62% of support tickets resolved by AI chatbot

The situation

The platform served 40,000+ active patients across 20 states. Support volume had tripled since a 2024 insurance-coverage expansion, and the team had grown from 6 to 14 agents just to keep up with Tier-1 volume — scheduling, password resets, insurance eligibility, "when does my therapist get back to me."

Leadership wasn't going to 2× headcount again. They needed the queue to stop growing.

What we built

A HIPAA-safe chatbot on two surfaces: the patient portal (text widget + voice) and a dedicated SMS number. Both backed by the same retrieval-augmented generation (RAG) pipeline reading their help-center KB + a curated set of policies.

Intent classification first, generative second. Every inbound message is first classified (scheduling / password / insurance / clinical-question / billing / other) by a small open-source classifier. Scheduling and password requests skip the LLM entirely and call the platform's own APIs. Insurance and billing route to an LLM with retrieval against the KB. Clinical questions NEVER go to the LLM — they route to a human immediately with "this is a medical question, a person will be with you in a moment."

Human handoff is one tap away. Every bot surface has a persistent "Talk to a person" button. The bot never argues with a user who wants a human.

Audit logging per HIPAA. Every message + classifier output + LLM response + resolution status logged to an encrypted append-only store, retained per the client's BAA requirements.

Results after 90 days

• 62% of Tier-1 ticket volume resolved by the bot. Zero clinical escalations mishandled.
• Median resolution time for bot-resolved tickets dropped from 14 minutes (human queue wait) to 38 seconds.
• CSAT on bot resolutions landed at 4.6/5 after 60 days of tuning.
• The support team reallocated 84 hours/week to retention outreach + clinical coordination — work that actually moves patient outcomes.

What we'd do differently

The hardest single decision was the clinical-handoff guardrail. Early versions let the bot answer soft clinical questions ("is it normal to feel anxious?"). We turned that off entirely within two weeks — the liability and trust risk wasn't worth the efficiency gain. That discipline is why the rollout survived legal review.