Christos Melissourgos

AI voice agents look simple in a demo and collapse in production — and the reasons span every layer, not just one. The LLM itself breaks (hallucinated policies, regressions on model upgrades, dialect and accent gaps, misread intents). The telephony layer breaks (SIP quirks, failed transfers, voicemail detection). The knowledge layer breaks (stale FAQs, missing edge cases). The action layer breaks (wrong CRM field updated, broken workflow triggered). And the feedback loop is usually missing entirely, so none of the above gets caught until a customer complains.

This talk introduces a five-layer build framework — Scope, Voice, Brain, Hands, Loop — that maps every production voice AI agent onto the same anatomy, regardless of vendor or stack. Each layer has its own failure modes, its own build order, and its own minimum-viable version.

Drawn from two years of running voice agents in production across collections, e-commerce, booking, and support, the framework gives teams a shared vocabulary for scoping, building, and evaluating voice AI — whether they build in-house, hire a vendor, or combine both.