# How to choose a voice agent platform

> Choose a voice agent platform on evidence: call types, phone system, data rules, latency and cost per minute, with platforms compared as of October 2026.

- Author: Pranav Mahaveer, Founder, SBD
- Updated: 2026-10-30
- Level: Intermediate. Time to do: One to two weeks
- For: Owners, operations and technical leads choosing how to put an AI agent on a phone line
- You will end with: A shortlist of two platforms tested on your own calls, with measured latency, an all-in cost per minute and the compliance design written down
- Canonical: https://smallbaydigital.com/guides/how-to-choose-a-voice-agent-platform/

## Steps

1. **List the calls you want to automate.** Group a month of calls by reason, count each group and note which ones need a system lookup or a booking.
2. **Work out how calls will reach the agent.** Decide between a new number, forwarding, a SIP connection to your phone system, WhatsApp or a website button.
3. **Write down the constraints.** Languages, where data must be stored, healthcare or payment rules, and whether any calls are outbound.
4. **Shortlist two platforms.** Use the constraints to pick two candidates from the table below, plus an open-source option if the phone system or data rules require it.
5. **Test both on your own calls.** Build the same small agent on both, run scripted test calls over a real phone line and measure latency from the recordings.
6. **Price it all-in and plan the exit.** Add platform, transcription, model, voice and phone costs per minute, and check you can move prompts, tests and integrations elsewhere.

> **The short answer.** Pick the platform last. First decide which calls the agent will take, how those calls reach it, and which rules apply. Those three answers narrow the field to two platforms. Build the same small agent on both, test it over a real phone line, measure the delay from the recordings and add up every cost per minute. As of October 2026, platform fees are 3 to 8 cents a minute and all-in costs usually land between 6 and 20 cents.

This guide is for choosing the technology. For what voice agents can do in a business, and how we build them, see our [voice agents page](/solutions/voice-agents/), and to hear one working, try the [live voice agent demos](/voice-agents/). Prices, versions and rules below were checked on 30 October 2026, and they change often.

## What a voice agent is made of

A phone call reaches the agent through a phone provider or your own phone system. Speech to text turns the caller’s words into text as they speak, and a turn detector decides when they have finished. A language model reads that text with your rules and knowledge and decides what to say or which tool to call. A voice model speaks the reply. Each of these layers can come from a different company, which is why the platform choice matters less than it looks: the platform mostly decides how easily you can swap the parts.

There are two designs. A cascaded pipeline runs speech to text, a model and a voice in sequence. It lets you use any model, gives you a transcript of every turn and lets you change one part without touching the others. A speech-to-speech model hears and speaks directly. It sounds more natural and handles interruptions well, but you have fewer choices of model and voice, and OpenAI’s flagship realtime model costs several times more per minute. Most business calls suit the cascaded design.

## The platforms, as of October 2026

| Platform                | Published price                                                           | Where it fits                                                                   |
| ----------------------- | ------------------------------------------------------------------------- | ------------------------------------------------------------------------------- |
| Retell AI               | $0.055 a minute plus voice and model; its own example is about $0.11      | Most small-business phone lines. QA built in; HIPAA agreement on Enterprise     |
| Vapi                    | $0.05 a minute plus each provider at cost                                 | Full control of each component. HIPAA is a $2,000 a month add-on                |
| ElevenLabs Agents       | $0.08 a minute plus model and phone                                       | Voice quality, many languages, WhatsApp. States US, EU and India data residency |
| Telnyx Voice AI         | $0.05 a minute with transcription and voice; about $0.06 all-in in the US | Lowest cost US lines. Billed in 60-second steps                                 |
| Deepgram Voice Agent    | $0.075 a minute bundled                                                   | EU endpoint and self-hosted containers                                          |
| Cartesia                | $0.03 a minute plus the model                                             | Low platform fee; on-premises through sales                                     |
| Bland AI                | $0.12 to $0.14 a minute bundled, phone extra                              | Simple bundled pricing for outbound. Daily call caps on lower plans             |
| LiveKit Agents, Pipecat | Open source; pay for hosting and parts                                    | Existing phone systems, data residency, high volume                             |

Synthflow now sells annual enterprise contracts only, and Hume is closing its voice API on 13 November 2026. Enterprise contact-centre vendors such as Sierra, Decagon, Parloa and PolyAI sell to much larger companies.

## 1. List the calls you want to automate

Take a month of calls from your phone system or call log and group them by reason: booking, order status, opening hours, new enquiry, complaint. Count each group and note how long the calls last. Mark which ones need a lookup or an action in another system, and which ones need a person no matter what.

Start with one group that is frequent, repetitive and low risk. After-hours calls and overflow are good first candidates, because anything the agent handles there is a call that would otherwise have been missed.

## 2. Work out how calls will reach the agent

There are four routes. A new number from Twilio or Telnyx is the simplest. Forwarding from your existing number works with almost any provider. A SIP connection to your phone system keeps transfers to your staff inside the same system, but it is where most of the integration effort goes: Microsoft Teams needs Direct Routing, and 3CX cannot connect straight to most hosted platforms, so it needs a bridge such as LiveKit’s SIP gateway. WhatsApp Business calling lets customers call your WhatsApp number; businesses cannot start calls from US or Canadian numbers, and the account needs a messaging limit of at least 2,000 recipients a day.

Talk to whoever runs your phone system before you shortlist anything. The route can rule platforms in or out on its own.

## 3. Write down the constraints

List the languages and accents your callers use, and collect a few recordings. Phone audio is compressed, and one 2026 study found it roughly doubled transcription errors and hit long numbers hardest, so test on phone audio, not on studio samples.

Write down where data must be stored and which sector rules apply. Healthcare in the US needs a business associate agreement with every vendor that hears or stores patient data. Card payments should never be spoken to the agent; send a payment link instead. In the EU, people must be told they are talking to an AI, a duty that has applied since 2 August 2026. Outbound calls need prior consent in the US, under the TCPA since the FCC’s 2024 ruling on AI voices, and in the UK under PECR.

## 4. Shortlist two platforms

Match the constraints to the table. If your phone system needs a SIP bridge, or data must stay in your own cloud, include LiveKit or Pipecat. If you need several languages or WhatsApp, include ElevenLabs. If cost per minute on US lines is the main concern, include Telnyx. For most other small-business lines, Retell or Vapi is a sensible starting pair.

## 5. Test both on your own calls

Build the same small agent on both: one call type, one tool, the same instructions and knowledge. Write twenty test scripts that include interruptions, background noise, a long reference number and a caller who asks for something the agent should not do. Run them over a real phone line, not the browser test panel.

Measure the delay from the call recordings, from the moment the caller stops to the first sound of the reply. An independent test in 2026 found hosted platforms at 1.3 to 1.7 seconds at the median on real calls, about half a second slower than their dashboards showed. Record the median and the slowest 5% of turns. Read the calls where the agent was interrupted most; that is where turn-taking problems show.

## 6. Price it all-in and plan the exit

Add every cost per minute: platform fee, transcription, model, voice and the phone line, plus per-call charges such as answering-machine detection on outbound calls. Some providers bill in whole minutes, which raises the real cost of short calls. Multiply by your call volumes from step 1. Most hosted stacks come to $0.06 to $0.20 a minute.

Then check the exit. Providers change quickly: in the past year one voice company closed after an acquisition and another is ending its API. Keep your prompts, knowledge, test scripts and integrations in your own accounts, in a form you could rebuild on another platform within days.

## Common mistakes

- Choosing on the demo voice. On a phone line the gap between premium and mid-priced voices mostly disappears.
- Trusting dashboard latency. Measure from recordings.
- Testing in the browser only. Phone audio, carriers and transfers behave differently.
- Letting the agent confirm actions it has not verified. It should wait for the system’s reply.
- Putting permissions in the prompt. Give the agent tools that cannot do the wrong thing.
- Starting with outbound sales calls. Consent rules are strict, and inbound calls are easier to get right first.

## Checklist

- [ ] A month of calls grouped by reason, with counts
- [ ] The first call type chosen, with a hand-over rule
- [ ] The route into the agent agreed with whoever runs your phone system
- [ ] Languages and accents listed, with sample recordings
- [ ] Data location and sector rules written down (HIPAA, PCI, GDPR)
- [ ] Outbound calls only to people with recorded consent
- [ ] Two platforms built and tested on the same script
- [ ] Latency measured from call recordings, at the median and the slowest 5%
- [ ] Cost per minute calculated with every component, on your call lengths
- [ ] An AI disclosure in the first sentence and a recording notice
- [ ] Card payments kept off the voice channel
- [ ] Prompts, knowledge, tests and integrations stored in your accounts

## Sources

- Retell AI, [Pricing](https://www.retellai.com/pricing)
- Vapi, [Pricing](https://vapi.ai/pricing)
- ElevenLabs, [Agents platform](https://elevenlabs.io/agents)
- Telnyx, [Voice AI agent pricing](https://telnyx.com/pricing/voice-ai-agents)
- Deepgram, [Pricing](https://deepgram.com/pricing)
- Cartesia, [Pricing](https://www.cartesia.ai/pricing)
- Bland AI, [Pricing](https://www.bland.ai/pricing)
- OpenBenchmarks, [Voice agent latency on real phone calls](https://openbenchmarks.com/voice-agent-latency/lowest-latency-voice-agent-platform)
- Meta, [WhatsApp Business calling](https://developers.facebook.com/documentation/business-messaging/whatsapp/calling)
- LiveKit, [Solving end-of-turn detection](https://livekit.com/blog/solving-end-of-turn-detection)
- Caller.digital, [Narrowband telephony speech benchmark, India 2026](https://caller.digital/blog/8khz-narrowband-telephony-asr-tts-benchmark-india-2026)
- Jones Walker, [EU AI Act Article 50 timing](https://www.joneswalker.com/en/insights/blogs/ai-law-blog/yes-august-2-still-matters-the-eu-approved-a-high-risk-ai-delay-but-most-trans.html)
- SBD, [AI voice agents](https://smallbaydigital.com/solutions/voice-agents/)

## Questions

### Which voice agent platform is best?

There is no single best one. As of October 2026, Retell suits most small-business phone lines, Vapi suits builds that need every component chosen separately, ElevenLabs is strongest on voice quality, languages and WhatsApp, Telnyx is the lowest cost for US lines, and LiveKit or Pipecat suit existing phone systems, data residency and high volume.

### How much does a voice agent cost per minute?

Usually $0.06 to $0.20 a minute all-in, counting the platform, transcription, model, voice and phone line. Speech-to-speech on the most expensive realtime models runs to $0.35 to $0.45 a minute.

### How fast should a voice agent reply?

Aim for about a second between the caller finishing and the agent speaking. An independent test of five hosted platforms on real phone calls measured 1.3 to 1.7 seconds at the median, and the platforms’ own dashboards reported about half a second less.

### Should we use a speech-to-speech model?

Only when sounding natural matters more than control. A cascaded pipeline (speech to text, a language model, then a voice) lets you use any model, keeps a transcript of every turn and lets you swap parts on their own. Keep a cascaded fallback if you do use speech to speech.

### Can we connect a voice agent to our existing phone system?

Usually, over SIP or by forwarding. Microsoft Teams needs Direct Routing. 3CX authenticates differently from most hosted platforms, so it needs a SIP bridge such as LiveKit’s gateway.