This article is part of our AI Front Desk hub. Want help applying it? Talk to eSEOspace.
An AI phone agent turns the caller's speech into text, sends that text to a language model and reads the model's reply aloud. That loop repeats for every turn until the call ends or moves to a person.
This article is part of our AI Front Desk guide. Platform details and prices were checked against each vendor's documentation on 9 November 2026, and they can change.
Key Takeaways
- A voice agent has three core parts: a transcriber, a language model and a voice.
- A phone connection and your business data sit around those three parts.
- Every step adds delay, so response time depends on the whole chain.
- Speech recognition makes mistakes, so the agent needs instructions for handling them.
- Platforms can bill each part separately, so a headline per-minute price may not be the full bill.
The three core parts
Vapi's introduction splits a voice agent into three parts. It lists speech-to-text, a large language model and text-to-speech. Other platforms use different labels for the same jobs.
| Part | Job on the call | What can go wrong |
|---|---|---|
| Speech-to-text | Writes down what the caller said | Mishears names, numbers and addresses |
| Language model | Decides what to say or do next | Gives a wrong or off-script answer |
| Text-to-speech | Reads the reply aloud | Mispronounces words or sounds flat |
We explain the broader idea in how AI agents work. A phone agent is one version of that pattern with a live conversation on top.
What happens during one turn
Here is one turn of a call, in order.
- The call arrives on a phone number connected to the platform.
- The transcriber turns the caller's words into text as they speak.
- The system decides the caller has finished their turn.
- The model reads its instructions, the conversation so far and any retrieved business facts.
- The model writes a reply or asks to use a tool, such as a calendar.
- The voice engine reads the reply aloud over the phone line.
Then the cycle starts again. A three-minute call can include many of these turns.
Where business facts come in
Retell's knowledge base documentation describes one approach. Before the agent responds, Retell uses the transcript so far to find relevant passages, then adds them to the prompt. We cover what to load in what to put in your AI receptionist's knowledge base.
Where tools come in
A tool lets the agent do something, such as check a calendar or transfer the call. Retell's best-practice notes warn about the gap while a tool runs. They say: "Without a parameter carrying something to say, the agent goes silent exactly when the caller is waiting."
Why latency matters
Latency is the pause between the caller finishing and the agent starting to reply. Each step in the chain adds to it, including transcription, the model, any lookup and the voice.
Vendors publish their own figures. Vapi's introduction lists "Sub-600ms response times" as a capability, and Retell's docs say knowledge base retrieval should generally add under 100 milliseconds. These are vendor statements, and we did not measure them.
Real calls can behave differently from a demo. Before you commit, phone a trial agent from an ordinary mobile and ask it something that needs a lookup.
Where phone agents struggle
Transcription errors. Retell's notes say: "The transcript comes from speech recognition and will contain mistakes." Have the agent read back phone numbers, addresses and spellings before it saves them.
Long answers. Retell's notes also say: "Long turns feel like being lectured and are painful to interrupt." Keep each reply to one idea.
Screen habits. Retell warns: "Everything the model writes gets read aloud, which changes what a good prompt looks like." Its notes say to ban bullet lists, headers and emoji in the prompt.
Handoffs. A transfer can fail if nobody picks up. Retell's call transfer node includes a separate path for that outcome, and your plan needs one too. See when an AI should hand off to a person.
What each part costs
Build platforms can show the parts as separate lines. These are published prices on 9 November 2026.
| Item | Published price | Source |
|---|---|---|
| Retell voice agents | $0.07 to $0.31 per minute | Retell pricing |
| Retell voice infrastructure | $0.055 per minute | Retell pricing |
| Retell platform voices | $0.015 per minute | Retell pricing |
| Retell telephony | $0.015 per minute; no charge for SIP trunking or custom telephony | Retell pricing |
| Retell knowledge base | +$0.005 per minute; first 10 knowledge bases free | Retell pricing |
| Vapi hosting | $0.05 per minute on the no-package tier, models and transport extra; other packages are priced differently | Vapi pricing |
Retell's page doesn't say plainly whether its $0.07 to $0.31 range includes telephony. Model prices also vary by the model you pick.
Hypothetical arithmetic, not a vendor figure. Add Retell's infrastructure, platform voice and telephony lines, plus its listed GPT 4.1 nano rate of $0.0032 per minute. That comes to $0.0882 per minute, or about $0.26 for a three-minute call, before add-ons.
Retell also bills the whole call. Its pricing FAQ says you are charged during silence and hold time. It adds: "Each call is tracked to the nearest second."
Our AI receptionist pricing guide compares packaged products as well.
Recording, transcripts and disclosure
An AI phone agent creates text from each call, and some products also record audio. Rules on consent, recording and telling callers about AI vary by place. This is general information, not legal advice, so start with our AI Front Desk guide and ask a qualified attorney about your setup.
Next step
Map one call type from start to finish before you choose a platform. Our Retell, Vapi and Bland comparison covers build options. If you need an agent connected to your own systems, our AI agent development team can scope it with you.
Frequently asked questions
Is an AI phone agent just a chatbot with a voice?
The model works in a similar way, but a phone call adds transcription, a generated voice and live timing. A chatbot reads typed words. A phone agent has to work from an imperfect transcript while the caller waits.
Can the agent do things, or only talk?
It can act when it has tools connected, such as a calendar or a transfer function. Without those connections, it can only talk and take details. Check which tools a product supports on your plan.
Why does the agent sometimes get a name or number wrong?
The transcriber can mishear a word, and the model then works from that error. Retell's documentation says transcripts will contain mistakes. Reading details back to the caller reduces the risk of saving the wrong one.
What happens to the cost when the call is transferred?
It depends on the platform. Retell's pricing page says: "Once the call is transferred, the AI voice agent fee stops." It also says: "Only the telephony fee continues for the remainder of the transferred call."
Bland's pricing page lists separate transfer-minute rates, and says customers who bring their own telephony do not pay them.
How fast should an AI phone agent respond?
We don't have an independent benchmark to give you. Vendors publish their own numbers, and those may not match your calls. Judge it by test calls on real phones, with the lookups your calls will need.
