05 — CASE STUDY / 2026
Synthflow — Voice Agent
A case study on building a reliable AI voice agent: call flows, hand-off to humans, and the quality bar that makes people trust it.
How this study was done
Context
Voice agents fail quietly: callers hang up, intents get misread, and the business never sees why. I started from the call itself and what a caller expects within the first ten seconds.
Approach
I mapped the conversation flows end to end, separated model failures from product failures, and defined where the agent must hand off to a human instead of guessing.
What I proposed
A containment-versus-satisfaction framing, guardrails on high-risk intents, and an evaluation loop that scores real calls rather than scripted demos.
How to read the deck
The deck above walks through the caller problems, the flow design, the quality metrics, and the risks of over-automation.