You hear about it everywhere. Depending on who is talking, artificial intelligence is about to replace everyone or is good for nothing. You run a workshop, a practice or a shop, and you simply want to know whether there is anything to do on Monday morning.
There is something to do, and it is modest. Here is what the official sources say in mid-November 2026, without predictions.
ChatGPT, Gemini, Copilot: what they do well, and what they do badly
ChatGPT, Gemini, Copilot and Claude are assistants you write to as you would to a person. The Federal Office for Cyber Security (FOCS) sums up what they can do in 2026: produce text, images and code, and translate almost in real time. For a small business, that means a first draft of an email, a summary of a long set of regulations, a translation that still needs checking.
They are already in the pockets of your customers and your staff. In spring 2025, according to the Federal Statistical Office, 43% of the population aged 15 to 88 (in French; also available in German and Italian) said they had used a generative AI application, and 31% used one for work.
Their weakness lies in how they work. As FOCS points out, a language model produces text from statistical relationships, and the result may be right or wrong. Google says so itself in the Gemini help pages: such tools can "present inaccurate information as factual". The wrong answer arrives in the same confident tone as the right one, and nothing flags a legal article that does not exist or a deadline that was made up.
The test. Ask an assistant three questions from your own trade to which you know the answer perfectly: a standard, a deadline, a common practice. Count the mistakes. You will know how much credit to give it on subjects you know less well.
Will AI replace jobs? Exposed does not mean eliminated
In May 2025 the International Labour Organization (ILO) published a study carried out with NASK, Poland's national research institute. It estimates that one job in four worldwide is in an occupation exposed, to some degree, to generative AI. In high-income countries such as Switzerland, the share rises to 34%. Clerical jobs remain the most exposed.
At the other end of the scale, the ILO working paper observes that exposure falls when tasks are physical. For craft trades, machine operators and gardeners, the experts consulted gave lower scores than the workers surveyed: automation runs into limits as soon as a task requires handling an object or manual dexterity.
The ILO stresses that its figures measure potential exposure, not job losses, and that most jobs will be transformed rather than eliminated, because human input is still needed. Its researchers also revised some of the 2023 scores downwards, judging them too optimistic after two years of experimentation. Others went up, in web development, media, finance and IT.
In a small business, then, what is exposed is mainly the office: quotes, payment reminders, minutes, data entry. Work on site, in the workshop or in the garden, which is done by hand, is far less so.
First risk: what you paste into ChatGPT or Copilot
The Federal Data Protection and Information Commissioner (FDPIC) has been making the point since November 2023: the data protection law in force since 1 September 2023 is directly applicable to AI-based data processing.
In October 2026 the FDPIC and FOCS ran a joint awareness campaign on the risks of generative AI at work, in particular what FOCS calls oversharing of data. Copying a whole email into an assistant is enough to hand over a customer's name, their situation, sometimes their health. And health data are sensitive personal data within the meaning of the law (Art. 5(c) revFADP), subject to stricter requirements. For a medical or therapy practice, further rules specific to the health sector also apply.
FOCS adds that many tools may use what you type to train future models, unless this is excluded by contract, and that information once entered cannot be retrieved. Google states that a subset of Gemini chats is checked by human reviewers.
FOCS recommends anonymising before you type, choosing the business version of a tool over a private account, and refusing training on your data wherever possible.
The test. Open your assistant's history, and have your staff do the same. Look for customer names, diagnoses, salaries, bank details. Then open the settings and look for the option that concerns model training.
Second risk: better-written scams, thanks to AI
The FOCS semi-annual report, published in August 2026, counts fraud attempts as 39% of the reports received in the first half of the year. The office notes that AI now lets scammers generate emails, text messages and even websites in an instant that look more real than the real thing. Fake invoices remain one of the most common pretexts for sending malware.
The report also describes a CEO fraud in which the criminals imitated a manager's voice to obtain payments. According to FOCS, around ten seconds of recording is now enough to clone a voice. The office says these are still isolated cases, but expects that to change.
Spelling mistakes and clumsy wording are no longer enough to spot a fake email, and a familiar voice no longer proves anything. A simple procedure protects you better than your eyes or ears.
The test. Agree today with the person who makes your payments: no change of bank details and no urgent transfer is carried out on the strength of an email or a call, even one in your own voice. You hang up and call back on the number you know: that is the reflex FOCS recommends.
On the technical side, a strict DMARC record makes fake emails that use your exact address fail at the email providers that apply your policy. We explain how in our article on SPF, DKIM and DMARC.
Being found when customers ask Google or an AI
Since March 2025, Google has shown AI Overviews for some search results in Switzerland, in French, German, Italian and English, initially for signed-in users aged 18 and over. Other customers put their question straight to an assistant.
These answers draw on what has been published. Google's documentation is explicit on several points:
- No special optimisation is needed, and no "AI" file has to be added.
- The page must be indexed and eligible to appear in search results with a snippet.
- Important information must be available as text.
- Your Google Business Profile must be kept up to date.
If your website and your Google profile give different opening hours, the machine chooses for you.
The test. Ask two different assistants: "What do you know about [business name] in [town]? Opening hours, services, phone number." Note every mistake, then work out where it comes from: your website, your Google profile, an old directory.
What the law says about AI, in Switzerland and in the EU
On 12 February 2025 the Federal Council announced that it intends to ratify the Council of Europe Convention on AI and to continue regulating sector by sector, for example in healthcare or transport. OFCOM sets out the approach: legislative amendments will be made on a sector-specific basis wherever possible, and general cross-sector regulation will be limited to areas linked to fundamental rights, such as data protection.
On 28 October 2026 the Federal Council took stock. The consultation on the draft is planned for spring 2027 (in French; also available in German and Italian) and will cover transparency, data protection, non-discrimination and supervision, among other things. For now, then, Switzerland has no general AI law. The data protection law, however, already applies.
According to its Article 2, the EU AI Act covers those who place AI systems on the European market, those who use them within the EU, and companies outside the EU whose AI output is used there. The Confederation's SME Portal adds that the Act's transparency rules have applied since August 2026: labelling deepfakes, informing people who are talking to a chatbot.
A business in Biel/Bienne that writes its quotes with an assistant, for Swiss customers, does not in principle fall within this scope. The question arises if you sell a product containing AI in the EU, or if your website offers a chatbot to customers in France or Germany. In Switzerland too, the FDPIC points out that people have the right to know whether they are writing to a machine.
Three steps for this week
- A written rule on data. One page, applying to everyone, you included: what never goes into an AI tool. Customer names together with their situation, health data, salaries and staff files, bank details, passwords, documents a customer has entrusted to you. FOCS recommends that businesses set a rule of this kind.
- Drafts, never decisions. The assistant prepares, someone who knows the trade checks. Every figure, legal article or deadline is verified at the source before anything is sent.
- Your portrait as a machine sees it. Run the test above once, then correct your website or your Google profile wherever the answer is wrong.
AI is moving fast, and some of the official pages cited here will have changed within a year. The three steps above do not depend on any of them.