From dial to dispatch.

Each bord site has at least one Twilio DID, country-matched. plombiersgeneve.ch uses a Swiss +41. balitennislesson.com uses an Indonesian +62. The number on the site is the number Twilio rings. For Switzerland specifically, bord chains the Swiss mobile (Digital Republic eSIM, CHF 13/mo) to a +41 22 539 45 12 Twilio DID, then to ElevenLabs. So even calls to the personal Swiss number land on the AI agent.

ElevenLabs has nine production agents. Each agent has:

• A voice (Eric for Swiss, others per market).
• A model (gpt-4o where dispatch fidelity matters; gpt-4o-mini elsewhere).
• A prompt with FAQ and dispatch rules.
• One tool: dispatch_lead.
• An initial_wait_time = 3s so it doesn't talk over the "hello".

No closing phrase before the tool fires. If the agent says "goodbye" before calling dispatch_lead, the lead is lost. That single rule needs explicit emphasis in the prompt.

1. 01
   dispatch_lead is called

   Carries: customer name, phone, language, requested service, location.
2. 02
   Lead row written to Supabase

   Status = pending. UTC timestamp. Source site URL. Conversion tracker linked.
3. 03
   Provider lookup

   Service category × city → roster query → first available match.
4. 04
   Three-way notify

   Provider gets WhatsApp template, email, SMS. Customer gets a confirmation SMS. localclark@gmail.com gets a copy of everything.
5. 05
   Post-call summary email

   ElevenLabs sends a separate post-call email with full transcript, sentiment, next-step suggestion.

The two failure modes that actually happen: (a) the agent fires dispatch_lead but the provider phone is wrong; (b) the agent never fires dispatch_lead because the prompt got too soft. Both are caught by the quality scanner described in the parent chapter.

Production-critical: this whole pipeline runs on Vercel and Twilio, never on the founder's Mac. Lost 24h of dental leads in April 2026 when one polling job ran on launchd and the laptop slept. That rule is now hard.