Autonomous Voice AI Agents
Firetell Voice AI Agents run on native realtime multimodal voice models—streaming audio directly between callers and AI without brittle multi-hop delays—so your workspace can answer inbound hotlines, qualify leads, book appointments, send SMS confirmations, and warm-transfer callers to human teams 24/7.
1. Creating a Voice Agent with AI Prompt Generation
- Navigate to Console → Voice Agents and click New Voice Agent.
- In the prompt composer, describe what your assistant should do in plain language:
"Create a bilingual English/Vietnamese receptionist for Lotus Dental Clinic. Greet callers warmly, answer pricing and insurance questions from our Knowledge Base, collect their preferred date and phone confirmation for bookings, and transfer dental emergencies immediately to extension 101."
- Firetell streams a complete, structured System Instructions draft in real time.
- Want adjustments before creating? Type follow-up instructions (e.g., "Ask for the patient's full name one question at a time and keep responses under 2 sentences") to refine the generated prompt interactively.
- Enter an Agent Name (3+ characters, e.g.,
Lotus Clinic Receptionist) and click Create Voice Agent.
2. Prompt Engineering & Dynamic Runtime Variables
Inside the Voice Agent editor, the left pane controls your assistant's Greeting Message and System Instructions (Prompt).
Dynamic Variables
Click any variable pill above the editor to copy and inject live call metadata into either your Greeting Message or System Instructions:
| Variable | Resolved Value at Call Time | Example Output |
|---|---|---|
{{caller_name}} | Matched customer name from Workspace Contacts (falls back gracefully if unknown) | Minh Nguyen |
{{caller_number}} | Caller's E.164 phone number or internal extension | +14155552671 |
{{called_number}} | Inbound Firetell DID / hotline number dialed by the caller | +18005550199 |
{{current_date}} | Current date in workspace timezone (YYYY-MM-DD) | 2026-11-12 |
{{current_time}} | Current local time in workspace timezone (HH:mm) | 14:30 |
{{current_day}} | Current day of the week | Thursday |
Greeting Message (greeting_message)
- Fixed / Templated Greeting (Up to 250 characters): Spoken immediately when the call connects—for example:
Hello {{caller_name}}! Thank you for calling Acme Support. How can I help you today?
- Dynamic AI Greeting (Leave Blank): If left empty, the AI generates a contextual opening dynamically based on your System Instructions and Conversation Memory (ideal when greeting returning callers differently from first-time callers).
System Instructions (10–10,000 characters)
Structure voice prompts for natural spoken dialogue rather than written essays:
# Role & Persona
You are Clara, the front-desk voice concierge for Grand Harbor Hotel. Speak warmly, clearly, and concisely—keep each turn to 1–2 short sentences so the caller can respond naturally.
# Context
- Today is {{current_day}}, {{current_date}}, and the local time is {{current_time}}.
- Caller phone number: {{caller_number}} (Name on file: {{caller_name}}).
# Core Responsibilities
1. **Room & Policy Inquiries**: Use your Knowledge Base to answer questions about check-in/out times, parking, breakfast, and room rates accurately.
2. **Reservation Assistance**: Ask one question at a time to collect check-in date, number of nights, and guest count. Offer to send a booking link via `send_sms`.
3. **Human Escalation**: If the caller asks for a manager, group event pricing (5+ rooms), or billing disputes, politely tell them you are connecting them now and invoke `transfer_call` to `101`.
# Guardrails
- Confirm phone numbers, dates, and spellings back to the caller before calling external tools.
- Never fabricate policies or discounts. When the caller says goodbye and needs nothing else, say a brief farewell and invoke `end_call`.
3. Model, Voice & Audio Call Controls
Configure realtime speech and call safety parameters in the right-hand sidebar:
| Setting | Range / Options | Guidance |
|---|---|---|
| Provider & Model | Realtime multimodal models (displays $rate/min badge) | Shows exact per-minute credit rate directly beside each model option. |
| Voice & Primary Language | Neural voices with gender/style tags across 30+ languages | Match Primary Language to your main caller audience for optimal STT accuracy and natural prosody. |
| Temperature | 0.1 – 2.0 (Default 0.8) | Use 0.4–0.7 for strict policy/compliance workflows; 0.8–1.0 for expressive, conversational concierge flows. |
| Speaking Speed | 0.5x (Slow) – 2.0x (Fast) (Default 1.0x) | 0.95x–1.05x works best for general support; 1.1x feels brisk for high-volume dispatch. |
| Noise Reduction | On / Off (Default On) | Filters street, vehicle, and cafe background noise on the caller's audio leg for cleaner speech recognition. |
| Conversation Memory | On / Off (Default Off) | When enabled, injects recent call summaries/transcripts for returning {{caller_number}} callers so the agent remembers prior issues (requires active Call Transcriptions). |
| Call Recording | On / Off (Default Off) | Records dual-channel stereo audio (Caller + Voice Agent) on both live DID calls and browser Test Calls. |
| Max Call Duration | 10 – 3,600 sec (Default 180 sec) | Hard safety cap that automatically terminates runaway or unattended calls to protect your credit balance. |
4. Tools & Function Calling
Under Tools / Actions, grant your Voice Agent autonomous telephony and backend capabilities during live conversations.
Built-in System Tools
Click Add Tool (or Choose Built-in Tools) to enable native telecom controls:
| Built-in Tool | Default | Parameters | When the Agent Uses It |
|---|---|---|---|
transfer_call | Enabled | to (E.164 number +1415..., extension 101, agent ID, or team ID), reason | Warm/blind handoff to a human specialist, department queue, or external phone line. Instruct the agent to speak a brief hold phrase in the same turn. |
end_call | Enabled | reason | Cleanly hangs up the SIP/WebRTC session as soon as the conversation concludes or the caller says goodbye. |
hold_call | Enabled | action (hold | resume) | Places the caller on hold or resumes audio while consulting external systems. |
send_sms | Optional | message (required), to (defaults to caller's number if omitted) | Texts booking links, addresses, order confirmations, or reference numbers to the caller's mobile phone mid-call. |
collect_digits | Optional | prompt, max_digits (1–20), timeout_seconds | Captures reliable keypad DTMF input for PINs, account numbers, or credit/member IDs in noisy environments. |
send_dtmf | Optional | digits (e.g., 1234# or 0) | Emits DTMF tones on outbound calls to navigate remote IVR trees or extensions. |
Custom Webhook Tools
Click New Tool under Custom Webhook Tools to connect your CRM, ERP, booking engine, or internal APIs:
- Tool Name & Description: Use
snake_case(e.g.,check_order_status) and write a precise description telling the LLM when to call the tool and what it returns. - HTTP Endpoint & Method: Configure
GET,POST,PUT,PATCH, orDELETEto your HTTPS URL (e.g.,https://api.yourcompany.com/voice-tools/orders). - Custom Headers: Attach authentication headers such as
Authorization: Bearer <token>orX-API-Keyso your backend verifies requests. - Parameters & Required Fields: Define typed arguments (
string,number,boolean, optionalenumvalues) that the Voice Agent extracts from the caller before firing the request.
5. Versioning, Browser Test Calls & Publishing
Firetell isolates draft experimentation from live production calls using immutable version snapshots (v1, v2, v3, …):
- Create a Draft Version: Open the version dropdown in the top bar and click New Version. Your live
publishedversion continues serving customer calls uninterrupted while you edit prompts, models, Knowledge Bases, or tools on the draft. - Run a Browser Test Call: Click Test Call in the top-right toolbar to speak directly with the currently selected version through your computer's microphone and speakers—ideal for validating latency, pronunciation, RAG answers, and webhook tool execution without dialing a PSTN number.
- Promote to Production: Click Save, then click Publish to switch live inbound traffic (from bound Phone Numbers and Call Flows) to the new version immediately.
For programmatic management, see the Voice Agents REST API.