *always* deferred to a human.
What we built
- A symptom intake form that uses a structured-output LLM call to classify the enquiry into a specialty + urgency band.
- A routing layer that hands the enquiry to the correct branch + specialist queue.
- A review log so a senior clinician can audit every classification daily.
- A safety net: any enquiry that mentions chest pain, breathing trouble, or self-harm bypasses the AI entirely and gets escalated to a human within 60 seconds.
Outcomes
- Booking-to-consultation match rate up from 71% to 96%.
- Average time-to-correct-specialist down 4.2×.
- Zero clinical-advice incidents in the first 90 days of operation.
What we didn't do
We deliberately did not ship a chatbot that gives medical advice. The platform's only job is to route — every clinical word comes from a human.