Skip to content

Voice AI for Customer Support in 2026: What Works, What Doesn't

Voice agents fit reminders, simple status calls and after-hours triage. Compare Vapi, Retell, Bland and PolyAI prices and work out a call's real cost.

Asad AliFounder & CEO
Updated
Reading time
12 min read

TL;DR:

  • A voice agent is three layers (speech-to-text, a language model, text-to-speech) plus telephony. Platforms such as Vapi, Retell, Bland and PolyAI bundle them.
  • It fits repetitive, bounded calls: reminders, simple status questions and after-hours triage. It is a poor fit for troubleshooting, emotional calls and regulated advice without a person.
  • With prices checked on November 12, 2026, the worked examples below land between about $0.11 and $0.15 per minute including telephony, before any human follow-up.
  • Latency and interruption handling decide whether a call feels natural. Measure them on real calls, not demos.
  • Voice is a channel, not a strategy. Start with one call type and measure how many calls finish without a person.

Voice AI is ready for some support calls and not others. The voices are good enough; what decides success is whether the call type is bounded, whether the agent can hand off cleanly, and whether the per-minute cost still works once you add the calls a person has to finish. This guide covers the stack, current platform prices, what works, what does not, and the cost math.

How this was put together: desk research of the official pricing pages for Vapi, Retell AI, Bland AI, PolyAI, Voiceflow, ElevenLabs, Deepgram, Cartesia, OpenAI and Twilio, checked on November 12, 2026. Cost tables are arithmetic from those prices with stated assumptions, not measured deployments. Chatsy publishes this article and does not sell a voice product.

The 2026 voice AI stack

A voice agent is three layers and a phone line:

Speech-to-text (STT): turns the caller's words into text. Examples with published prices: Deepgram lists Nova-3 streaming at $0.0077 per minute (pay as you go, monolingual), and OpenAI lists gpt-transcribe at $0.0045 per minute and live transcription at $0.017 per minute. AssemblyAI also sells streaming transcription.

Language model: decides what to say and which tool to call. Voice platforms usually let you pick a fast, low-cost model for the conversation and pass its cost through.

Text-to-speech (TTS): speaks the reply. ElevenLabs lists Flash and Turbo at $0.04 per 1,000 characters, which its own table converts to about $0.04 per minute; Deepgram lists Aura-2 at $0.030 per 1,000 characters; Cartesia sells Sonic through monthly credit plans at about 750 to 800 credits per minute of audio.

Telephony: the phone number and minutes. Twilio lists US local numbers at $0.0085 per minute to receive calls and $0.0140 per minute to make them, plus $1.15 a month per local number.

Orchestration platforms bundle these layers and handle turn-taking, interruptions and phone plumbing:

  • Vapi: developer-first. A $0.05 per minute hosting fee, with models passed through at cost.
  • Retell AI: pay as you go from $0.07 to $0.31 per minute depending on the voice, model and telephony you choose.
  • Bland AI: one per-minute rate that includes the model, transcription and voices; telephony is billed separately.
  • PolyAI: publishes monthly plans with included minutes, plus usage-based pricing and enterprise contracts.
  • Voiceflow: a visual builder that deploys to voice and chat; business pricing is by request.

Model providers also sell voice directly. OpenAI lists gpt-live-1 at $0.05 per minute, with backend model and tool use charged separately, and Deepgram sells a Voice Agent API from $0.075 per minute.

Voice AI platform comparison

Prices from each vendor's pricing page, checked November 12, 2026. Verify before signing.

PlatformPublished pricingWhat the rate includesTelephonyBest for
Vapi$0.05 per minute hosting; Core plan $29 a monthHosting only; STT, model and TTS passed through at costIts own SIP and WebRTC free; carriers billed by the providerTeams that want to choose every layer
Retell AI$0.07 to $0.31 per minute; $10 free creditVoice infrastructure $0.055 per minute, plus chosen TTS and model$0.015 per minute (US via Twilio) or bring your ownInbound support with a managed stack
Bland AIStart $0.14 per minute ($0 platform fee); Build $0.12 per minute plus $299 a monthModel, transcription and voicesBilled separately at pass-through costHigh-volume outbound and simple inbound
PolyAIEssentials $49 a month (250 minutes, then $0.30); Plus $199 (1,000 minutes, then $0.25); Scale $299 per location (1,500 minutes, then $0.20); Enterprise customPlan minutes on its platformBring your own carrier over SIP at no chargeLocation-based businesses and larger contact centers
VoiceflowBusiness pricing by request; agencies on usage-based billingBuilder and deploymentNot publishedTeams without engineers building flows visually

Read the "what the rate includes" column before comparing headline numbers. A $0.05 hosting fee and a $0.14 all-in model rate are not the same thing.

What voice AI actually does well today

Three kinds of call suit a voice agent, as long as there is a clean path to a person.

Outbound: reminders and confirmations

Low stakes, high volume and scripted enough that the agent cannot wander far.

Examples:

  • Appointment reminders a day ahead: confirm, reschedule or cancel.
  • Failed payment notices: "your card was declined; would you like a link to update it?"
  • Delivery windows: "your package arrives between 2 and 4; will someone be home?"
  • Follow-up surveys after a support case.

Check the rules for automated outbound calls in each country you call, and keep consent records. Measure how many calls confirm, reschedule or end in a transfer before you scale.

Inbound: routing and simple FAQ

The agent answers, identifies the reason for the call, and either answers from approved content or routes to a queue.

Examples:

  • Opening hours, location and "are you open today?"
  • Order or claim status, if a tested lookup tool exists
  • Walking a caller through a documented password reset

These replace tier-zero call routing. Track the share of calls that end without a transfer on your own lines; it varies too much by business for any industry number to be useful.

After-hours triage

When the alternative is voicemail, the bar is low. A call comes in at 11 p.m.: the agent decides whether it is urgent (lockout, outage, safety), pages an on-call person if it is, and books a callback if it is not. Write the urgency rules down and test them with real past calls.

What voice AI doesn't do well yet

Complex multi-turn troubleshooting

Voice is harder than chat for diagnosis. The caller cannot paste an error message, and the agent cannot show a screenshot or link. "My router has three blinking lights" works far better in chat, where the agent can send steps and images.

Emotionally charged conversations

Cancellations after a bad experience, bereavement, or a complaint that has been escalating for weeks. Tone matters more on the phone, and an always-polite synthetic voice can make things worse. Route these to a person quickly, and test the signals you use (a request for a person, repeated phrases, specific keywords) on real recordings.

Regulated workflows without human oversight

Health, financial and legal advice. A voice agent can collect details and route to a licensed person; it cannot be that person. Use it for intake and triage, with people giving the advice.

Accents, noise and less common languages

Transcription quality varies by accent, line quality and background noise, and major languages are better supported than less common ones. Test on recordings from your real callers before committing.

Latency is everything

The biggest factor in whether a voice agent feels natural is the delay between the caller finishing a sentence and the agent starting to speak. There is no universal threshold, so measure it on real calls and listen to them.

The delay is the sum of these parts:

LayerWhat adds delayWhat to try
End-of-speech detectionWaiting to be sure the caller has finishedTune the voice activity settings on real calls
Speech-to-textStreaming transcription of the last wordsUse a streaming model and the nearest region
Model first tokenModel size, prompt length, reasoning settingsUse a fast model for conversation turns
Text-to-speech first audioSynthesis start timeUse a low-latency voice model and stream audio
Network and telephonyRound trips between providers and the carrierKeep providers in the same region

Reasoning-heavy models and long prompts are the usual cause of slow turns. Many teams use a fast model for the conversation and call a heavier model only for specific steps.

Turn-taking and interruption

Good conversational mechanics matter as much as intelligence. A natural agent detects when the caller is finishing, starts its reply at the right moment, and stops immediately when the caller interrupts, then responds to what they said.

Orchestration platforms handle this for you; it is a large part of what you pay them for. If you assemble raw STT and TTS yourself, budget real engineering time for turn-taking and test it with real callers.

Cost math in practice

Take a mid-sized deployment: 500 calls a day at 3 minutes each is 1,500 minutes a day, or 45,000 minutes over 30 days. Three ways to build it with today's published prices:

BuildPer-minute componentsPer minute45,000 minutes
Retell AI$0.055 voice infrastructure + $0.040 ElevenLabs Flash + $0.024 GPT 5.4 mini + $0.015 telephony$0.134$6,030
Bland AI (Start)$0.14 all-in rate + $0.0085 Twilio local inbound$0.1485$6,682.50
Vapi$0.05 hosting + $0.0077 Deepgram Nova-3 + about $0.04 ElevenLabs Flash + $0.0077 to $0.0452 model (Vapi's calculator range for OpenAI) + $0.0085 Twilio$0.1139 to $0.1514$5,125.50 to $6,813

On PolyAI's Plus plan, the same volume would be $199 plus 44,000 extra minutes at $0.25, or $11,199, so compare its higher-volume plans and enterprise pricing at this scale.

For a human comparison, 45,000 minutes is 750 hours of talk time a month. At 75% occupancy over 160 working hours, one person handles about 120 talk hours, so the volume needs about 6.25 full-time agents. At $20 an hour, that is about $20,000 a month in wages before benefits, tools and management.

The voice numbers assume the agent finishes the call. If only 20% of calls end without a transfer, you pay for the voice minutes and for a person on the other 80%. Measure your completion rate on a pilot before you compare budgets.

Not For You

Voice AI is the wrong tool when:

  • Your callers prefer a person and you have not asked them. Survey before you build.
  • Your call volume is low, for example a few dozen calls a day. Setup and testing will not pay back quickly.
  • Your support is mostly diagnostic or technical. Move it to chat or email, where you can send steps and screenshots.
  • Your calls involve emotional or compliance-heavy topics (mental health, hospice, legal triage, financial advice). Use voice AI for intake only, if at all.
  • Many callers speak a language your chosen stack handles poorly. Pilot carefully with real recordings or wait.

FAQ

What is voice customer support? Any support where the customer talks to your team by phone, voice IVR or voice notes. Voice AI customer support means an AI agent handles some or all of that conversation, usually with a person available for transfer.

Is there AI customer service that works on the phone? Yes. Platforms such as Vapi, Retell AI, Bland AI and PolyAI sell phone agents with published pricing, and PolyAI lists named customers on its own site. Whether one works for you depends on your call types, so pilot on one bounded call type first.

Will customers know they're talking to an AI? Assume many will, and disclose it at the start: "Hi, I'm an AI assistant. I can help with order status and scheduling. Say 'agent' at any time to reach a person." Callers handle disclosure better than they handle finding out later.

Is voice AI more expensive than chat AI? Per conversation, usually yes. In the Retell example above, a 3-minute call costs about $0.40 ($0.134 × 3). A 10-turn chat on Claude Sonnet 5.5 at 2,000 input and 200 output tokens per turn costs about $0.06 in model tokens. Voice can still win if the call resolves something a chat would drag out over a day, so compare on your own data.

Bottom line

Voice AI works for high-volume, bounded calls with a clean handoff. The mistake is treating it as a general replacement for your phone team. It is a channel, not a strategy.

Start with one call type: reminders, order status or after-hours triage. Measure completion rate, transfers and cost per completed call, then add the next call type.

Chatsy does not answer phone calls. It is a chat, email and help-center product: agents answer from your sources in a website widget and hand conversations to your team through live chat on Hobby and above. If you are deciding between voice and chat for a workflow, read AI agents vs chatbots, the live chat response time benchmarks and where support AI is heading. Try Chatsy free or see pricing.

All articles
  • AI & Technology

    Multimodal AI in Customer Support: What's Real in 2026

    Images, voice and screen content can help support teams. See which uses hold up, what an image or a voice minute costs today and where people must review.

    May 20, 2026 · 11 min read

  • AI & Technology

    MCP for Customer Support: What It Is and How to Use It in 2026

    Model Context Protocol connects support agents to tools such as Slack and Stripe. See which servers are maintained, the permission risks and a rollout.

    May 16, 2026 · 11 min read

  • AI & Technology

    AI Chatbot Evaluation and Testing: A Support Team Playbook for 2026

    How to evaluate a customer support chatbot before launch using a 25-question golden set, four scoring dimensions, and the right tooling stack.

    May 20, 2026 · 12 min read

Build an agent from your support content

Add approved help content, test its answers, and decide when a person takes over.

No credit card required