The old chatbot had one job — deflect tickets — and it did it by getting in the way. It answered confidently whether or not it knew, made customers repeat themselves, and hid the “talk to a human” button three clicks deep. Deflection numbers looked great on a dashboard. CSAT told the real story.
The problem, precisely
In healthcare support, a wrong answer isn’t just annoying — it can be a compliance problem. And a support experience that traps people erodes trust faster than slow replies ever did. The goal was never “deflect more.” It was “deflect the questions we can answer well, and get everything else to a person quickly and gracefully.”
The approach
Route first, answer second. Every message is classified before the system decides what to do: answer from the docs, ask a clarifying question, or hand off. Answers are grounded in and cited from the company’s own knowledge base — no citation, no answer. Escalation is always one obvious click away.
Deflection you’re proud of comes from answering well, not from making the exit hard to find.
What we built
# routing
classify → intent + risk tier
answer → grounded, cited from own docs only
clarify → ask when the question is underspecified
escalate → visible escape hatch, warm handoff
measure → per-path CSAT + deflection quality
How we knew it worked
We instrumented CSAT per routing path, so “deflection” and “satisfaction” stopped being in tension on the dashboard. Clean deflection settled at 31%, CSAT rose 11 points, and escalations reached a human in under a minute — with a red-team suite standing guard against invented policies.