How Does an AI Receptionist Work?

The six-step technical flow.

1. 1. Inbound call hits a forwarded number
   You forward your existing business line (or a dedicated Twilio number) to the AI receptionist platform. The customer dials your normal number; the call routes to the agent within one to two rings.
2. 2. Speech-to-text (Deepgram, Whisper, etc.) transcribes
   The caller's audio is streamed to a speech-to-text service like Deepgram Nova-3. Crisp uses Deepgram for sub-300ms transcription latency, which is critical to making the conversation feel real-time.
3. 3. LLM generates the response (Claude Sonnet 4.6 on Crisp)
   The transcribed text plus a configured system prompt (your hours, services, pricing tiers, FAQs, escalation rules) is sent to a large language model. Crisp uses Claude Sonnet 4.6 via OpenRouter. The model returns the receptionist's next line.
4. 4. Text-to-speech (Cartesia, ElevenLabs) speaks back
   The LLM response is sent to a TTS engine that renders it as natural-sounding audio. Crisp uses Cartesia Sonic-3, one of the cheapest premium voices that consistently wins blind tests against pricier alternatives.
5. 5. Orchestration (Vapi) ties it all together
   Vapi orchestrates the round-trip, manages call state, handles interruptions and barge-in, and exposes webhooks for booking, escalation, and CRM hand-off. Without orchestration, the call sounds choppy and the agent can't book appointments.
6. 6. Webhook fires to your CRM or calendar
   When the agent books an appointment, a webhook fires to Jobber, Buildertrend, Housecall Pro, Google Calendar, or whatever CRM you use. The lead arrives qualified, scoped, and on your calendar without a human touching it.

Premium stack vs budget stack: where the gap shows up.

All AI receptionists use roughly the same architecture (the four components above). The gap between a $49/mo tool and a $425/mo agent is in the components chosen. A budget tool runs an older LLM with a synthetic-sounding TTS voice. A premium agent runs Claude Sonnet 4.6 with Cartesia Sonic-3. The first sounds robotic and gives wooden answers. The second handles unexpected questions, captures nuance, and books bookings the first tool would drop.

Where it shows up most:

• Emotional callers. A burst pipe at 2am, an anxious dental patient, a homeowner with a flood. Budget LLMs respond robotically; premium LLMs read the emotion and respond calmly.
• Multi-step bookings. "I need a furnace inspection but also want a quote on a new water heater while you're here." Budget LLMs handle one or the other; premium LLMs handle both.
• Edge questions. "Do you do permit work?" "What's your hold-harmless policy?" "Are you HCRA-registered?" Budget LLMs guess; premium LLMs check the configured knowledge base and answer correctly or escalate.

What the conversation actually sounds like.

Listen to Demo 2 on the Crisp homepage. The agent picks up, runs the qualifying conversation (scope, budget tier, timeline, address), books a site visit on the configured calendar, and texts you a summary the moment the call ends. The whole call takes 90 seconds to three minutes. The caller does not have to repeat themselves to a human later; the agent captured everything.