Before you hire someone to build an AI voice agent

The short answer

An AI voice agent is a chatbot on the phone. It answers calls, understands what the caller says, replies in a spoken voice, and hands the call to a person or takes an action such as booking a slot.

Before you hire someone to build one, decide three things: which calls it should handle (and which it must pass on), what happens when it gets stuck, and whose accounts every part runs on. Then test it by calling it, from real phones, the way your customers will. Voice is priced higher than chat for good reasons, and most of them show up only on a real call.

What a voice agent is made of

A chatbot is one system. A phone agent is usually five, joined together:

1. A phone number and line, from a telephony provider. Calls come in here.
2. Speech to text, which turns the caller’s words into text.
3. The AI model, which decides what to say, using your documents and rules, the same as a chatbot.
4. Text to speech, which turns the answer back into a voice.
5. Actions and handoff: transferring the call to a person, booking in your calendar, looking up an order.

Some platforms bundle these; some developers assemble them from separate services. Either way, ask which parts exist, which company provides each one, and which account each sits on. Each can have its own bill and its own limits, which is one reason voice costs more to run than chat.

Why voice is harder than chat

The AI agents and chatbots page notes that voice agents are priced higher because delay and interruptions are hard. In practice that means:

• Delay. Every reply passes through speech recognition, the model and a voice. A pause that is fine in a chat window feels like a dropped call on the phone.
• Interruptions. People talk over each other on the phone. The agent has to stop speaking when the caller starts, and not treat “mm-hm” as an answer.
• Noise and accents. A caller in a car or a shop is harder to understand than typed text. A misheard name or number becomes a wrong booking.
• No links, no scrolling. A chatbot can show a link or a list. A voice agent has to say it, and the caller has to remember it.
• Confidence. A wrong answer spoken in a calm voice sounds more certain than the same answer typed. The never-do list matters even more.

Decide these before you message anyone

Items 1 to 4 are the same as for a text chatbot, and what to give a chatbot developer before they start covers them in detail. If you already have a chatbot, most of this pack exists.

What moves the price

I have no reliable price list for voice builds, and the right figure depends on the parts above. What moves it:

Factor	Why it adds cost
Number of call types	Each needs its answers written and tested by voice
Actions in live systems	Each connection needs building and a test in the real system
Handoff rules	After-hours paths, callbacks and message-taking add branches
Languages and accents	Each language needs its own testing, and voices differ
Call volume	Usage is billed per minute or per call, across several providers
Compliance needs	Recording, consent and data storage rules add setup and review

Ask each developer to quote the build and to estimate the running cost per 100 calls separately. The second number is the one you will pay every month.

Running costs and limits

Every part of a voice agent is billed by use, and every paid account has limits. This is where unattended systems stop quietly.

One of my own overnight jobs hit a monthly spending limit on the AI account it ran on. It produced its usual output, then ended with an error. Another job would have shared that same account, so I kept that second job switched off rather than have two jobs draw on one limit. A phone line is less forgiving than an overnight job: if a limit is reached at 11 on a busy day, calls go unanswered.

So ask for three things: every account in your name, a spending limit you chose, and an alert to you well before any limit is reached, not after.

How to test it before you accept it

Call it. Not the developer’s demo recording, and not a test console: real calls from real phones.

• Ten of your common questions, asked in your customers’ words, including one with a name and one with a number to write down.
• Three questions it cannot answer. It should say so and offer a person or a callback.
• Three never-do questions. It should decline every time.
• One interruption: start talking while it is speaking.
• One call from a noisy place, and one with a long silence.
• One handoff during working hours and one after hours. Check that a person actually receives the call or the message.
• Then read the transcripts and the usage for those calls.

The scoring method in how to test an AI chatbot before you accept it works for voice as well: decide the pass mark before you call, and count wrong answers, not just good ones.

Where it goes wrong anyway

• Callers hang up at the first pause. Test the delay on a mobile network, not only on the developer’s connection.
• The handoff number changes. Staff leave, hours change. Someone has to update the agent; agree who, and at what rate, as in the maintenance guide.
• The usage bill surprises you. Watch the cost per call for the first month, with the limit and alert already set.

Wrap-up

A voice agent is a chatbot with a phone line, a voice and several more bills. Decide which calls it takes and which it passes on, write the never-do list, put every account in your name with a limit, and accept it only after you have called it yourself from a real phone. If your customers mostly write rather than call, start with a chatbot; everything you prepare for it carries over.