AI voice agents: one that makes the calls, one that takes them

Problem

Trades businesses miss calls while they're on a job, and cold-calling them by hand doesn't scale.

What I built

I built both agents and the n8n workflows behind them. Every prompt version is run against test calls and reviewed; one review caught and fixed time-zone double-booking and spelled-out email capture.

• Outbound (“Ava”): an ElevenLabs voice agent rings trade-business owners and books a 15-minute call straight into Google Calendar, with a Meet link. It always says it's an AI when asked, and handles opt-outs on the spot.
• An n8n daily caller runs on a morning schedule, with hard caps per run. It picks the next lead from the sheet and starts the call.
• A post-call handler parses each conversation, writes the outcome back to the sheet, alerts me in Slack on a positive call, texts the demo link via Twilio, then triggers the next call.
• Inbound (AI receptionist): an ElevenLabs agent on a Twilio number that answers as the business's own receptionist, built as one template where each business's services, hours, emergency policy and booking rules are plugged in as variables. It never invents company facts.
• It answers questions about services and pricing, checks availability, and books, reschedules and cancels appointments, taking name, address, postcode, problem and phone number without asking twice.
• Emergencies override the normal flow (a gas smell stops the booking script), and it says it's an AI whenever it's asked.

Result

• The outbound caller has been in production since Oct 15, 2026.
• The receptionist runs as a live demo. No client deployment or results are published yet; numbers get added only when they're real.