Bringing AI into your business

Where AI pays off and where it does not

AI pays off where language turns into work: drafting, summarising, sorting, translating. It does not pay off where someone has to take responsibility, where a mistake lands directly with a customer, or where the work needs hands rather than words. Drawing that line before any tool gets bought is the cheapest step of the whole undertaking - the most expensive AI projects are the ones that start in the wrong place.

For the office side of a small business, that means: the reply to a customer enquiry, the cover letter for a quote, the minutes from a meeting, the summary of a long report. Here AI delivers a usable first draft in seconds, which a person sharpens and stands behind. In the workshop, on the building site, in care work, the work itself stays untouched; what changes is the paperwork around it.

The other side matters just as much. AI takes no responsibility: what it writes sounds confident even when it is wrong, and it does not notice. It knows nothing about your business, your customers or your prices; everything it should know, you have to give it. And it replaces no decision - whether the quote goes out is not the tool's call.

Deep DiveWhat AI is and what it can do: the calm primer this guide builds on→

The first use case

The most tempting question is the wrong one: what could we do with AI? It produces lists of twenty ideas, none of which ever ships. The right question is smaller and less comfortable: which single case goes first?

Three questions separate a workable first case from all the others. Does the task repeat - does it come up every week, not twice a year? Is it made of language or text - writing, summarising, sorting? And does a person check the result before it leaves the company? Only a task that answers yes to all three is fit to go first. Everything else is not forbidden - it is simply not first.

Why so strict? The first question secures the payoff: what repeats weekly pays back every invested hour; what happens twice a year never does. The second secures the fit: language models are strong at language, the name says so. The third protects the business: as long as a person reviews every result, a poor AI draft costs a few minutes, not a customer.

Two guardrails come on top, and no switch replaces them. Not the core business first: the first case should take load off your company, not operate on its heart. And not the most delicate process: the opening case should manage without especially sensitive personal data - chapter 4 spells out what falls under that.

Here is what comes out of the switch in practice:

The cover letter for a quote. The numbers you calculate the way you always have. But the letter that explains what is in there and why: AI delivers a draft in seconds that takes you two minutes to bring onto your own tone.

From meeting notes to minutes. Bullet points in, cleanly structured minutes out, with open items and owners. One of the most rewarding tasks there is, because it comes up every week and nobody enjoys it.

The job ad. Five facts about the role become a complete draft. And we draw the line right here: having the ad written, yes - having incoming applications judged by an AI is a different business, with personal data and the risk of judging people wrongly, and it does not belong in the first months.

Sorting incoming enquiries. Every request summarised and classified: urgent or routine, quote or support. The inbox does not get smaller, but it becomes readable.

Just as important is what falls out of the switch. Payroll repeats, but a mistake there reaches your staff unchecked. The final check of your own work is precisely the step that cannot be delegated - to anyone, a model included. And everything legally binding, from contracts to terminations, stays draft work with a human signature underneath, however fluent the draft sounds.

At the end of this step, a single sentence goes on paper, and it is deliberately unspectacular: we are trying AI for the minutes, Andrea owns it, every result gets reviewed. How that sentence becomes a pilot that actually proves something is the next chapter.

The pilot: start small, measure honestly

A pilot is not a gut feeling with an end date. It is a small promise to yourself: one task, one named owner, a fixed period, and written down in advance what you will accept as evidence of value. All of it fits on one page.

The named owner matters more than the tool. Pick someone who wants to do this, not someone who happens to have capacity - curiosity beats availability. The period: four to six weeks. Long enough for an honest picture across several weekly cycles, short enough that nobody can wait it out.

The heart of the page is the measuring line, and it gets written BEFORE the start. "We save time" is not a measurement, it is a hope. A measurement means: write down how long the task takes today, count during the pilot, and ask the team at the end whether they want to keep going. It takes no more than that - but no less either, because without that line the gut feeling decides in the end, and the gut feeling is always right.

At the end of the period a decision falls, and it is a real one: scale it or stop it. The dangerous option is the third one, letting it drift without a decision - a pilot without an end is a subscription without a right to cancel. Stopping is a result too: you then know something about your business you previously only suspected, and it cost you six weeks instead of a bad investment.

Data and data protection

This is where the mines are, and every one of them is marked. The ground rule first: Switzerland's revised data protection act has applied since September 2023 and is written technology-neutral - it covers AI tools exactly as it covers any other software. Responsible for your customers' personal data is you, not the provider.

In practice it helps to think in three tiers. Without personal data - a text about your services, a manual, the job ad - an AI tool is as uncritical as a word processor. With ordinary personal data, meaning names, addresses, correspondence, you need a provider bound by contract as a processor. And especially sensitive data belongs in a tool only after its own review - the law lists the categories: health, religious and political views, biometrics, social assistance, criminal proceedings. For high-risk processing it requires a data protection impact assessment up front.

The good news: for most small businesses, almost everything comes down to the contract. And the questions it has to answer are known - the Swiss federal data protection commissioner lists them for cloud services, and they apply to AI tools one to one:

This is exactly where the free account and the business contract part ways. Free accounts are made for private individuals: what happens to the inputs stays open, or hangs on a setting somebody has to find. Business plans from the large providers typically commit to not using your inputs to train their models, and they govern retention and deletion. "Typically" means: read it, don't assume it - with a serious provider, the commitment is there in writing.

That leaves the location. Switzerland or the EU as a data location keeps things simple; the US has worked since September 2024 through certified providers - but then the certification is a checkpoint, not a cosmetic detail.

Deep DiveWhere your data lives and what that means for access - our page on data location→

Status and sources. Verified in July 2026: the Swiss data protection act (articles 5, 9, 16, 22) and the federal commissioner's pages on AI and data protection, cloud services and data processing by employers. This is practical orientation, not legal advice - when in doubt, the question belongs with a specialist.

Bringing the team along

The starting point: in most companies AI is already in use, just not officially. The Swiss federal statistics office found in spring 2025 that a good four in ten people in Switzerland have used generative AI, almost a third of the population for work. The figure measures people, not companies - but it supports a realistic assumption: someone at your company is already trying it. The only question is whether with or without rules.

Bans change none of that; they only push it into the shadows - onto the private phone, into the private account, where none of your rules apply. Rules bring it out. One page is enough to start, and it answers four things: which tools are allowed, what may go in and what never does, that a person reviews every result before it leaves the company, and who to ask. The page stays deliberately short and grows with what you learn.

And then: time. "Try it on the side" is the quietest way to bury an initiative. Whoever runs the pilot gets hours for it, not encouragement. Take the curious ones first - almost every team has one person who secretly knows this already; make them the go-to instead of the exception. And when the pilot is over, that group tells the rest what worked and what did not. That carries more weight than any outside training.

Choosing tools and what they cost

One tool the company truly masters beats five subscriptions nobody quite knows. The vendor zoo is the most common trap after a good start: for every new problem a new tool promises the answer, and after a year six subscriptions are running and two are used.

The choice itself is less dramatic than the comparison charts suggest. Four criteria are enough. First, the contract: everything from chapter 4 - without it a candidate is out, however good it is. Second, administration: set up accounts for the team centrally and revoke access when someone leaves. Third, its place in the day: a tool that sits where you already work, in the writing program or the inbox, gets used; one that needs its own tab gets forgotten. Fourth, language: the draft has to arrive in your German, not in translation German.

The costs come in three blocks, and they are not the same size. Licences are the smallest: a business plan costs per person roughly what a mobile plan does. The biggest block is time - the hours of the person running the pilot, the rounds in which the team learns, the cases that get done twice at the beginning. Budget only the licences and you have budgeted the smallest part. And the forgotten block is upkeep: the policy needs to stay current, the tools keep changing, and someone has to answer the questions that start coming. One hour a week for that role is more realistic than none.

The honest limits

What AI still cannot do in 2026: carry responsibility, stay reliable on facts, know your business without being taught it. The first is the most important limit and the most lasting - whatever leaves the company has a person as its sender, and that person stands behind it. The second means, in daily work: an AI text that sounds convincing can still be wrong, and it looks confusingly like the right one. Reviewing is therefore not a transitional caution you drop later; it is part of the procedure. And the third disappoints most reliably: without your templates, your prices, your tone, every tool writes average. It only gets good once you give it what "good" means at your company - and nobody takes that work off your hands.

How you notice the introduction is working is unspectacular: the team uses the tool without being reminded. The measuring line from the pilot still holds in month three. And the questions inside the company get more concrete - no longer "are we allowed to?", but "can we do the enquiries with it too?".

And when do you leave it? When no case at your company passes the three questions from chapter 2 - that happens, and it is no disgrace. When the data side is unresolved and nobody has time to resolve it. Or when the only reason is that everyone else is doing it. Introducing AI because everyone does is the same mistake as rebuilding a website because it looks old: the case is missing. It may come next year - and then this guide will still be here.

Questions from practice

What business owners ask us about bringing in AI - answered the way we answer them.

Can we put customer data into an AI tool?

That depends on the tool, not on the AI. A business account whose contract governs the data processing and rules out use for training can handle customer data in principle. A private free account cannot. And especially sensitive data, health information for instance, needs its own review before any tool ever sees it.

Do we need a written AI policy?

Yes, and one page is enough to start: which tools are allowed, what may go in and what never does, and that a person checks every result before it leaves the company. That page is a start. It grows with what you learn along the way.

Will AI replace jobs here?

In small businesses it shifts work rather than cutting jobs: less typing of first drafts, more reviewing and deciding. Anyone promising you blanket staff savings is doing the maths on your expectations, not on your business.

Is the free version enough to get started?

For private finger exercises, yes. For the business, no. The difference sits in the contract: business plans typically commit to not using your inputs to train the models - exactly the point free accounts leave open or bury in a setting. The difference is one subscription; the awkward questions afterwards cost more.

What does the data protection act actually require?

Switzerland's revised data protection act has applied since September 2023 and makes no exception for AI: you remain responsible for the personal data you process. In practice that means binding the provider by contract, knowing where the data flows, and running a data protection impact assessment before any high-risk processing. Chapter 4 walks through it.

What should we do next week?

One single case. Take a task that repeats every week, consists of text and gets checked by a person, and write it down in two sentences: we are trying X, Y owns it, every result gets reviewed. That is the entire starting effort - the pilot in chapter 3 takes it from there.