Skip to main content

Building AI Phone Agents in 2026: Architecture, Costs and Platforms

How to build an AI phone agent in 2026: the reference architecture, per-minute costs from published rates, a worked receptionist, and which platform fits.

Founder of IImagined.ai

Published
Apr 5, 2026
Updated
Oct 11, 2026
Reading time
13 min read
Quick answer

An AI phone agent is five layers working in a loop: telephony, speech to text, a language model, text to speech and the tools the model can call, with an orchestration layer handling turn-taking and transfers. On published rates checked October 2026, a hosted agent runs about $0.08 to $0.15 a minute. Pick a hosted platform if nobody will own code, a framework such as LiveKit if a developer will, and design the handoff to a human before anything else.

An AI phone agent is five layers working in a loop: a carrier delivers the call, speech to text transcribes the caller, a language model decides what to say or which tool to call, text to speech speaks the reply, and an orchestration layer manages turn-taking, interruptions and transfers. You can rent the whole loop from a hosted platform for roughly $0.08 to $0.15 a minute on published rates, or assemble it yourself on an open-source framework for a lower platform fee and more engineering.

Prices and features checked October 2026 against the pricing pages for Vapi, Retell AI, Bland AI, LiveKit Cloud and Twilio Voice (US), plus the Vapi, Retell and LiveKit Agents docs. Every cost on this page is arithmetic from those published rates. We have not logged calls for it, so there are no latency or success figures here.

This is the hub for our voice AI guides. It lays out the reference architecture, prices a minute and a month from the vendors' own numbers, walks through a receptionist build, and points to the guide for each platform. If you already know you want a side-by-side of the platforms, go straight to the best voice AI platforms for developers.

How an AI phone agent works: the reference architecture

Every phone agent, whatever the vendor, runs the same loop once per turn of the conversation. Learn the loop and each platform's dashboard becomes a set of familiar choices with different labels.

One turn of a call
  1. 01
    Telephony

    A carrier or SIP trunk streams the caller's audio in

  2. 02
    Speech to text

    A transcript, plus a decision that the caller has finished

  3. 03
    Language model

    Reads the prompt and history, then replies or calls a tool

  4. 04
    Tools

    Calendar, CRM or your API returns a result as text

  5. 05
    Text to speech

    The reply becomes audio and plays back to the caller

  6. 06
    After the call

    Transcript, recording and outcome go to your systems

The orchestration layer sits around that loop and does the parts callers notice most: deciding when someone has stopped talking, stopping playback when they interrupt, keeping a tool call from leaving dead air, and transferring to a person. It is the main thing you are paying a hosted platform for.

LayerIts jobWhat you chooseHow it is billed
TelephonyGets call audio to and from the phone networkA platform number, or your own carrier over SIPPer minute, plus number rental
OrchestrationTurn-taking, interruptions, tool calls, transfersA hosted platform, a framework, or your own codePer minute, or your servers
Speech to textTurns the caller's audio into text as they speakProvider, model, languagePer minute of audio
Language modelDecides what to say and when to call a toolModel size, prompt, tool listPer token, quoted per minute by platforms
Text to speechSpeaks the replyProvider and voicePer character or per minute
Tools and dataCalendar, CRM, order lookup, your APIBuilt-in integrations or webhooks to your backendYour own hosting

There are two ways to build the middle of the loop. The table above describes the cascaded pipeline. The alternative is a single speech-to-speech model that takes audio in and returns audio out.

Two ways to build the middle of the loop
Cascaded: speech to text, model, text to speech
  • Three parts you can swap one at a time
  • Any text model, so tool calling behaves as it does in a chat app
  • A full text transcript at every step, which makes debugging easier
  • The cheaper route on published per-minute prices
Speech-to-speech: one realtime model
  • One model hears the audio and speaks the reply
  • Fewer parts to configure and fewer vendors to pay
  • Tone of voice reaches the model, not only the words
  • Costs more: $0.07 to $0.38 a minute on Retell's menu

Start cascaded unless you have a reason not to. It is the standard setup on the hosted platforms, the cheapest on their price lists, and the easiest to fix when a call goes wrong, because you can read exactly what the model was given.

What does an AI phone agent cost per minute?

Between about $0.08 and $0.15 a minute on a hosted platform, from the published rates. The shape of the bill differs more than the total: Vapi charges a hosting fee and passes models through at cost, Retell prices every model and voice per minute, Bland AI bundles the AI parts into one rate, and LiveKit Cloud charges a small session fee and leaves the models to you.

Illustrative inputs: 1,000 inbound minutes a month, about 330 three-minute calls, on a US local number from each platform, with no add-ons. LiveKit uses its Inference prices for Deepgram Nova-3 ($0.0048), GPT-5.4 mini ($0.0030) and Inworld Realtime TTS 2.0 Flash ($0.0090), before its $5 monthly inference credit. Bland lists telephony separately, so its row adds Twilio's US inbound rate. Engineering time is excluded everywhere, which flatters the options you have to build.

StackWhat the price list saysPer minute1,000 minutes a month
LiveKit Cloud, Ship plan$50 plan includes 5,000 agent minutes; LiveKit number at $0.01 a minute after 100; Inference models at $0.0168$0.076 at this volume$75.80
Vapi, usage only$0.05 hosting, plus transcriber, model and voice at provider cost; Vapi number free$0.082 to $0.129$82 to $129
Retell, GPT 5 mini$0.055 voice infrastructure, $0.015 voice, $0.008 model, $0.015 telephony$0.093$95 with the $2 number
Retell, GPT 5.4 miniAs above with a $0.024 model$0.109$111 with the $2 number
Bland AI, Start plan$0.14 covering model, transcription and voice; telephony assumed at Twilio's $0.0085$0.1485$148.50
Monthly cost at 1,000 inbound minutes (USD)
LiveKit Cloud, Ship plan
$75.80
Vapi, lower end
$82
Retell, GPT 5 mini
$95
Retell, GPT 5.4 mini
$111
Vapi, upper end
$129
Bland AI, Start plan
$148.50

Arithmetic from published rates, not measured calls. The LiveKit figure includes its $50 plan fee and assumes you write and run the agent code. Source: Vapi, Retell AI, Bland AI, LiveKit and Twilio pricing pages, checked October 2026.

Three things move these numbers more than the choice of platform.

  • The model. On Retell's list a small text model is under a cent a minute and a frontier one is $0.16 or more. Vapi's calculator shows the same spread, $0.0077 to $0.0452.
  • The voice. Retell prices its platform voices at $0.015 a minute and ElevenLabs voices from $0.04 to $0.10.
  • Volume against fixed fees. LiveKit's plan fee and Bland's $299 Build plan only pay off above a certain number of minutes. At 100 minutes a month, the plans with no monthly fee win everywhere.

If you assemble everything yourself, the carrier sets the floor. Twilio lists $0.0085 a minute to receive a call on a US local number and $0.0044 a minute for Media Streams, the feature that hands call audio to your own server, so about $0.013 a minute before any model runs. For outbound campaigns the sums change again, because unanswered dials cost money: our Bland AI vs Vapi vs Retell comparison prices a 1,000-dial campaign per connected call. For the model rates underneath all of this, see the LLM API pricing comparison.

AI voice agent for business: where it fits and where it does not

A voice agent earns its place on calls that are frequent, short and follow a pattern. It is a poor fit for calls where the caller needs judgement or sympathy more than speed.

  • Good fits: booking and rescheduling, opening hours and directions, order or appointment status, routing a call to the right person, after-hours message taking, and a first pass at lead qualification.
  • Poor fits: complaints, anything medical, legal or financial that amounts to advice, emergencies, and conversations where the answer depends on context the agent cannot look up.

The practical test is whether you could write the call down as a short script with a few branches and a clear point where a person takes over. If you can, an agent can run it. If the script keeps growing exceptions, keep a human on that line and give the agent the simple calls around it.

A worked example: the receptionist

Take a small clinic, a made-up one, that wants the phone answered when the front desk is busy. The agent has one job, booking, and three ways out: answer a simple question, take a message, or transfer.

ToolThe model calls it whenIt returns
check_availabilityThe caller asks for a timeFree slots near the requested time
book_appointmentThe caller has confirmed name, service and slotA confirmation the agent reads back
take_messageThe question is outside the agent's briefA reference, and a note sent to staff
transfer_callThe caller asks for a person, or sounds distressedA live handoff during opening hours
One booking call, turn by turn
  1. 1
    Greeting

    The agent names the business and says it is an AI assistant. One sentence, then it listens.

  2. 2
    Intent

    The caller asks for a check-up next Tuesday afternoon. The model picks out the service and the time window.

  3. 3
    Availability

    The model calls check_availability and offers two slots. It says a short holding phrase while the tool runs.

  4. 4
    Details

    It collects the name and a callback number, one question at a time, and reads the number back.

  5. 5
    Confirmation

    It repeats the service, day and time. Only after a yes does it call book_appointment.

  6. 6
    Close

    It gives the confirmation, asks if there is anything else, and ends the call. The transcript and booking go to staff.

Four decisions make this work, and none of them is the model. Keep replies short, because a paragraph that reads well sounds endless on a phone. Ask one question at a time. Read back anything that will be written down. And cap the call length, so a confused loop ends in a transfer and not a long bill. Tool calls follow the same pattern as function calling in a chat app, which our guide to Claude API function calling covers in detail.

At the illustrative 1,000 minutes a month from the table above, this receptionist costs somewhere between $76 and $149 a month to run on published rates, before anyone's time to build and look after it. To build this exact agent on a hosted platform, with Google Calendar as the booking tool, follow the Vapi tutorial.

How to build an AI calling agent: three routes

The routes differ in who owns the orchestration layer. Choose by who will maintain the thing in six months, not by which demo sounded best.

  1. A hosted platform. Vapi, Retell and Bland AI run the loop for you. You write a prompt, choose models and a voice, attach a number and connect tools. This is the right first build for almost everyone.
  2. A framework. LiveKit Agents is open source under the Apache 2.0 licence. Your Python or Node.js code joins a call as a participant and the framework handles turn detection and interruptions. You can run it on LiveKit Cloud or host it yourself; the LiveKit Agents guide covers both, with the volume at which it starts to pay.
  3. Your own stack. A carrier number, a media stream to your server, and model APIs you call directly. You pay no platform fee and own every problem, including the hard realtime ones. Take this route only when the phone agent is your product.
Which route fits
High, steady volume
A hosted platform on a paid plan. Price the monthly fee against the per-minute saving first.
A framework such as LiveKit, on its cloud or your own servers. The platform fee is where the saving is.
Low or uneven volume
A hosted platform, pay as you go. Nothing to maintain, live in days.
A hosted platform through its API. Move to a framework when the minutes justify the work.
No developer on the team
A developer owns it

The platforms, and the guide for each

Four names cover most builds in 2026. The prices are the platform's own fee, checked October 2026; models, voices and telephony are extra except where a row says otherwise.

PlatformWhat it isListed platform priceBest fit (our view)
VapiHosted orchestration; you choose each model$0.05 a minute hosting, models at costA developer who wants control without running infrastructure
Retell AIHosted platform with a priced menu of models and voices$0.055 a minute voice infrastructure, then per-minute model and voiceSupport and booking lines that need predictable costs
Bland AIHosted platform with one bundled AI rate$0.14 a minute on Start; $0.12 on Build at $299 a monthSales teams that want one rate and built-in batch calls
LiveKit AgentsOpen-source framework you run in code, on LiveKit Cloud or your own servers$0.01 per agent session minute on CloudEngineering teams with high, steady volume

One detail catches people on day one: Vapi's free phone numbers are US numbers and inbound only, per its docs. For outbound or international calls you import a number from a carrier.

The FCC's summary is on its ruling page. Whatever the law requires where you operate, say at the start of the call that the caller is speaking to an AI assistant and that the call may be recorded. Callers who find out halfway through tend to stop trusting the rest of the conversation.

Start here: a first agent in a week

If this is your first phone agent, do these in order. The reading path is one guide per step.

  1. Write the call down. One job, the questions it needs answered, the tools, and the point where a person takes over.
  2. Build it on a hosted platform with the Vapi tutorial, on free credits and a free inbound number.
  3. Price your real volume with the cost table above, then check a second platform in the platform comparison.
  4. Connect what happens after the call with the n8n webhook guide.
  5. Revisit the route at volume with the LiveKit Agents guide.
Before real callers reach it
  • The agent says it is an AI and, where required, that the call is recorded
  • A transfer to a person works, and there is a plan for when nobody picks up
  • Every tool has been tested with a taken slot, a wrong date and a tool error
  • A maximum call length is set, so a stuck call ends
  • The calendar and phone number are the real ones, in the right timezone
  • Someone reads the first 20 transcripts and fixes the prompt from what they find
  • You know the monthly cost at your expected minutes, and what doubles it

A working agent is a feature. Selling it, as a product or as a service to local businesses, needs the parts around it. Our AI SaaS Builder program has no voice lessons, but it teaches those parts: tool use with the Claude API, a Supabase database, deployment, pricing models and Stripe billing.

AI phone agents: FAQ

What is an AI phone agent?

An AI phone agent is software that answers or places phone calls and holds a spoken conversation. A carrier delivers the call audio, a speech-to-text model transcribes the caller, a language model decides what to say or which tool to call, and a text-to-speech model speaks the reply. An orchestration layer handles turn-taking, interruptions and transfers to a person.

How much does an AI phone agent cost per minute?

On published rates checked October 2026, a hosted agent costs roughly $0.08 to $0.15 a minute all in. Vapi's calculator shows $0.082 to $0.129 a minute, Retell lists an all-in range of $0.07 to $0.31 depending on the model and voice, and Bland AI charges $0.12 to $0.14 a minute plus telephony. LiveKit Cloud charges $0.01 per agent session minute, with models and telephony on top.

Can I build an AI calling agent without code?

Yes, for a first version. Vapi, Retell and Bland AI all have dashboards where you write a prompt, pick a voice, attach a phone number and test a call, and LiveKit has a browser-based Agent Builder for prototypes. You need code when the agent has to call your own backend, when you embed calls in an app, or when you want to own the infrastructure.

What is the difference between a cascaded pipeline and speech-to-speech?

A cascaded pipeline uses three models in a row: speech to text, a text language model, then text to speech. A speech-to-speech model takes audio in and returns audio out in one step. Cascaded is cheaper and lets you swap each part. On Retell's price list, text models start at $0.0016 a minute while its speech-to-speech options run from $0.07 to $0.38, checked October 2026.

Do I need consent to use an AI phone agent?

For outbound calls in the US, usually yes. The FCC confirmed in a 2024 declaratory ruling that AI-generated voices count as artificial voices under the Telephone Consumer Protection Act, so those calls need the called party's prior express consent. Inbound lines raise different questions, mainly call recording and telling callers they are talking to an AI. Rules vary by state and country, so take legal advice.

Can an AI phone agent transfer a call to a human?

Yes, and every production agent should. Retell documents warm transfer, LiveKit supports warm and cold transfers, and Bland AI bills transfer time at $0.04 to $0.05 a minute unless you bring your own Twilio account. Design the transfer first: when the agent hands off, what it tells the person picking up, and what happens when nobody answers.

Which AI voice agent platform is best for a small business?

Choose by who will build it. If nobody on the team writes code, a hosted platform with a dashboard, a phone number and a calendar tool gets a receptionist live fastest: Vapi, Retell or Bland AI. If a developer will own it and call volume is high and steady, LiveKit's open-source framework has the lowest platform fee. Test a real call on two of them before you commit.

Every guide in this hub

Free · no card needed

Start with the free Starter Kit.

A unit economics calculator from AI SaaS Builder, plus four more working files.

Starter Kit5 files · free
All Access · all four programs · $99/mo

The agent answers the phone. Now build the product around it.

All Access includes AI SaaS Builder. It does not teach voice platforms; it teaches what sits around one: tool use and structured output with the Claude API, Supabase, deployment on Vercel, pricing models and Stripe billing. The other three programs, live coaching and the private community are in the same subscription.

See All Access: $99/mo →30-day money-back guarantee
Free

Building a phone agent this month?

Join the free Discord to compare platform bills, prompts and transfer setups with other builders.