What an AI voice assistant can do for your business, and what it cannot

Most pages about voice assistants list what the technology can do and stop there. That is half the question. The other half, the one that decides whether you will regret it or not, is the list of calls that should not be handed to a machine.

This article gives both. Nothing that follows is a promise for later: if something does not work well today, it is written here as such.

Where Swiss SMBs stand

The Swiss economy is made of very small outfits. According to the FSO business structure statistics (STATENT 2023, published in August 2025), Switzerland has 624,219 SMBs, or 99.7% of market economy businesses, of which 561,952 are micro-enterprises, or 89.7% of all market economy businesses. In a business that size, the person answering the phone is also the person on the building site, at the chair or in the kitchen.

AI adoption, meanwhile, is moving fast. The AXA study on the SME labour market 2025, picked up by the Confederation's SME portal, measures 34% of Swiss SMBs deliberately using AI in their processes, against 22% a year earlier. Of those, 57% say they are saving time, 2% have cut staff and 10% have created jobs. That last figure is worth holding on to when the subject is telephony: so far, in practice, AI has mostly moved work around rather than removed people.

On speech recognition specifically, Europe is at the very beginning. The Eurostat figures for 2025 cover businesses with 10 to 49 employees.

So one business in six has a foot in AI, and one in seventeen in speech recognition. The third bar measures language generation, which is very largely written: it is not telephony.

You are not behind.

What a voice AI genuinely does well

Answering, every time

This is the core function, and everything else follows from it. A call comes in at 7.30 in the evening, a second while you are already on the line, a third on a Sunday: each one finds someone at the other end. Not a voicemail box, a conversation. The difference is not cosmetic: according to platform data from Invoca (2021, figures covering their own US customers, so vendor data), 26% of calls from prospective customers go unanswered, and only 2% of those callers leave a message. Voicemail is not a safety net.

Answering common questions, from your information

Opening hours, address, parking, lead times, payment methods, services, what you do and what you do not. The assistant answers from the information you gave it, not from general knowledge. That is the distinction that matters: an AI improvising a price is a problem, an AI reading out your rates is a service.

Booking, moving and cancelling an appointment

The assistant writes into the calendar you already use, offers the free slots, and sends the confirmation back. The demand is there: in its KMU Digital Pulse 2025 (an intervista survey of 1,660 people in Switzerland, plus an analysis of 47,079 Swiss SME websites), localsearch measures that 77% of the population wants to find and book SME services online, while 3% of Swiss SMBs make that possible, and that 56% say online booking decides which provider they pick. To be read at the appropriate distance: localsearch is a Swisscom subsidiary selling digital services to those same SMEs. The same study notes that 39% of respondents could imagine booking an appointment with an AI voice assistant.

The underlying trend is not Swiss, but it points the same way: in 2024, 39.8% of people in the EU27 had booked an appointment with a health professional through a website, against 33.5% in 2022. Switzerland is not in that dataset.

Filtering, routing, transferring

A good share of what rings is neither a customer nor an appointment: cold calling, a delivery, a question meant for somebody else. The assistant identifies the reason, routes it to the right person, and puts the call through to you when that is warranted. See what that looks like in a garage or a property management company, where sorting the inbound flow often counts for more than booking appointments.

Sending you a written summary

After every call, an email: who called, why, what was agreed, what is left to do. You read in fifteen seconds what a voice message would have taken you two minutes to listen to. It is the feature customers talk about most after a month, and rarely the one that convinced them.

Several calls at once, and several languages

A human front desk does neither. Three simultaneous calls at midday are handled like a single one. And a caller who switches to French, German or English is followed without a transfer and without an apology. In a restaurant, those two points alone are often worth the service.

The hours when nobody answers

The argument is simpler than people think, and you can check it by reading the website of the human competition. Telib, a Swiss phone answering service run by people, publishes its hours: Monday to Friday, 8am to 7pm. That is about 55 hours a week, already far more than a practice or a workshop covers on its own.

A week has 168 of them. The rest of the time, nobody answers: Tuesday at half past seven in the evening, Saturday morning, Sunday, public holidays. Those are precisely the hours when someone who has just broken a tooth, locked themselves out or found a leak dials a number, then the next one.

What it does not do well

Here is the part most providers do not publish: the situations where a voice assistant is the wrong tool, and where you have to hand over.

A complaint

Someone who is unhappy is not calling for information. They are calling to be heard, and often to get something done about it. A machine can take the message cleanly, but it can neither acknowledge fault nor decide what you are going to do. An assistant that tries to see a complaint through to the end turns an annoyed customer into a lost one.

An angry caller

The tone rises, and from there every automated sentence makes it worse. The rule that works is simple and effective: at the first sign of irritation, transfer, without insisting and without going back through the menu. It is a rule you set once.

Anything calling for judgment or a commercial concession

A discount, an exceptional deadline, an edge case that resembles nothing you planned for, a quote that depends on three things the caller does not know yet. Those are decisions, not answers. A well configured assistant says that a person will call back, notes what needs to be known, and stops there.

The line dates from 2022 and still holds.

A caller who does not want to talk to a machine

It happens, and it is not a problem to be solved. Some people hang up at the first sentence, others ask for a person straight away. What works is a short, announced way out. Hiding the fact that this is an AI is the worst option, commercially as much as legally: we set that out in our article on data protection.

Who handles what

Reduced to a table, with the call types an SMB actually receives.

How to tell whether it is worth it for you

To settle it, ask yourself the questions below. None of them asks for figures you do not have.

• How many calls arrive when nobody can answer? Look at last week's missed calls on your phone. If the answer is two, keep the setup you have.
• What do those calls look like? If they are mostly questions and appointments, an assistant handles them. If they are mostly ongoing cases and complaints, it does not.
• Who answers today, and at what cost? That is the useful comparison, and it is rarely about the technology. We made it in detail between an assistant and a phone answering service in a dedicated article, and the prices of the three plans are on our pricing page: CHF 99, 299 or 599 a month, excl. VAT, for 500, 1,500 or 3,000 minutes.

The honest conclusion is a split, not a replacement. It is also what Gartner says after surveying customer service leaders:

Applied to an SMB's phone, that gives a simple rule: the assistant takes everything that comes in, and you decide what makes its way up to you.