AI training for companies: what to teach your employees.

In Switzerland, more than half of people who use AI at work have already used it against company rules. Banning it doesn’t help. Good AI training for companies fits into half a day of practice and one page of rules.

By Alessandro ScerpellaManaging Director 9 September 2026

74%

of people who use AI at work in Switzerland have relied on a result without checking it.

KPMG / University of Melbourne, 2025

At a glance

Among people who use AI at work in Switzerland, 52% have used it against company rules and 74% have relied on a result without checking it.
Banning it isn’t enough. What works is a few written rules and half a day of practice on the tools you already use.
Personal and business accounts aren’t the same thing: the contract is different, and so is what happens to your data.
If AI output ends up in the EU, the AI Act requires measures to train the people who use it. Since July 2026, in a lighter form.

Tuesday morning at a medical practice with four staff. The medical secretary needs to summarise a three-page report for a letter. She opens ChatGPT on her phone, using her personal account, and pastes in everything: the patient’s name, date of birth, diagnosis. Two minutes later the summary is ready, and it’s well written too.

Nobody at the practice knows, and nobody has ever told her it isn’t allowed. The practice is fictional; the figures below show how often AI gets used like this, without rules. That’s what AI training for companies is for: what to teach, in what order, and which rules to write down before the course.

Your staff already use AI

According to the Federal Statistical Office, 47% of the Swiss population said they used generative AI in 2025, meaning tools like ChatGPT, Gemini or Copilot that write text and answer questions. In Europe, only Norway and Denmark have higher rates1.

Some of that use finds its way into the office. In a survey by Innofact for Comparis in March 2026, 31.4% of chatbot users said they use them for everyday work, to summarise or write texts. A year earlier the figure was 26.6%2. Among Swiss workers surveyed by KPMG and the University of Melbourne, 65% say they use AI at work regularly and intentionally3.

Even if nobody at your company has decided anything, AI is already there: in the browser, on phones, in private accounts. The question is how it’s being used.

The risk is using AI without rules

KPMG and the University of Melbourne surveyed 48,000 people in 47 countries between November 2024 and January 2025, around a thousand per country. Here’s what the Swiss country profile says about people who use AI at work3:

52%

have used AI in ways that go against company rules or guidelines.

KPMG / University of Melbourne, 2025

63%

have made mistakes in their work because of AI.

KPMG / University of Melbourne, 2025

In the same profile, 74% say they have relied on AI output without checking whether it was correct, and 69% that they have passed off content made with AI as their own. These figures are higher than the global average: worldwide, 44% have used AI against company rules3.

These mistakes aren’t down to the tool. They come from two things nobody has explained: what you can paste in, and what needs checking before you use it.

Banning AI doesn’t get it out of the office.
It just moves it onto personal phones.

The federal government takes the same view. In its 2025/2026 annual report, the FDPIC (Federal Data Protection and Information Commissioner) says it secured agreement that staff implementing the federal administration’s AI strategy will be trained and given access to learning materials4.

In-house AI training: the half-day programme

For most SMEs, a well-run half day is enough, working on the company’s real documents and with the tools you already use. Here’s how we structure it.

How it thinks (30 minutes). A language model doesn’t look for the right answer: it writes the most likely continuation. That’s why it confidently gets names, figures, dates and legal articles wrong. Once people understand this, they stop trusting it blindly.
What not to paste (45 minutes). Health data, salaries, customer data with full names, passwords, documents covered by professional secrecy. People learn to strip out the data first: “Client A” instead of the name, no dates of birth, no IBANs.
Asking well (60 minutes). Context, an example, the format you want back. We work on your own cases: the payment reminder that needs writing in German, the reply to a complaint, the minutes of Monday’s meeting.
Checking (45 minutes). Every number, name and quotation is verified against the source. There’s only one rule: if you can’t check it, you don’t use it.
House rules (30 minutes). We read them together, discuss the grey areas and decide who answers questions after the course.

With breaks, that makes four hours.

One use per department, and what to check
Department	A useful application	What to always check
Administration	A polite draft payment reminder, in German or French	Amounts, due dates, client name
Sales	Replying to an enquiry received by email	Prices and terms: they come from the price list, not from the AI
Site and workshop	A voice message turned into a work report	Hours, materials, site name
Management	Summarising a long contract	The clauses, read in the original text

Our own examples, to give you an idea. The real programme is built around your tools.

If you need more time, we also run courses for companies and professionals, online, in person or tailor-made.

When you don’t need a course

If nobody at your company uses AI and you have no plans to, one line in your internal rules is enough. And if AI only works inside an automation that’s already been built, not everyone needs to learn to write prompts: the person approving the output just needs to know what to look for.

House rules, on one page

Write the rules before the course, not after. One page is enough. Any longer and nobody will read it again.

The rules on one page

What never gets pasted in. Health data, salaries, customer data with names, passwords, documents covered by professional secrecy. If needed, anonymise it first.
Which tool. Only the one the company has chosen, with the company account. Write down the tool’s name.
What gets checked. Facts, names, numbers, dates and legal references, always against the source.
Who signs. Any text made with AI goes out under the name of someone who has read all of it.
With customers. If a customer is talking to an AI assistant, in a chat, on WhatsApp or on the phone, they must be told.
Who answers questions. One person, named in full.

On the customer rule, the FDPIC points out that anyone talking to a language model has the right to know whether they are communicating with a machine5.

The medical practice, with the rules in place

Tuesday’s medical report

The report contains health data, which the revised Swiss Federal Act on Data Protection (FADP, nLPD in Italian) classes as sensitive personal data (Art. 5)6. On top of that, the Swiss Criminal Code makes it an offence, prosecuted on complaint, for doctors and their assistants to disclose secrets confided to them in their professional capacity (Art. 321)7. The medical secretary counts as the doctor’s assistant.

With the one-page rules in place, the same Tuesday goes like this: the report doesn’t go into a personal chat. If the practice wants to use AI for summaries, it chooses a tool with a data protection agreement, decides with the doctor what can go into it and first asks a lawyer whether professional secrecy allows it. The doctor reviews and signs the summary.

Personal or business account: not the same thing

In EY’s May 2026 survey, 29% of respondents said private accounts such as ChatGPT can be used at their company, and for 8% they are the only way to use AI8. EY’s sample, however, includes a lot of large companies.

Personal and business accounts compared
Aspect	Personal account	Business account
Who signs with the provider	The employee	The company
Are chats used to train models?	With ChatGPT they can be, unless you switch the option off	No: ChatGPT Business and Enterprise unless you opt in, Microsoft Copilot with a work account
Data processing agreement	No	Yes: included with Microsoft, signed on request with OpenAI
Who decides who sees what	The employee	The company

Sources: OpenAI910 and Microsoft11, pages accessed on 9 September 2026.

If you stick with a personal account, at least turn off model training: in ChatGPT you’ll find it under Settings, in Data controls. OpenAI warns, however, that if you give a response a thumbs up or down, the whole conversation may be used to train its models, even after you’ve switched the option off9.

For customer data, a personal account isn’t enough. The FADP requires anyone processing data on your behalf to be bound to you by a contract and able to guarantee data security (Art. 9)6. With a personal account, the contract isn’t between the company and the provider. We go into this in our article on ChatGPT and the FADP.

If you work with the EU: what the AI Act requires

The AI Act is the European Union’s regulation on artificial intelligence. It isn’t Swiss law, but it may still apply to you: it also covers businesses outside the EU when the output produced by the AI system is used in the EU (Art. 2)12. An example: an agency in Chiasso that uses AI to prepare texts and offers for Italian clients.

Since 2 February 2025, Article 4 has required providers and users of AI systems to take care of their staff’s AI literacy, meaning the skills to use it sensibly. The original wording required measures to ensure, “to their best extent, a sufficient level” of AI literacy12. Regulation (EU) 2026/1744, in force since 27 July 2026, rewrote the article13.

“Providers and deployers of AI systems shall take measures to support the development of AI literacy of their staff [...]. This obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual.”

Regulation (EU) 2026/1744, new Art. 4 of the AI Act EUR-Lex, 2026

In plain terms: the obligation remains, but it asks for measures, not a certified level. The “deployer” is whoever uses an AI system in a professional capacity, which includes your company. Training built around your tools, with written rules and a list of who attended, is the simplest way to show you’ve given it proper thought. No course, ours included, makes you compliant on its own.

Switzerland doesn’t have an AI law yet. The Federal Council is preparing a draft to put out for consultation by the end of 202614, and in the meantime data protection law applies. This article isn’t legal advice: for your specific case, ask your fiduciary or a lawyer.

If you’d like a course for your team, on site or online, get in touch. We build it around your tools and your documents, starting from the one-page rules.