Building Voice Agents for Small Business

Why voice agents?

Most small businesses lose 20-40% of inbound calls. Staff are busy, it’s after hours, or the phone just rings. A virtual receptionist that answers every call, books appointments, and routes urgent issues pays for itself the first time it captures a lead that would have been lost.

This guide walks through building one with reaatech’s voice-agent platform.

Before you start

You’ll need:

• A Twilio account with a phone number
• An LLM API key (OpenAI, Anthropic, or any provider)
• Node.js 20+ and pnpm

No AI expertise required. If you can configure a webhook and edit a TypeScript config file, you can run this.

Step 1: Clone the repo

$ git clone https://github.com/reaatech/voice-agent
$ cd voice-agent
$ pnpm install
$ cp .env.example .env

Step 2: Configure your environment

Edit .env with your Twilio credentials and LLM provider:

TWILIO_ACCOUNT_SID=ACxxxx
TWILIO_AUTH_TOKEN=xxxx
TWILIO_PHONE_NUMBER=+15555550123
LLM_PROVIDER=openai
OPENAI_API_KEY=sk-xxxx

You can use any OpenAI-compatible endpoint. The platform supports Anthropic, Groq, and local models too.

Step 3: Define your agent’s behavior

The agent’s behavior is defined in a simple TypeScript config. Here’s the default receptionist:

export const receptionistConfig = {
name: "receptionist",
greeting: "Thanks for calling. How can I help you today?",
systemPrompt: `You are a professional receptionist for a small business.
  - Answer questions about hours, location, and services
  - Book appointments using the calendar tool
  - Route urgent calls to the right person
  - Always be courteous and concise`,
tools: ["calendar", "transfer", "faq"],
maxTurns: 12,
idleTimeoutMs: 30_000,
};

Step 4: Start the agent

$ pnpm dev
Voice agent running on http://localhost:3000
Twilio webhook configured at /api/incoming-call

Step 5: Point Twilio at your agent

In the Twilio console, set your phone number’s voice webhook to:

https://your-domain.com/api/incoming-call

Call your number. The agent picks up on the first ring.

What you’ll notice immediately

The agent handles the basics well: hours, location, services. It books appointments with surprisingly natural conversation flow. Callers who don’t realize they’re talking to AI say “thank you” and hang up satisfied.

What it won’t handle perfectly out of the box:

• Niche industry terminology (you’ll want to add to the FAQ tool)
• Multi-step edge cases (“I need to reschedule, then also ask about billing”)
• Callers who ramble for 90 seconds without stating their need

These improve as you add to the FAQ and tune the prompt.

Production considerations

Cost

At 500 calls/month with average 3-minute conversations, you’re looking at roughly $40-60/month in LLM costs plus Twilio’s per-minute rates. The DIY tier uses your own API keys — no platform tax.

Prompt injection

The platform includes prompt injection guards at the tool-use layer. If a caller says “ignore your previous instructions,” the guardrail intercepts it before it reaches the LLM. This is on by default.

Monitoring

Every call generates structured telemetry: duration, turns, tools used, confidence scores, cost breakdown. These flow to your observability stack via OTel. You’ll see if call quality degrades when you update the prompt or switch models.

Scaling

The DIY setup handles 5-10 concurrent calls on a $20/month VPS. Going beyond that is about horizontal scaling — the architecture is stateless at the application layer. State lives in the call context object, not in the process.

When to go from DIY to built-for-you

The DIY tier works well when:

• You have someone comfortable editing a TypeScript config file
• Your call volume is under 200/month
• Your use case fits the standard receptionist template

Consider having us build it when:

• You need custom tool integrations (your CRM, your calendar system)
• Your call volume is 500+/month and you need reliability guarantees
• You want someone else to own the initial setup and training
• You need HIPAA review or other compliance work

The architecture

Common failure modes and how we handle them

Hallucinated appointments. The calendar tool validates dates and times at the API level before confirming with the caller. If the LLM generates a date that doesn’t exist, the tool returns an error and the agent gracefully recovers.

Infinite loops. The maxTurns config (default 12) prevents conversations from running forever. After the limit, the agent politely wraps up and offers to transfer to a human.

Silence handling. After 30 seconds of silence (idleTimeoutMs), the agent checks in: “Are you still there?” After 60 seconds, it offers to take a message and hangs up.

Accent and dialect. The STT pipeline uses the default Deepgram model which handles most accents well. If you serve a specific dialect community, you can swap to a tuned model with one config change.

Ready to try it?

The repo is at reaatech/voice-agent. It’s MIT-licensed. Clone it, configure it, and take your first call this afternoon.

If you’d rather have us build and deploy it, book a conversation. First one’s free.