Voice Agents

What Is a Voice Agent?

A voice agent is an AI-powered virtual caller that can make and receive phone calls on behalf of your business. Under the hood, each agent is backed by a large language model (LLM) for conversation, a text-to-speech engine for speaking, and real-time speech-to-text for listening. You control the agent's behavior entirely through its configuration — no code required.

Agents are the core building block of CallTone. Every campaign, inbound call flow, and receptionist service is powered by an agent that you create and customize.

Creating an Agent

Navigate to Dashboard → Agents and click Create Agent. The creation form guides you through the following fields:

• Name — An internal label for your reference (e.g., "Q1 Sales Agent"). Not spoken to callers.
• System Prompt — Instructions that define the agent's personality, knowledge, and behavior. See best practices below.
• First Message — The opening line the agent speaks when a call connects.
• Voice — The text-to-speech voice the agent will use.
• Model — The LLM that powers the agent's responses.
• Capabilities — Optional toggles for appointment scheduling and receptionist mode.

After filling in the fields, click Save. Your agent is now ready to be assigned to a phone number and used in campaigns or for inbound calls.

System Prompt Best Practices

The system prompt is the single most important factor in determining how well your agent performs. Think of it as a detailed brief that you would give a new employee on their first day.

Be Specific

Vague instructions produce vague results. Instead of "Be helpful," write "You are a customer success representative for Acme Software. Your goal is to help the customer resolve their billing issue. If you cannot resolve it, offer to transfer them to a human agent."

Include Objection Handling

For sales agents, list the three to five most common objections and the ideal response for each. For example: "If the prospect says 'I'm not interested,' respond with: 'I completely understand. Many of our best customers felt the same way before seeing a quick demo. Would you be open to a five-minute walkthrough?'"

Set Clear Boundaries

Tell the agent what it must never do. Examples: "Never make up pricing or feature information. Never claim to be a human. Never provide legal or medical advice. If asked about something outside your knowledge, say you will have a team member follow up."

Structure the Conversation

Outline the flow of the conversation step by step. For a qualification call this might be: (1) Greet and introduce, (2) Ask about their current solution, (3) Identify pain points, (4) Present your value proposition, (5) Offer a next step (demo, meeting, resource). The agent will follow this order naturally.

Keep It Concise

While detail is important, avoid writing a novel. Aim for 200 to 500 words. The model performs best when instructions are clear and direct. Use bullet points rather than long paragraphs where possible.

Voice Configuration

CallTone supports multiple text-to-speech providers, giving you a wide selection of voices with different characteristics:

• Vapi — Built-in voices optimized for low latency. Best for use cases where fast response time is critical. Voices include alloy, echo, fable, onyx, nova, and shimmer.
• ElevenLabs — Premium voices with the most natural-sounding inflection and emotion. Ideal for customer-facing agents where voice quality is a top priority. Offers a large library of pre-made voices and the option to clone a custom voice.
• PlayHT — A wide variety of voices with good quality and competitive pricing. Strong multilingual support for agents that need to operate in multiple languages.

You can preview any voice before committing. Select a voice, click the play button, and listen to a sample. When choosing, consider your audience, brand personality, and the type of conversation the agent will have.

Model Selection

The language model determines how intelligently your agent responds. CallTone supports the following models:

• GPT-4o (Recommended) — The best overall model for most use cases. Handles complex conversations, nuanced objections, and multi-step reasoning with high accuracy. The preferred choice for sales, receptionist, and support agents.
• GPT-4o Mini — Faster response times at a lower cost. Ideal for simpler conversations such as appointment confirmations, FAQ handling, and basic lead qualification. Slightly less capable with complex reasoning.
• GPT-3.5 Turbo — The most economical option. Best for highly scripted conversations where the agent follows a rigid flow and does not need to improvise. Not recommended for complex sales or support interactions.

You can change the model at any time without losing any other configuration. Test different models to find the right balance of quality, speed, and cost for your specific use case.

Capabilities

Capabilities are optional features you can enable on a per-agent basis:

Appointment Scheduling

When enabled, the agent gains the ability to check your calendar for availability, propose time slots to the caller, and book confirmed appointments — all during a live call. Confirmation emails and reminders are sent automatically. You must configure your calendar settings under Settings → Appointments before enabling this capability.

Receptionist Mode

Turns the agent into a virtual front-desk receptionist. The agent can greet inbound callers, look up departments and staff members, transfer calls to the right person using Vapi's transferCall function, and take messages when staff are unavailable. Configure departments and staff under Dashboard → Receptionist.

Knowledge Base

You can upload documents that your agent can reference during calls. This is useful for providing the agent with detailed product information, pricing sheets, FAQs, or company policies that would be too long to include in the system prompt.

Supported file formats include PDF, DOCX, TXT, and CSV. Navigate to the agent's detail page and open the Knowledge Base tab to upload documents. The platform processes and indexes the content so the agent can retrieve relevant information on the fly when a caller asks a question.

For best results, keep documents focused on a single topic and use clear headings and formatting. The agent will reference these documents to provide accurate, grounded answers rather than generating information from its general training data.

Testing Your Agent

Before deploying an agent to a live campaign, always test it thoroughly. From the agent detail page, click Test Call and enter your own phone number. The platform will initiate a live call so you can interact with the agent exactly as a real prospect or customer would.

During testing, try the following:

• Let the agent deliver its first message and assess whether the opening is clear and engaging.
• Ask questions that are covered in your system prompt to verify the agent responds correctly.
• Throw common objections at the agent ("I'm not interested," "How much does it cost?") and see how it handles them.
• Ask something outside the agent's knowledge to verify it respects its boundaries and does not fabricate answers.
• If appointment scheduling is enabled, try to book an appointment and check that the confirmation email arrives.

After each test call, review the full transcript and AI-generated summary in the Calls section. Use what you learn to refine the system prompt, adjust the first message, or switch to a different voice or model.

Editing and Updating Agents

Agents can be edited at any time. Navigate to Dashboard → Agents, click on the agent you want to modify, and update any field. Changes take effect immediately for new calls — calls already in progress will continue using the previous configuration.

Common reasons to update an agent include:

• Refining the system prompt after reviewing call transcripts and identifying areas where the agent could perform better.
• Changing the voice or model based on caller feedback or cost considerations.
• Updating product information or pricing in the system prompt and knowledge base.
• Enabling or disabling capabilities like appointment scheduling as your business needs evolve.

We recommend reviewing agent performance weekly using the Analytics and Intelligence dashboards. Look for patterns in negative-sentiment calls, common objections the agent struggles with, and drop-off points in conversations. Continuous iteration is the key to maximizing agent effectiveness.

Next Steps

Now that you understand agents, explore these related topics:

• Getting Started — End-to-end setup walkthrough from account creation to your first call.
• Tutorials — Step-by-step guides for sales agents, campaigns, appointments, and receptionist setup.
• API Reference — Create and manage agents programmatically via the REST API.