Build your voice agent

Configure the AI agent’s model, voice, prompt, and call behavior — then test it with a live web call.

The Voice section of the campaign form is where your AI caller takes shape. An estimated cost per minute badge updates as you pick options.

The essentials

• Model — the AI model powering the conversation.
• Voice — the voice your contacts hear.
• Language — searchable picker; defaults to English (US).
• System prompt (required) — tells the agent what to say and how to behave. Start from a persona template (Outbound Sales & Reactivation, Appointment Booking) or write your own. If your campaign collects data, an auto-managed “Information to collect” block is inserted into the prompt — hit Regenerate after changing the fields.
• Welcome message — the agent’s opening line when the call connects.
• Personalization — once an audience is selected, insert {{merge}} tokens (like the contact’s name) into the prompt; tokens that don’t exist in your contact data get flagged.

Advanced settings

Collapsible panels on the right fine-tune behavior:

Panel	What you control
Speech	Speaking speed, creativity, interruption sensitivity, response delay, ambient sound
Call	Max duration, end-after-silence, voicemail detection — hang up silently or leave a message you write
Retry	Max attempts, interval between attempts, and which results trigger a retry (no answer, busy, voicemail, call failed)
Post-call analysis	A prompt that runs after each call to classify what happened
Knowledge base	Facts, FAQs, and reference URLs the agent can draw on
Transcription	Accuracy mode, vocabulary specialization, background-noise handling, boosted keywords

Test it live

Click Test Campaign once you have a name and system prompt. A floating Test Audio panel opens (it doesn’t block the form):

1. Run Test starts a live voice call in your browser — talk to the agent through your microphone.
2. The transcript streams turn by turn as you speak.
3. End the Call stops it; run it again as many times as you like.

The test uses your current, even unsaved, configuration — so you can tweak the prompt and immediately hear the difference. No real contacts are called.