AI Voice Agents

An AI voice agent holds the conversation itself: it speaks, listens, decides what to do, and can book an appointment or hand off to a person.

How it works

Caller audio
   │
   ▼
Speech to text  (Deepgram, or Groq Whisper)
   │
   ▼
Language model  (OpenAI, or Groq)  ──▶ tools: book, look up, transfer, callback
   │
   ▼
Text to speech  (ElevenLabs, Cartesia, or self-hosted Piper)
   │
   ▼
Caller hears the reply

The whole pipeline runs on the media server, beside Asterisk, so audio reaches it over the loopback interface instead of crossing a network. That is what makes the round trip fast enough to feel like a conversation.

Turn taking

The parts that make an AI agent tolerable to talk to rather than merely functional:

• Barge-in — the caller can interrupt, and the agent stops talking
• Silence detection — the agent notices when the caller has stopped and does

not wait awkwardly

• Filler injection — brief natural fillers cover thinking time so pauses do

not read as a dropped call

• Voice activity detection — Silero VAD on the media path distinguishes

speech from line noise

Choosing providers

Admin → Configuration → AI Settings. Language, speech and voice providers are selected per tenant, and a campaign can override the tenant default.

Piper matters if you would rather AI speech not involve an external vendor at all — it runs on our own infrastructure.

If your plan does not include AI features, none of these providers is contacted and the platform runs as a straight human-agent dialer.

Templates

Thirteen industry templates ship with the platform, each carrying a persona, a prompt, a call flow, available tools and a sample script. Start from the closest one and edit rather than writing a prompt from nothing.

Templates cover medical clinic, dental, legal, real estate, car dealership, car rental, financial services and insurance, home services, fitness, salon and spa, tourism, political, and general purpose.

What an agent can do mid-call

Agents have tools, not just words:

• Look up a customer by phone or email within the tenant, and use their

history in the conversation

• Check availability against a connected calendar
• Book an appointment, gated behind an explicit confirmation step so the

agent cannot book something the caller did not agree to

• Write a booking to a sheet where that is the workflow
• Transfer to a human agent
• Schedule a callback

Booking, honestly

Appointment booking depends on a working calendar integration for the tenant. If the integration is unavailable, the agent reports the failure rather than claiming success — an agent that cheerfully confirms a booking that never happened is worse than one that says it could not.

Bookings the AI could not complete land in Pending Bookings for a human to pick up, so nothing is silently lost.

Calendar booking from AI agents is an area under active work. Confirm the behaviour end to end in your own tenant before relying on it for production volume, and watch the Pending Bookings queue.

Testing an agent

Use the campaign Test action to run the agent against a number you control. Listen to the whole call, then read the transcript — the two together tell you whether the prompt is doing what you think.

What you get back

Every AI call produces a transcript, an outcome, a summary, and a cost breakdown across speech, language and voice usage. All of it lands on the call record alongside recordings from human calls.

Disclosure

Some jurisdictions require you to disclose that a caller is speaking to an artificial voice. That disclosure belongs in your prompt, and it is your responsibility — see TCPA and calling compliance.