Reselling AI Voice Agents: Costs, Pricing and What Breaks

Reselling AI voice agents works when you sell a narrow job, price it per minute with a real margin, and treat the phone call as the first step of a workflow instead of the whole product. The agencies I see getting burned sold a “virtual receptionist that handles everything” for a flat fee, then found out on the first busy Monday what live calls sound like.

Here’s what I’d tell an automation freelancer whose client just asked if they “do those AI phone things,” with real vendor prices and the telephony and consent details that decide renewals.

Which AI voice agent use cases actually work live

The use cases that survive real callers are narrow and forgiving: after-hours answering, message taking, call routing, and short inbound lead qualification. Appointment booking works when the calendar is simple. Open-ended customer service, outbound cold calling, and anything that needs perfectly captured spelled-out data tend to demo beautifully and then fall apart on live calls.

In a demo, you’re the caller, asking the questions the prompt was written for. A real caller is driving with the window down, talks over the greeting, and opens with a question about an invoice from last spring.

What holds up:

• After-hours answering. The bar is voicemail. An agent that answers the usual questions, takes a message and texts the owner a summary clears it on day one.
• Receptionist triage. Working out why someone called and routing them is a classification problem, and current models are good at classification.
• Inbound lead qualification. Five or six questions, a score, a handoff. Keep the script short.
• Appointment booking, with conditions. One calendar with fixed slots is fine. Multiple providers, rooms and reschedules produce edge cases the agent gets wrong while sounding confident.

What demos well but fails live:

• Outbound cold calling. The FCC ruled in February 2024 that AI-generated voices count as “artificial” under the TCPA, so robocalls using them need the called party’s prior express consent[8]. I won’t resell cold outbound to anyone.
• Capturing spelled data. Emails, street names and policy numbers get mangled on a phone line. Have the agent read values back, or text the caller a link to a short form.
• Complaints and billing disputes. An upset caller wants a person. Get them to one fast.

What reselling AI voice agents costs per minute

The cost of an AI voice agent is a stack of three layers: telephony (per-minute carrier charges plus number rental), the speech and language models, and a platform fee if you use a hosted voice platform. On published prices, hosted platforms with mid-range models land around 9 to 15 cents a minute. A self-built stack can come in under 5 cents before your own hosting.

Telephony

Twilio’s published US rates are $0.0085 a minute for inbound calls to a local number, $0.022 for toll-free, $0.014 a minute outbound, and $1.15 a month for a local number. Recording adds $0.0025 a minute, streaming audio to your own model server adds $0.0044, and letting Twilio run the speech layer through ConversationRelay costs $0.07 a minute[1].

The models

OpenAI lists gpt-realtime-2.1 audio at $32 per million input tokens and $64 per million output tokens, with the mini version at $10 and $20[5]. At one token per 100 ms of caller audio and one per 50 ms of model audio, that’s about 1.9 cents per minute of caller speech and 7.7 cents per minute of agent speech, and later turns cost more because the whole conversation goes back to the model every turn[6]. Google prices Gemini 3.8 Live by the minute: $0.005 for audio in and $0.018 for audio out[7].

Hosted voice platforms

• Vapi: $0.05 a minute for hosting, with transcription ($0.0095 to $0.0099), LLM ($0.0077 to $0.0452) and voice ($0.0146 to $0.0238) passed through at cost. Telephony is billed separately[2].
• Retell: $0.055 a minute for its voice engine and $0.015 for US telephony, plus an LLM that ranges from under a cent to $0.32 a minute. Their headline range is $0.07 to $0.31[3].
• Bland: $0.14 a minute on the free Start plan, or $0.12 on Build with a $299 monthly platform fee, covering the LLM, transcription and voice. Telephony is separate[4].
• Synthflow: no public per-minute rate. Custom enterprise plans start at $30,000 a year[20].

Vapi plus Twilio inbound comes to roughly 9 to 14 cents a minute. The DIY route (Twilio inbound, audio streaming, recording and Gemini Live) is about 3.8 cents even if the agent talks the whole minute, but now you run a realtime audio server around the clock and own every outage. Start hosted. Revisit once one client’s volume makes the difference worth it.

How an AI voice agent agency should price it

Price voice agents in three parts: a setup fee for building the agent and testing it against real calls, a monthly fee that covers monitoring and tuning plus a bundle of included minutes, and a per-minute overage with real margin over your cost. Never sell unlimited minutes. One busy season turns a flat fee into a loss.

Setup pays for the call flow, the integrations and test calls until it stops embarrassing you. The monthly fee pays for listening: someone reads transcripts every week and fixes the prompt, and that someone is you.

For a dental office doing 800 minutes a month on a hosted platform at about 13 cents, usage is roughly $104. I’d charge $1,500 for setup, $450 a month with 500 minutes included, and $0.35 a minute over. At 800 minutes that’s $555 a month against about $110 in cost. Those are my numbers, not a market survey.

The overage rate protects you. A client who doubles their call volume should double your revenue, not your bill. I go deeper on setup versus monthly fees in pricing automation retainers, and if the client would rather pay the model provider directly, see BYOK for agencies.

Telephony: numbers, subaccounts, caller ID and texting

Give each client their own phone number inside their own Twilio subaccount, agree in writing who owns that number, and register the client’s brand for A2P 10DLC before the agent sends a single text. If the agent also calls people back, get the client’s business verified so outbound calls carry the best caller ID attestation available.

Who owns the number

The number is the client’s business identity, so it belongs to them and leaves with them. Twilio subaccounts separate each customer’s usage, bill back to your parent account, and let you move a number between accounts, though A2P registrations and Trust Hub enrollments must be resubmitted after a move[9]. The broader question is in who should own client automations.

If the client already has a number, don’t port it on day one. Have them forward calls to the agent’s number, so switching the agent off is one setting on their side.

Caller ID and STIR/SHAKEN

STIR/SHAKEN is the caller ID authentication the FCC requires voice providers to use, so carriers can check that a call really comes from the number it shows[10]. It matters for outbound calls. Twilio signs calls at one of three attestation levels, and level A means the caller is known and has the right to use that number[11]. The business being attested is your client, not you, so verify per client. A callback that shows up as “Spam Likely” defeats the point.

A2P 10DLC if the agent texts

Booking confirmations and owner summaries are texts, and US business texting from local numbers falls under A2P 10DLC: register a brand and a campaign, or pay extra carrier fees on unregistered traffic. If you register on behalf of clients, each needs its own brand and campaign[12]. Start it during setup. It isn’t instant.

Handing off to a human without losing the caller

Every voice agent needs a way out to a person, and the handoff is where most live failures happen. Use a warm transfer where the agent briefs the human before connecting the caller, screen the transfer so a voicemail box can’t accept it, and decide in advance what happens when nobody picks up.

On Twilio, a completed dial means “the called party answered the call and was connected to the caller”[13], and a voicemail greeting answers a call just like a person does. The plumber’s cell rolls to voicemail, the agent thinks the transfer worked, and the customer ends up leaving the message they could have left in the first place.

The fix is a screen on the human’s side. Twilio can run instructions on the called party’s end after they answer but before the two sides connect, including a chance to decline[14]. “Press 1 to take a call about a leak” briefs the person, and a voicemail box can’t press 1. Answering machine detection is the alternative, but Twilio says it “will not always return the right answer”[15]. If nobody accepts, the agent comes back, takes a message, and creates a callback task.

Call recording consent: one-party vs all-party states

Most US states let you record a call with one party’s consent, but a group of states requires every party to agree, and California is the one that generates lawsuits. The safe default for a voice agent is to disclose recording in the greeting on every call, in every state. This isn’t legal advice. Have the client’s lawyer sign off.

California Penal Code 632 covers recording a confidential communication “without the consent of all parties,” with fines up to $2,500 per violation[16]. The bigger exposure is civil: Penal Code 637.2 lets an injured person sue for the greater of $5,000 per violation or three times actual damages[17]. Multiply by a few thousand calls a month.

Other all-party states include Florida, Illinois, Maryland, Massachusetts, Montana, New Hampshire, Pennsylvania and Washington[18]. The caller’s location matters, not only your client’s. I also have the agent say it’s an AI. Put your retention period for recordings and transcripts in the contract.

Latency: what callers will tolerate

Callers judge a voice agent by the gap after they stop talking. In human conversation, the most common gap between turns is under 200 milliseconds, across languages. An agent that takes a second or more feels broken, and callers start repeating themselves, which makes it worse. Aim for well under a second, measured on a real phone line.

The 200 ms figure comes from a study of turn-taking in ten languages[19]. You won’t hit it, but you must stay out of the range where the caller thinks the line dropped. The usual culprits are chained transcription, LLM and voice steps, an oversized model, and tool calls mid-turn. A two-second calendar lookup needs a filler line like “one sec, let me check.” If the caller interrupts, the agent has to stop talking. And test from a cell phone, because the browser demo lies.

The post-call automation is where the retainer lives

When the call ends, the post-call workflow should write it into the client’s CRM, attach a short summary and outcome, trigger the follow-up, and flag calls the agent handled badly. The voice agent is the part clients see, but this workflow is the part they’d miss most if you switched it off.

• CRM update. Structured fields (name, reason for call, lead score, outcome), not a raw transcript in a notes field nobody reads.
• Summary to the owner. A three-line text or Slack message within a minute of hang-up.
• Follow-up. A booking confirmation, or a callback task when a transfer wasn’t accepted.
• Quality review. Early hang-ups and failed transfers go into a weekly review list. That list is what your monthly fee pays for.
• Minutes per client. So overage billing isn’t guesswork.

At five clients that’s an operations job, covered in managing client automations. The pattern works on any stack today: have the voice platform send an end-of-call webhook to your automation tool and build the rest there.

It’s also why we’re building a Voice Agent node into TaskJuice’s workflow editor and testing it now. The call should be one step, with the CRM write and the follow-up right after it, in the same client workspace and run history.

Frequently asked questions

How much should an agency charge for an AI voice agent?

Charge a setup fee, a monthly fee with a minute bundle, and a per-minute overage. Set overage at two to three times your blended cost so model price changes don’t eat your margin. I’d start a small-business receptionist around $1,500 setup and $400 to $500 a month.

Can an AI voice agent make outbound calls to leads?

Only with care. The FCC treats AI-generated voices as artificial voices under the TCPA, so robocalls using them need prior express consent from the person called. Calling back a lead who asked for a callback is different from cold outreach, but get the consent language reviewed before you sell it.

Voice is the easiest automation to sell a small business, because they hear it working. Sell the narrow job and meter the minutes. Then put most of your effort into what happens after the call, because that’s the part that gets a client to renew.

References

[1] Twilio Voice Pricing, United States: www.twilio.com/en-us/voice/pricing/us

[2] Vapi Pricing: vapi.ai/pricing

[3] Retell AI Pricing: www.retellai.com/pricing

[4] Bland AI Pricing: www.bland.ai/pricing

[5] OpenAI API Pricing: developers.openai.com/api/docs/pricing

[6] Managing costs for realtime models, OpenAI: developers.openai.com/api/docs/guides/realtime-costs

[7] Gemini Developer API Pricing, Google: ai.google.dev/gemini-api/docs/pricing

[8] FCC Makes AI-Generated Voices in Robocalls Illegal, FCC: www.fcc.gov/document/fcc-makes-ai-generated-voices-robocalls-illegal

[9] Subaccounts, Twilio Docs: www.twilio.com/docs/iam/api/subaccounts

[10] Combating Spoofed Robocalls with Caller ID Authentication, FCC: www.fcc.gov/call-authentication

[11] Trusted Calling with SHAKEN/STIR, Twilio Docs: www.twilio.com/docs/voice/trusted-calling-with-shakenstir

[12] A2P 10DLC, Twilio Docs: www.twilio.com/docs/messaging/compliance/a2p-10dlc

[13] TwiML Voice: Dial, Twilio Docs: www.twilio.com/docs/voice/twiml/dial

[14] TwiML Voice: Number, Twilio Docs: www.twilio.com/docs/voice/twiml/number

[15] Answering Machine Detection, Twilio Docs: www.twilio.com/docs/voice/answering-machine-detection

[16] California Penal Code Section 632: leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=PEN&sectionNum=632

[17] California Penal Code Section 637.2: leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=PEN&sectionNum=637.2

[18] Recording Phone Calls and Conversations Under the Law: 50-State Survey, Justia: www.justia.com/criminal/recording-phone-calls-and-conversations-under-the-law-50-state-survey/

[19] Universals and cultural variation in turn-taking in conversation, Stivers et al., PNAS (2009): www.pnas.org/doi/10.1073/pnas.0903616106

[20] Synthflow Pricing: synthflow.ai/pricing