The Costs of an AI Voice Agent That Pricing Pages Don't Show

A decision-maker comparing two AI voice agent platforms is usually comparing two numbers: a price per minute and a monthly subscription fee. That's the information vendors publish, and it is accurate. The problem isn't that it's wrong — it's that it answers a different question than the one the decision-maker is actually asking. The pricing page answers "how much does a minute of conversation cost." The budget needs to answer "how much does it cost to handle a call correctly, over three years, inside my actual systems."

Between the two sit five categories of cost that pricing pages don't show. None of them is hidden in a malicious sense — they simply fall outside what the vendor is selling. If you're not yet sure what this technology actually covers, start with What is an AI voice agent?.

The pricing page describes a running cost, not a project cost

First thing to understand: the advertised price is almost always one layer among several, and vendors say so themselves.

On Vapi's pricing page, checked in September 2026, the $0.05/min rate covers platform hosting. Transcription, the language model and voice synthesis are billed on top, at supplier cost, with separate ranges shown for each. Telephony transport is yet another line item on its own.

The same logic applies at Retell: voice infrastructure is listed at $0.055/min, on top of which come voice synthesis, the model, telephony, and then a set of add-ons also billed by the minute — knowledge base, noise cancellation, guardrails, PII redaction. On that same page, the gap between the cheapest and the most expensive language model on the list runs from $0.0016 (GPT 5 nano) to $0.32 per minute at the standard tier (GPT 6 Astra): a 1-to-200 spread, on the same pricing grid, for the same product — and up to $0.64, or 1-to-400, once you factor in that same model's fast tier. Which model you pick weighs more heavily on the bill than which platform you pick. At ElevenLabs, the documentation states plainly: "The LLM model and any telephony are billed separately on top, based on usage."

The second lever, discussed less often, is concurrency: the number of simultaneous calls. Vapi charges $10 per extra line per month beyond the plan's included capacity, Retell charges $8. ElevenLabs applies an overage rate of $0.160/min — double the standard rate — once the limit is crossed. In other words, the number that actually determines your bill isn't the average price per minute, it's the shape of your call peak. A customer service line whose volume doubles between 8 a.m. and 10 a.m. doesn't pay the advertised rate: it pays for its peak, or it puts callers on hold.

Telephony is a project, not a line item

This is the cost most consistently underestimated, because it belongs to a discipline — telecoms — that AI project teams don't usually practice.

Three mechanics are enough to understand the size of the gap.

Destination changes everything. On Twilio's SIP trunking pricing page for France, checked in September 2026, termination to landlines starts at $0.0147/min. Mobile shows two rates on the same line depending on where the call originates: $0.1563/min at the standard rate, $0.0364/min "from the EEA" — and it's this second rate that applies to traffic originated from France, a 2.5x gap versus landlines. Premium-rate numbers, for their part, climb to $0.5473/min: 37 times the landline rate, on the same page, for the same product. An inbound helpline and an outbound reminder campaign calling premium-rate numbers therefore have nothing in common budget-wise, even at an identical platform price.

A transfer runs two meters at once. Twilio's documentation is explicit: each call leg is billed separately, and each leg is rounded up to the next minute. An inbound call transferred to a human agent therefore counts as two legs. This detail has a counter-intuitive consequence: the more often your AI voice agent fails, the more expensive the billed minute becomes, since the transfer minute stacks on top of the minutes the bot already used. Unit cost gets worse in exactly the situation where the project is already going badly.

How you connect to your existing stack shapes the architecture. Depending on the method used, a transfer to a human agent can be a full handoff (SIP REFER) or the bot can stay in the conference to keep transcribing — this second option keeps the platform's meter running for the entire human conversation that follows. On older infrastructure, media travels through centralized gateways using compressed codecs, which adds latency and degrades the audio the model receives. That degradation doesn't show up on an invoice. It shows up in the misunderstanding rate, and from there, in repeat calls.

Content preparation: the human cost nobody budgets for

An AI voice agent doesn't know your return policy, your on-call schedule, or your warranty terms. It reads whatever it's given.

Vendors bill for the container: Retell lists $0.005/min for using a knowledge base, and $8/month per base beyond the first ten. Nobody bills for the content, because nobody produces it on your behalf.

That work is rarely technical. It means identifying which documents are authoritative, resolving contradictions between versions, dating material, retiring what's obsolete, and structuring what's left. It therefore falls on business owners — precisely the people whose time was never blocked out in the project plan. How much of an AI project's time goes into data preparation is still debated, and the often-cited "80%" figure is itself contested: an analysis tracing its origin traces it back to more modest surveys — 60% in 2016, 51% in 2017, 73.5% of respondents saying they spend at least a quarter of their time on it in 2019 — and calls it a statistic repeated without a solid basis. None of those surveys, however, describes this line item as marginal.

What makes this cost mechanism dangerous is that it's recurring. A procedure that's changed but never re-indexed will keep producing confident, outdated answers: the retrieval system picks the passage closest to the question, with no way of knowing that one document has replaced another. Content preparation isn't a launch task, then — it's an ongoing governance cost. A housing association that updates its on-call instructions twice a year needs to budget two review cycles a year, not a one-off setup project.

Error recovery and oversight get billed twice

No AI voice agent resolves 100% of calls. The self-resolution rates published across the market vary widely, and it's worth naming where they come from: the 60-to-90% ranges you'll see quoted are published by vendors or firms tied to the industry, not by independent third parties. We don't repeat them as our own figures — which is also why we publish our testing methodology.

What is documented, on the other hand, is that the market bills oversight as a line item in its own right. On Retell's pricing grid, automated conversation quality assurance is listed at $0.10/min after 100 free minutes — nearly double the price of the voice infrastructure itself ($0.055/min). When a vendor charges more for checking the work than for doing it, that tells you where the real cost sits.

On top of that comes the human cost, which appears on no pricing grid at all: listening to sample calls, classifying failures, fixing prompts, and handling callers who call back because the first call resolved nothing. The right calculation isn't "cost per bot minute" but "fully loaded cost per request actually resolved," which adds up bot minutes, telecom legs, human agent time on transferred calls, and repeat calls. It's also the only honest basis for comparison against a human service, which French contact centers typically bill as a subscription plus a per-minute conversation rate.

The cost doesn't stop once the system goes live

An AI voice agent isn't a project that finishes: it's an asset that drifts.

Vapi itself identifies three triggers for regression, summarized in its note on evals: a prompt change, adding a tool, or a model change. And the way it describes the failure mode is worth quoting directly: "Nothing crashed. Nothing alerted. Conversions just quietly dropped." The conclusion the vendor draws is blunt: "Code has tests. Prompts and tools do not. That is the gap."

Model changes, in particular, don't run on your calendar. OpenAI's deprecation policy commits to at least six months' notice for generally available models, at least three months for their variants, and considerably less for preview models — and migrating to a new model still requires an application-side change. You therefore inherit a re-testing cycle imposed from outside, with no added business value, on a schedule you don't control. The same applies to integrations: every API change on the CRM or ERP side forces a round of rework.

Compliance falls in the same category. Since 2 August 2026, Article 50 of the EU AI Act requires that a person be informed they are interacting with an AI system, in a clear and distinguishable manner, at the latest at the time of the first interaction; the European Commission's FAQ notes that fines can reach €15 million or 3% of global annual turnover. For an AI voice agent, that disclosure has to be spoken out loud — which means it eats up billed seconds on every single call, both on the platform side and the carrier side. It's a compliance cost that lands, quite literally, on the usage bill.

In France, outbound calls to consumers are subject to an additional constraint that's more structural than budgetary: since 11 August 2026, Article L223-1 of the French Consumer Code, as rewritten by Article 13 of Law No. 2025-594 of 30 June 2025, bans telephone cold-calling without prior consent that is freely given, specific, informed, unambiguous and revocable — with the burden of proof resting on the business. An outbound B2C AI voice agent project with no traceable consent base isn't a cheaper project: it's a project with no addressable market. On call recording, the CNIL (France's data protection authority) sets rules on informing callers, limiting how long recordings are kept, and who is authorized to access them. We cover these obligations in Trust & compliance.

The strongest objection: "these costs exist on every IT project"

An integrator or a vendor could reasonably respond that none of this is specific to AI voice agents. Integration, data quality, oversight, maintenance: these line items weigh on any information-system project, and singling them out here could look like a selective prosecution. They'd also be right to add that a narrow scope — appointment booking, order status, call qualification before transfer — reduces them considerably.

That objection is valid, and the recommendation it implies is the right one. It doesn't dissolve the argument, though, for three reasons.

First, a narrow scope removes some line items, not all of them. Content preparation shrinks to almost nothing for a bot that just reads out opening hours. It removes neither the telephony wiring, nor the transparency obligations, nor the migration forced by a model being retired — the last of which applies just the same to a three-intent agent as to a three-hundred-intent one.

Second, the difference from a conventional project is one of kind, not degree. The central component is non-deterministic, and it's supplied by a third party who retires it on their own schedule. An accounting connector doesn't change behavior because a sentence got rephrased; a language model does.

Third, the argument "these costs exist everywhere" isn't reassuring — it's corroborating. When Gartner predicted, in a 25 June 2025 press release, that more than 40% of agentic AI projects will be canceled by the end of 2027, the reasons cited were escalating costs, unclear business value, and inadequate risk controls — not the technical failure of the conversational layer itself.

The right question, then, isn't "how much does the minute cost," but "how many of these line items does my use case actually let me avoid."

What to ask before you sign

Four questions are enough to surface the gap between the pricing page and the real budget: which minutes get billed during a transfer, and to whom; what rate actually applies at my peak concurrent-call level, not my average; who produces and maintains the knowledge-base content, and with what business-side time budget; and what happens, contractually, on the day the model you rely on gets retired.

No vendor today publishes a fully loaded cost per request resolved. As long as the billed unit stays the minute rather than the resolution, the gap between the pricing page and the real budget won't be a transparency failure: it will be a property of the business model. Whether a buyer large enough eventually forces the other metric into the open remains to be seen.

---

Disclosure: Call-bot.ai is published by Nehos Groupe - KBC SOLUTIONS, which designs and integrates conversational AI solutions for businesses and publishes its own platform, Alyas.ai (not mentioned in this article), and therefore has an economic interest in the integration costs described in this article. We do not resell any of the platforms mentioned and receive no compensation from them. The rates quoted were taken from vendors' public pricing pages in September 2026 and are subject to change.