Create a Voice Agent
Voice agents share the AI Agents section with chat agents and workflow bots. Each agent is independent — you can run multiple voice agents and pick a different default for inbound calls at any time.
Step 1 — Open the create dialog
From the sidebar, go to AI Agents → + Create agent.
In the dialog, pick Voice AI Agent as the channel. Voice agents have their own configuration shape (voice + languages + greeting) distinct from chat / workflow bots.
Step 2 — Pick a purpose template
Each template seeds the agent with a sensible default system prompt
- pre-selected tool set. You can override everything afterwards on the Persona / Tools tabs.
| Template | Use it for | Default tools |
|---|---|---|
| Sales | First-line lead qualification | capture_lead, assign_tag, schedule_callback, transfer_to_human |
| Support | Customer support intake | create_ticket, lookup_kb, transfer_to_human |
| Booking | Appointment booking (clinic, salon, etc.) | list_available_slots, create_booking, capture_lead |
| Custom | Build your own | None — toggle each tool individually |
Step 3 — Configure the persona
Six tabs run across the top of the agent edit page:
- Overview — name, status, default-agent toggle
- Persona — voice, languages, greeting, system prompt
- Knowledge Base — upload docs the agent answers from
- Tools — which actions the agent can take during a call
- Routing — fallback phone, working hours, follow-up template
- Activity — call history + tool fires per call
Voice
Pick a voice provider:
- Edge TTS (free, Microsoft) — broad language coverage, neutral accent. Default for new agents.
- Sarvam Bulbul v2 / v3 (paid) — Indian-language tuned, more natural for Hindi / regional languages. Used in Premium quality mode + as fallback.
Then pick a specific voice from the catalog (e.g. en-IN-NeerjaNeural,
hi-IN-MadhurNeural, meera, arvind). Click any voice’s preview
button to hear a 5-second sample before saving.
Languages
Tag every language the agent should accept. The first language is used for the greeting. Mid-call language switches happen automatically — when a Hindi-speaking caller switches to English, the agent follows.
Common picks:
- English + Hindi — broadest reach across India
- English + Hindi + Tamil/Telugu/Bengali/Marathi — region-specific
Greeting
The first line the caller hears. Keep it short (under 15 words):
“Hi, this is the AI assistant for Acme Dental. How can I help you today?”
Avoid:
- Greetings longer than 5 seconds — callers tend to start talking before you finish
- Multi-language greetings — pick one. Mid-call switching happens automatically.
System prompt
The voice agent’s “instructions” — what its job is, what tone it takes, what it should refuse. Example for a dental clinic:
You are the AI receptionist for Acme Dental in Bandra, Mumbai.
Your job is to:
1. Greet callers warmly + ask how you can help.
2. Answer questions about services + opening hours from the
knowledge base. Quote exact prices verbatim — never round.
3. Capture the caller's name + phone when they want a callback or
to book an appointment.
4. Use the capture_lead tool when the caller wants to be followed
up with. Use schedule_callback when they specify a time.
5. Transfer to a human when the caller asks for the doctor or
sounds upset.
Keep responses to 1-2 short sentences. This is a phone call, not
a chat. Match the caller's language naturally.A handful of platform-supplied “ground rules” (no inventing prices, say “let me check that and call back” when unsure, etc.) are prepended to your system prompt automatically.
Step 4 — Save + test
Click Save and then Test Call in the top-right. Talk to the agent in your browser to verify the voice + greeting + behaviour before pointing real callers at it. Test calls don’t charge AI Credits and have a 5-minute hard cap.