AI receptionist cost: what an AI voice agent really costs a small business in 2026

An AI receptionist (also called an AI voice agent, voicebot, or callbot) costs about $0.08 to $0.15 per minute of conversation when you assemble it on a specialized platform, and €0.21 to $0.49 with a packaged offer, plus the subscription. For a small business that takes 1,000 minutes of calls a month, usage runs from about a hundred to a few hundred euros or dollars. But that’s not where the budget is decided: the real cost is connecting it to your software. Without that, the agent knows neither who’s calling nor which time slot to offer.

Published prices contradict each other from site to site, and many articles quote outdated rates. So on October 9, 2026, I went back to the vendors’ official pricing pages. Amounts exclude tax and are in US dollars or euros depending on the vendor, unconverted: compare like with like. They change fast, so check them on the day you decide.

What does an AI receptionist handle, and where does it stop?

An AI phone answering system picks up, understands what the caller says, answers based on what you’ve given it, and takes action: it books an appointment, creates a request, transfers the call to the right person, or leaves a structured message. A fictional example, from the demo on my home page: a customer calls for Julien at 12:40 p.m., Julien is busy, the AI assistant picks up, notes the request (“see the product in a demo next Tuesday”), and offers a time slot.

What it doesn’t do: answer everything. It answers what it’s been set up for (hours, lead times, order tracking, appointment booking) and hands off the rest, with a summary of the call. It’s an AI voice agent, which makes it an agent in the true sense: no two conversations follow the same path, so it has to decide the next step. If you’re torn between an agent, a chatbot, and a simple automation, my article AI agent vs. chatbot vs. automation lays out the difference.

What does an AI voice agent cost per minute, block by block?

A phone conversation with an AI runs through five building blocks, each billed by the minute: telephony (the number and the call), speech-to-text transcription, the language model that understands and decides, text-to-speech that replies, and the platform that orchestrates it all. Some “speech-to-speech” models replace the three middle blocks with a single one.

Two benchmarks published by the vendors themselves: Vapi’s calculator puts 1,000 minutes a month at $82 to $129, all building blocks included; Retell’s default example comes to $0.11 a minute. My own math, based on these rows: an AI voice agent assembled on a platform costs about $0.08 to $0.15 a minute, and up to $0.40 or more with a high-end speech-to-speech model, which sounds more natural but costs more. This doesn’t include the build, the connection to your software, or ongoing support.

AI receptionist pricing: what do packaged offers cost?

Alongside the platforms you assemble yourself, business phone vendors sell a voice agent that’s ready to configure. Two of them publish clear prices:

Two caveats. At AirAgent, the FAQ says €0.30 a minute for any overage, while the price grid shows tiered rates: ask which one applies. At Aircall, the pricing page also mentions minute bundles and volume prices that don’t quite match the help page: ask for a written quote. Other business phone providers advertise an AI agent or AI receptionist, but I didn’t find a euro price on their French pages clear enough to quote.

Key point Compare in the same currency, at the same volume, and check what’s being counted: minutes, calls, or licenses. In this market, a price without a date is worthless.

AI receptionist cost for my business: three scenarios, 300 to 3,000 minutes

Here’s the monthly usage cost, excluding the build and the integration, for three call volumes. The math is mine, based on the prices listed above.

For the platform, add a phone number ($1.35 to $2 a month). For Aircall, the licenses also cover your team’s phone system: if you already pay for them, only the minutes matter. For AirAgent, the average call length is an assumption: the plans count calls, not minutes.

This table shows one thing above all: at these volumes, an AI receptionist’s usage cost stays modest. The difference between the options lies elsewhere.

Buy a subscription or have your AI receptionist custom-built?

One provider, Code2b, publishes its own benchmarks for a custom-built AI receptionist: setup “from around €2,000,” then €200 to €500 a month for hosting, support, and maintenance. That’s a ballpark, not a standard: it depends mostly on how many tools need connecting. Here’s how I compare the two routes with a business owner:

What moves the budget, either way: the integration (an agent that finds the caller in the CRM, the order status in the ERP, and an open slot in the calendar is worth far more than a smart answering machine, and that’s the job of an API integration); transfers to a person, and the rules and hours that govern them; languages; the hours covered and what happens outside them. If your order management runs on Sage in France, my article on the Sage 100 API covers what’s possible.

What rules apply to an AI receptionist in France and the EU in 2026?

I’m not a lawyer, and this isn’t legal advice. Three sets of rules apply to a voice agent in France, and it’s worth keeping them apart.

Telling callers they’re talking to an AI (EU AI Act, Article 50)

Since August 2, 2026, an AI system designed to interact with people must be built so they know they’re talking to an AI, unless it’s obvious, and this must be made clear “at the latest” at the first interaction. On a call, the greeting takes care of it: “You’ve reached the AI assistant for…” The regulation provides for fines of up to €15 million or 3% of worldwide revenue for this kind of breach; for a small business, the lower of the two amounts is the cap. The same transparency duty applies to a text chatbot, as my article on customer service chatbots and the EU AI Act explains.

Recording and transcribing calls: what France’s CNIL says

The guidance from the CNIL (France’s data protection authority) on monitoring and recording calls dates from 2009 and is flagged as “being updated,” but its principles remain the benchmark: no permanent or systematic recording, a spoken notice at the start of the call, a right to object, recordings kept six months at most, and analysis documents one year. In November 2025, in a decision about a customer service operation, the CNIL again found it excessive to record every call when a sample is enough to meet the quality goal. A voice agent needs to transcribe in order to understand; keeping every recording is a separate question, to settle with your data protection officer or your legal counsel.

Outbound calls: prior consent required since August 11, 2026

French Law No. 2025-594 of June 30, 2025 (Article 13) rewrote Article L223-1 of the French Consumer Code: since August 11, 2026, it’s illegal to cold-call a consumer who hasn’t given prior consent. The law targets outbound prospecting to individuals. It doesn’t cover a phone system that receives your customers’ calls. That’s my reading: have your counsel confirm it if you also call individuals.

What are an AI voice agent’s limits, and how do you keep them in check?

Voice is still harder than text. A benchmark published in March 2026 (τ-Voice) found that the voice agents tested completed 31% to 51% of tasks in clean conditions, and 26% to 38% with background noise and varied accents, compared with 85% for the best text agent; according to the authors, 79% to 90% of failures come from the agent’s behavior. These are test tasks, not small business calls, but the order of magnitude calls for caution. Transcription has its flaws too: a 2024 study of Whisper found that about 1% of transcriptions contained entire made-up sentences.

Hence three guardrails I put on every voice agent:

• a short script and a written scope: what the agent handles and what it doesn’t;
• answers drawn from your documents and your software, not from the model’s memory;
• a transfer or a message as soon as a request falls outside the scope, with the call summary sent to the right person.

As for human review, both modes work for voice too: at first, your team can confirm the appointments the agent books; once things run smoothly, they’re booked automatically, and anything outside the rules always goes back to a person. In service-call businesses like field service and maintenance, this is often where companies start.

Key point An AI receptionist answers what it’s been set up to answer and hands off everything else. That’s what to tell your customers, and what to demand from the tool.

What I don’t recommend

• A voice agent that never hands off. A customer stuck in a loop hangs up, and you’re more likely to lose them than with voicemail.
• Automated outbound prospecting. In France, calling individuals with an AI has required their prior consent since August 11, 2026, and your brand suffers either way.
• Recording every call “just in case.” The CNIL considers it excessive when a sample is enough.
• Choosing based on the per-minute price. At 1,000 minutes a month, the gap between two offers matters less than the connection to your calendar and your CRM.
• Trusting the missed-call statistics that get passed around. Many figures quoted about unanswered calls have no primary source, or come from old US studies. Count your own missed calls over a month: that’s the only number that matters.