What Is an AI Voice Agent? How It Works, What It Costs and When to Give It a Face

Every voice agent runs the same three steps for each turn of the conversation, streaming as it goes. Many chain a separate model for each step, a speech-to-text, LLM and text-to-speech pipeline, as LiveKit's Agents documentation describes it. Speech-to-speech models, such as OpenAI's Realtime API, work directly with audio and do all three in one.

1. 01

   Listen

   Speech recognition turns the caller's voice into text as they speak, and turn detection decides when they've finished their thought.

2. 02

   Think

   A large language model reads the conversation and its instructions, decides what to say, and can call tools: look up an order, book a slot, update a CRM.

3. 03

   Speak

   Text-to-speech turns the reply into a natural voice and streams it back in pieces, so the caller hears the start while the rest is still being generated.

Around that loop sit the parts that decide how human it feels, which voice agent frameworks such as LiveKit Agents handle for you: turn detection, so it doesn't cut people off; interruption handling, so it stops when someone speaks; and instructions that tell it who it is, what to say and when to hand over to a person.

• Customer support: Answers product questions, troubleshoots and hands off to a person when needed.
• Sales and lead qualification: Greets visitors, asks what they need and books a demo when they're a fit.
• Reception and scheduling: Answers common questions, takes messages and collects appointment requests.
• Tutoring and language practice: Explains one idea at a time and lets learners practise speaking out loud.
• Screening interviews: Asks every candidate the same structured questions and records the answers.
• Sales training: Plays a buyer so reps can practise calls and objections out loud.

See each one with an example and its cost on our AI voice agent use cases page.

Latency

How quickly it answers. Long pauses make a conversation feel like a phone tree.

Turn-taking

Whether it waits for people to finish a thought, or jumps in at every pause.

Interruptions

Whether it stops talking the moment someone speaks, as a person would.

Tools

Whether it can act during the call (check, book, update) rather than only talk.

Transcripts

Whether every conversation is recorded as text you can review and send to your own systems.

Most platforms charge per minute of conversation. Self-serve rates run from about $0.06 a minute (Cartesia) to $0.14 (Bland AI), on prices checked on 24 October 2026, but the headline rate often leaves out the language model, the voice or the phone line, which are billed on top. Agents with a realistic avatar from most vendors cost more, from about $0.11 to $0.50 a minute, because the video is rendered in real time. Talking Avatar Realistic Avatars voice agents cost $0.06 a minute with the avatar, voice and AI included.

See what Retell, Vapi, Bland and ElevenLabs charge per minute →

On a phone line, voice is all there is. When the conversation happens on a screen, a realistic face that speaks the agent's words can keep people's attention and make the agent feel like someone, not something: useful for website support and sales, tutoring, screening interviews and onboarding.

With Talking Avatar Realistic Avatars you can do this two ways: build a complete voice agent with an avatar in the dashboard, without code, or give a voice agent you already run a face through the API for $0.02 a minute.