AI voice agent: how they work and who should build one

Direct answer

An AI voice agent is software that holds a spoken conversation on the phone: it listens, turns speech into text, decides what to say or do, and answers in a generated voice. You can build one on a developer platform such as Vapi, Retell, or Bland and pay per minute, or buy a finished product that someone else builds and runs. Build if you have developers and an unusual call flow. Buy if you run a business and just need the phone answered well.

What an AI voice agent is

"Voice agent" is the category term. An AI receptionist is one kind of voice agent, aimed at answering a business line. Others handle appointment reminders, order status lines, after-hours support, surveys, or sales calls. The underlying technology is the same.

The "agent" part means it can do more than read a script. It can look things up (a customer's open job, store hours, an order status) and take actions through connected tools, within whatever limits the builder sets. Those limits are the hard part, and most of this page is about them.

How one works

Every voice agent stitches together the same pieces:

Telephony. A phone number and the connection that carries the call audio. Usually a carrier service or a SIP trunk.
Speech-to-text. Turns the caller's words into text as they talk.
The reasoning model. Reads the conversation, your instructions, and any looked-up data, and decides the next reply or action.
Text-to-speech. Speaks the reply in a generated voice.
Turn-taking. Detects when the caller has finished, handles interruptions, and keeps silences short. Bad turn-taking is why some agents feel robotic.
Tools and data. Calendar, CRM, order system, knowledge base. Each one the agent can reach is also something it can get wrong.
Records. Transcripts, recordings where permitted, and call outcomes.

Developer platforms package parts 1, 2, 4, and 5 and let you plug in the model and tools. Finished products package everything and hide the parts.

Developer platforms: Vapi, Retell, and Bland

These are the best-known build-it-yourself platforms. All three are fair choices for a technical team. Pricing below is from each vendor's page as of October 2026.

Platform	Advertised pricing	What that covers
Vapi	$0.05/min hosting, prepaid	Vapi's platform. Transcription, model, voice, and telephony are billed at provider rates on top
Retell AI	$0.07 to $0.31/min, pay as you go, $10 free credits	Voice infrastructure at $0.055/min plus voice, model, and telephony, which vary by choice
Bland	$0.14/min on the free-platform Start plan, $0.12/min on Build ($299/mo platform fee)	Bundled per-minute talk time, with transfer time billed separately

A few things to know before comparing those numbers:

The headline rate is rarely the full rate. Vapi passes model, voice, and phone costs through separately. Retell's range depends on which voice and model you pick. Bland bundles more into one number. Price out a full call for your setup.
Concurrency matters. Retell includes 20 concurrent calls and charges for more. Bland's plans set concurrency and daily caps. A busy Monday morning can hit a limit you never noticed.
You own the behavior. These platforms give you the controls. Whether the agent discloses it's an AI, refuses to quote prices, or books appointments without asking is up to you to build and test.

Who they're right for: software companies adding voice to their product, agencies building agents for several clients, and larger businesses with developers and an unusual call flow. If you want full control and have people to maintain it, these are strong tools.

Done-for-you voice agents

The other route is a finished product. Self-serve AI receptionists like Dialzara (from $29 a month for up to 60 minutes) and Rosie (from $49 a month for 250 minutes) let you fill in your business details and forward your number. Others, including TeamShift, set it up for you. See our AI receptionist guide for a wider list.

You give up some control. You get something that works this week, someone else handling updates, and pricing you can understand without a spreadsheet.

Build vs buy: how to decide

Build on a platform if:

You have a developer who can own it for the long haul, not just the launch.
Your call flow is unusual, like a multi-step intake that queries your own database.
You're putting voice inside a product you sell.
Call volume is high enough that per-minute savings outweigh engineering time.

Buy a finished product if:

You run a service business and want calls answered, not a software project.
Nobody on your team wants to debug speech timing or prompt behavior at 9 p.m.
Your needs are standard: answer questions, take messages, collect job details, get requests to you.

A rough check: if you can't name the person who will read transcripts and fix the agent every week, buy.

What any voice agent must get right

Whichever way you go, hold it to the same bar:

Disclose that it's an AI at the start of the call.
Say "I don't know" when the answer isn't in its information, and offer a callback.
Keep consequential actions behind a person. Quotes, refunds, payments, bookings, and order changes should become requests someone approves. Our guide to human review on AI calls covers why.
Offer a way out to callers who won't talk to an AI.
Keep a record of every call.
Mind the rules for outbound calls. The FCC ruled in February 2024 that AI-generated voices count as "artificial" voices under the TCPA, so robocalls using them need prior express consent unless an exemption applies. That matters most for agents that make calls rather than answer them.

Where TeamShift fits

TeamShift is on the buy side. Its team sets up an AI voice agent on your business line (your forwarded number or one TeamShift sets up) that answers day, night, and weekends. It says it's an AI assistant at the start of every call, answers callers from a business brief you write, and offers a callback when something isn't covered.

It takes messages and job details. Bookings, quotes, refunds, payments, and order changes are handed to you as requests to approve, and the caller is told someone will confirm. Missed and after-hours calls get voicemail and a text alert to you. Each call has a record (outcome and notes) in TeamShift. Callers don't get lookups in your tools. Only you, after verifying with a code texted to your phone, can have it check connected tools. Billing is usage-based with a free workspace and a maximum quoted up front, with no per-minute plan. See pricing and trust.

The difference from a standalone voice agent is what happens after the call. The same team that answers the phone also drafts the follow-up email for your OK, prepares estimate follow-ups through Sales for you to approve, and tracks the job. If you need outbound calling, live call transfer, or multiple languages, a developer platform or another vendor is the better fit, because TeamShift doesn't do those.

FAQ

What is an AI voice agent?

It's software that talks with people on the phone using speech recognition, a reasoning model, and a generated voice. It can answer questions, collect information, and take actions through connected tools within limits its builder sets.

How much does an AI voice agent cost?

As of October 2026, developer platforms advertise from about $0.05 to $0.31 per minute depending on platform and components, before your own build time. Finished small-business products start around $29 to $99 a month on vendor pricing pages.

What's the difference between Vapi, Retell, and Bland?

All three let developers build voice agents and bill by the minute. Vapi charges a hosting fee and passes model, voice, and phone costs through. Retell prices by the components you choose. Bland bundles more into a single per-minute rate with optional platform fees.

Should a small business build its own voice agent?

Usually not. Unless you have a developer who will maintain it, a finished product gets you a working phone line faster and with less risk.

Is an AI voice agent the same as an AI receptionist?

An AI receptionist is one kind of voice agent, focused on answering a business's inbound calls. Voice agents also cover reminders, support lines, surveys, and outbound calls.