AI Voice Agent for Sales Calls with Voice AI Platform Pricing and How to Pick One

What actually makes an AI voice agent sound human on a sales call

Four things decide whether a prospect keeps talking, and only one of them is the voice itself.

Latency is the first and the one most teams underestimate. Human conversation turns over in roughly 200 milliseconds. Every extra half second before the agent responds reads as hesitation, and a full second reads as a bad connection, which is when people say hello twice and hang up. The second is barge-in, meaning the agent stops talking the instant the prospect starts. An agent that finishes its sentence over someone is unmistakably a machine, and it is the single most common tell. Third is how it handles the unexpected: not the objections you scripted, but the person asking what company you are calling from, or saying something the model has no branch for. Graceful recovery beats a wider script tree. Fourth is prosody, the rhythm and emphasis of the speech, which is where synthesis has improved most and where buyers now spend the least time worrying.

The test that matters is not whether it passes for human in a demo, it is whether the conversation reaches a useful outcome. Our agent discloses that it is AI in the opening line, which sounds like it should hurt and largely does not, because the reason people hang up is that a call is irrelevant or badly timed, not that it is automated. What disclosure does buy you is a call that stays inside the FCC position, confirmed in February 2024, that AI-generated voices fall under the TCPA. The detail is on our TCPA compliant AI calling page.

• Response latency: human turn-taking is roughly 200ms, and delay reads as a bad line
• Barge-in: the agent must stop the moment the prospect speaks
• Recovery on unscripted questions matters more than a bigger script tree
• Prosody and pacing, where voice synthesis has improved most
• AI disclosure in the opening line, per the FCC position of February 2024

Build your own voice agent or buy one that already works

There are two genuinely different products sold under this name, and buying the wrong one wastes a quarter.

Developer platforms give you the pieces. You pick the speech-to-text, the model and the voice, wire up telephony, write the conversation logic, and run the whole thing yourself. Pricing is per minute and looks cheap on the pricing page: Vapi publishes $0.05 a minute for agent hosting, with speech, model and voice costs passed through on top, so the headline is a floor rather than a quote. Retell AI publishes $0.07 to $0.31 a minute depending on the model and voice you pick. Bland.ai publishes $0.14 a minute with no platform fee on its entry tier. These are real options and we compare them honestly, including where they win, on our Vapi, Retell AI and Bland.ai pages. What the per-minute rate does not include is the engineering time to build the agent, the iteration to make it actually book meetings, and someone owning compliance.

The done-for-you option is the other end. The agent, the script, the objection handling, the calendar integration and the compliance controls arrive configured, and you supply the list and the qualifying criteria. That is what we sell, at a flat managed fee rather than per minute, and the honest rule of thumb is that platforms win when the conversation is unusual enough that you need to control every branch, while a managed agent wins when the job is standard outbound qualification and the constraint is that nobody has time to build it. The third option, which is neither, is renting people. We priced that out from the agencies own pricing pages on cold calling services, and the broader category comparison is on best AI cold calling software.

• Developer platforms: you assemble speech, model, voice, telephony and logic yourself
• Vapi publishes $0.05/min hosting with speech, model and voice passed through on top
• Retell AI publishes $0.07 to $0.31/min depending on model and voice
• Bland.ai publishes $0.14/min with no platform fee on its entry tier
• Per-minute rates exclude build time, iteration and compliance ownership
• Managed agents suit standard outbound qualification; platforms suit unusual conversations

How much does an AI voice agent cost?

It depends entirely on which of the two products you buy, and the two are an order of magnitude apart at the same call volume.

Build-it-yourself platforms bill per minute and publish their rates. Vapi charges $0.05 a minute for agent hosting with speech, model and voice passed through at cost, or $0 for those components if you supply your own API keys, and includes 4 concurrent lines on usage only (10 on its $29 a month Core package) with additional lines at $10 each a month. Retell publishes $0.07 to $0.31 a minute and breaks the stack out: roughly $0.055 for voice infrastructure, $0.015 for text to speech, $0.015 for US telephony, and the model on top. The width of that range is almost entirely which model you pick. Bland charges $0.14 a minute on its zero-fee entry tier and is the only one of the group that puts the model, transcription and premium voices inside the per-minute rate rather than passing them through.

ElevenLabs, now branded ElevenAgents, prices in bundled minutes: $22 a month for 275 minutes on Creator, $99 for 1,238 on Pro, $299 for 3,738 on Scale. Additional minutes are $0.080 on every tier, which produces a genuinely counterintuitive result. At 1,000 minutes a month, Creator plus overage is $22 plus 725 times $0.080, so $80.00, against $99 for Pro. Moving up a tier buys concurrency and seats, never a cheaper minute.

Managed agents are priced as a service. Ours is a flat monthly fee with connected minutes included, published at $499, $1,490 and $3,900 a month, with a larger connected-minute pool on each higher plan. The comparison that matters is not the per-minute rate, it is total cost including the engineering time to build the agent, the weeks of iteration before it books anything, and someone owning compliance. We laid all nine platforms out with their published figures on AI voice agent pricing.

What is a voice AI platform, and do you need one?

A voice AI platform is developer infrastructure. It gives you speech to text, a model, text to speech, telephony and an orchestration layer, then leaves the conversation design, the integrations and the compliance to you. An AI voice agent is the finished thing that gets on the phone. Every one of the platforms below can build one; none of them arrives as one.

The question to ask is what your team is short of. If you have engineers and an unusual conversation, a platform is the right purchase, because you can control every branch, swap the model when a better one ships and keep your data inside your own accounts. Vapi's zero-cost pass-through with your own keys is genuinely attractive if you already hold model contracts.

If you are short of engineering time and the conversation is standard outbound qualification, a platform is a project rather than a product. The per-minute rate is the smallest line in that budget. Build a prototype in a week, spend a quarter making it book meetings reliably, and you have paid several times the annual cost of a managed agent in salary before the first booked call.

One detail that trips up buyers on both paths: concurrency. Every platform meters how many calls you can run at once, separately from minutes. Vapi now includes 4 on usage only and 10 on its $29 Core package, and sells more at $10 a line. Retell includes 20 free and charges $8 per additional line a month. Bland caps concurrency at 10, 50 and 100 by tier and, less obviously, caps calls per hour at 100, 1,000 and 1,000, so its top tier's 5,000-a-day allowance runs into the same hourly ceiling as the middle one. If you dial in bursts, model concurrency before you model minutes.

What to check in an AI voice agent platform comparison before you sign

Five things separate these products, and price is not the most important one.

Latency is first, because it decides whether the call survives. Human conversation turns over in about 200 milliseconds. Test the actual round trip on your own telephony, at your own concurrency, not in the vendor's demo, because the demo is not running 40 simultaneous calls.

Second is what the per-minute rate includes. Bland's rate includes the model, transcription and premium voices. Vapi's and Retell's do not. Two quotes that look 10 cents apart can be identical or double, depending on which components are inside the number, so normalize before comparing.

Third is data retention, which is quietly a compliance question. Vapi keeps call history for 14 days on its entry tier and chat for 30, and sells HIPAA at $2,000 a month; zero data retention now comes with its $29 Core package. If you need either, they are not add-ons you bolt on later at small-team pricing.

Fourth is who owns the compliance surface. On a platform, you do. The FCC confirmed in February 2024 that AI-generated voices fall under the TCPA, which means disclosure, consent, DNC suppression and calling-window enforcement are your build, not the vendor's. That work is on our TCPA compliant AI calling and TCPA compliance software pages, and it is the single most underestimated line item in a build-versus-buy comparison.

Fifth is whether the vendor will sell to you at all. Some voice and dialer vendors in this market carry seat floors, minimum commitments or annual-only contracts. Check that before you invest evaluation time. The comparison against traditional dialing is on AI dialer, and against human SDR cost on AI SDR.