AI voice and CRM integration

Why bolt-on diallers fail when transcripts, dispositions and AI actions must land on the same lead record.

Media vs product plane

SIP/RTP stays on FreeSWITCH (or CPaaS). Product state lives in PostgreSQL: leads, conversations, call_sessions, workflow runs.

Real-time path

SIP → FreeSWITCH → ESL → voice-orchestrator → STT/LLM/TTS adapters → tool layer (update CRM, schedule task) → async persist. Long LLM calls must not block the media thread.

Degradation

STT failure → DTMF or transfer to human
LLM failure → fallback model or scripted branch
Storage failure → stop new recordings, continue call handling per policy

Permissions

Recording and transcript access use separate permission codes from generic CRM read—see COM-008 in scope docs.

Evidence: case study voice section. Narrower SKU: AI receptionist product.

Next step

Configure modules, budget band and timeline on the product page—or read engineering evidence on the case study.

Open scope estimator Technical case study