Essay

An AI answering your phone is only as good as your calendar

Since October 20, an AI that answers the phone, in Dutch too, is a building block costing a few cents a minute. The voice is solved. What that AI can actually do depends on your own records.

October 26, 2026 · Levy van Bremen · Lees in het Nederlands

On October 20, the French company Mistral AI released Voxtral 3: a set of speech models plus a toolkit that lets developers build an AI agent for the phone. That is software that holds a conversation on its own and gets things done along the way. It listens, thinks, looks something up and talks back. Dutch is included, and calls can run through servers in the EU or on your own hardware. According to Mistral, each turn in the conversation takes less than a second, and the AI stops talking as soon as the caller cuts in. With the smaller of its two reasoning models, it costs 4 US cents a minute. A five-minute call comes to twenty cents.

Mistral is not alone. In September ElevenLabs, known for its AI voices, launched Reception, an AI receptionist for small businesses such as plumbers, estate agents and hair salons. It picks up when the team is busy or the business is closed, and sends a summary after every call. A day after Mistral, the Korean company Channel Talk launched a voice AI that not only talks to customers but also calls them itself.

Talking on the phone to a computer that sounds like a person has become cheap and ordinary. My argument: from now on, something else makes the difference, namely what that voice is allowed to look up and record during the call, and when it hands over to a person. A business whose calendar, prices and job status are not in a system is buying a polite answering machine.

A voice without access is an answering machine

Look at the examples Mistral gives itself: verifying a caller's identity, checking a bill, booking an appointment. All three are actions in another system, through a connection: a fixed entrance through which one piece of software may look into another or record something there. Reception, too, books appointments straight into a connected calendar.

Think of the questions customers actually call with. Can I come in on Thursday? When is the engineer coming? Is my car ready? Has my order arrived? An AI can only answer them if the answer is somewhere it can reach. If the schedule hangs on a planning board in the workshop, lives in the owner's head or sits in a group chat, one sentence remains: “I'll pass it on.” An answering machine could already do that.

What gets solved quietly today

This is the work most businesses underestimate. Whoever answers your phone today fills gaps without anyone noticing. They know the boss doesn't take appointments on Fridays, that one regular customer always calls about the same boiler, and that “next week” from the painter usually means the week after. None of it is written down. An AI only knows what has been written down and what it is allowed to look up.

So if you want an AI to answer the phone, you first have to get that quiet knowledge out of people's heads. Which calendar is the real one, if there are two? Which prices can you say out loud, and what do you say when the honest answer is “it depends”? Who can be disturbed for an emergency, and who never? It is dull work. It also produces something valuable regardless of the phone: a business that keeps running when the person who usually answers is ill or on holiday.

The strongest objection: people want a person

The obvious objection is that customers don't want to talk to a machine, and the numbers back it up. Of 503 consumers surveyed by research firm Metrigy, 84.7 percent would rather deal with a person than an AI agent. Even when assured their problem will be solved, 80.1 percent still choose a person. In February and March, Gartner surveyed 3,566 business and consumer customers. Half say their dealings are easier when companies use AI, but 87 percent say it is essential that there is then always a way to reach a human.

Dutch regulators take the same line. Last year the ACM, the Dutch competition and consumer authority, and the Dutch Data Protection Authority wrote that consumer law requires businesses to communicate directly and effectively, and that to them this means human contact must always be possible. They explicitly mention the phone. Since August 2 of this year, a caller must also be told at first contact that they are talking to AI; the European AI Act requires it. And the regulators warn that these AI models tend to tell people what they want to hear, improvise and create false certainty.

I take that objection seriously, but it compares the wrong things. For a small business, the AI's real competitor is the phone nobody answers at half past six, or the one that rings while the engineer is up on a roof. At those moments, the caller chooses between a machine and voicemail.

Reception lets you set up exactly that: your team answers first, and the AI takes whatever is left over or comes in after closing time. A Gartner analyst puts it bluntly: don't make AI the mandatory first step for every question. Also look at what people do find AI useful for. In the Metrigy survey, about half mention being routed to the right person, and almost half mention confirming orders and scheduling or rescheduling appointments. Those are exactly the tasks that only work if the AI can look into your systems.

The improvising is the more serious point. On the phone, a confident “yes, Thursday works” costs more than “I don't know, we'll call you back”. The most important sentence for an AI on the phone is therefore the one in which it admits it doesn't know something. I see this in my own work. In Gemeenteregels, an app of mine that summarises municipal rules, I rewrote the text this week for the moment the summary finds no answer. It looks like a detail. It decides whether people dare to believe the answers the app does give.

What you need to decide now

The temptation is to start with the vendor and the voice. Start with your own phone. For one week, write down what people call about, and next to each question note where the answer lives: in a system software can reach, on paper, or in someone's head. Only the first group can be handled by an AI today. The rest is homework, and that homework makes your business stronger even if you never let an AI pick up.

Then decide three things before switching anything on. Which questions always go to a person? How does a caller reach a staff member with one sentence or one key press? And in what form does the summary reach you, so nobody has to ask the customer the same thing twice? Start with the hours when nobody answers now.

If less than half of the answers are in a system, make your calendar and customer list reliable first. The voice is already here. It can wait.

Sources

Is there work in your business that comes back every week? That is what I build software for.

Book a call →