
Ninety seconds on the architecture of an AI sales agent I put on a live WhatsApp channel for a DTC medical-device business, 13 days from idea to production. The model reasons and our ERP holds the facts. Every message passes a gate, the agent pauses itself when a human replies, and a person can take over any conversation in one click. It recovered about 45% of warm leads. Together with better ad targeting, return on ad spend went from 1.8x to 7x.
Narrated in my own voice, cloned with ElevenLabs.
Read the full case study, with the architecture
Transcript
Someone messages at 11 at night, ready to buy. By morning, the lead is cold. So I put an AI sales agent on our live WhatsApp channel, 13 days from idea to production. The model was the easy part. The guardrails were the product.
Everything runs through one middleware. One webhook in, one process to trust. The model reasons, and the ERP holds the facts. Rule one: the model never authors a fact of record. Prices, stock, names and addresses all come from the ERP through tool calls.
Every message passes a confirmation gate before it sends. The moment a human replies, the agent pauses itself. A stoplight console shows every conversation. Green runs on its own, amber needs a look, and red goes to a person, with one-click takeover.
Where it matters, nothing is free text. 13 handoff reasons, 17 event types, and one append-only audit trail.
It runs around the clock and recovered about 45% of warm leads. Together with better ad targeting, return on ad spend went from 1.8x to 7x. If you want AI on a customer channel without betting the brand on it, let's talk.