Here is what an $18,000 AI call center for a local business really is. Seven agents. A dozen platforms running underneath: Twilio, Vapi, Google Calendar, HubSpot. It doesn't replace the phone line. It runs it. A call comes in, it answers and books. A missed call gets a text back before the caller dials the next guy. Every call lands in the CRM. And it reports what is actually turning into money. Below is how it works, written for two audiences at once: the business owner who wants to know what it does, and the engineering team who needs to know how it's built.
Module 1: Call Answering and Booking
What it does: it picks up on the first ring. At 2am, on a Sunday, during the busiest hour, every single time. It doesn't read a script, it reads the business: services, prices, hours. It answers the question, and if the caller is ready, it books them straight into the calendar while they are still on the line. No voicemail to call back. A booked appointment before they even hang up.
How it's built: the Call Answering and Routing layer runs on Vapi for the voice engine and conversational AI, ElevenLabs for natural voice synthesis, speech-to-text for transcription, and a phone line over a Twilio SIP trunk. Real-time reasoning by Claude decides intent on the call. The Appointment Booking module then uses a scheduling engine with Google Calendar sync, booking links, SMS confirmation, and slot negotiation, writing every booking into the shared calendar.
Module 2: Missed-Call Text-Back
What it does: most businesses add this last. Add it first. This is where the money leaks. Every missed call is a customer already dialing a competitor. So the second a call goes unanswered, it fires a text: "Sorry we missed you, want us to book you in?" A missed call stops being a lost customer and becomes a booked one.
How it's built: the Missed-Call Text-Back module runs automated text delivery through Twilio, workflow automation, a scenario builder, and a missed-call queue, with message copywriting generated against the brand voice and contact and deal sync into HubSpot. Nothing sits waiting for someone to notice it.
How seven agents work as one receptionist
Two things make this a real system instead of a chatbot. The first is one shared brain that every agent reads from and writes to: the knowledge base, the calendar, and every past call. The second is real access to the tools, which is the difference between a bot that talks and an agent that books.
For the dev team: the Shared Brain is the data layer. A human authors the /knowledge files once (business-info, services-pricing, hours-policies, call-scripts, faq-answers, cancellation-policy, brand-voice, escalation-triggers, and more), while Calendar and Call History are live records the system generates on its own. Supporting modules cover FAQ and Knowledge (answer generation, vector search, and a RAG pipeline over source documents), Lead Qualification (qualification logic, company enrichment, intake records, and a multi-step flow), and CRM Logging and Follow-Up (CRM sync, call transcription, task automation, and call summarization). Analytics and Call Reporting then serves dashboards, sentiment analysis, and executive reporting to the owner.
The Orchestrator: supervised, not autonomous
Sitting on top is one Orchestrator. It doesn't answer calls. It watches all seven agents, enforces the rules, and decides when the owner is needed. The pipeline is Collects, Scores, Routes, Drafts, Books, Approves. Routine calls resolve end to end. Only edge cases and exceptions reach the owner.
This is the point most people miss. The owner is still in it. They are just not the one picking up the phone. The tricky calls wait for a human. Everything else is handled. It is not autonomous. It is supervised. That boundary, routine work automated end to end and judgment calls escalated to a person, is exactly what makes a system like this safe to put in front of real customers.