ChatGPT at work and the Swiss nFADP: what you can do and what to avoid
Your team is already using ChatGPT, with or without permission. The right question is not "if" but "how": what you can put into an AI tool and what you can't, under the nFADP (the Swiss data protection act).
Your team is already using ChatGPT. With your permission or without it. I see it in almost every SME I work with in Ticino: someone rewrites their emails, someone summarises documents, someone pastes in "just a small piece" of a contract. So the right question is not whether to use AI in your company, but how: what you can put into an AI tool and what you can't, under the nFADP, Switzerland's data protection law.
In this article I try to answer as concretely as possible, with examples from SMEs and a policy you can adopt tomorrow. If you need the bigger picture of the law, I have written a practical guide to the nFADP for SMEs: here I focus on the specific case of ChatGPT and AI tools. One necessary disclaimer: this is not legal advice, it is the experience of someone who helps SMEs use these tools every day.
The starting point: the data you enter leaves the company
When you paste a text into ChatGPT, that text is processed on the provider's servers, outside your perimeter. And in the consumer versions, the free or personal ones, the data you enter can be used to train the models, unless you explicitly turn this off in the settings. Business and enterprise versions, and API access, generally work the other way round: the provider commits to not using your data for training. But it needs to be verified tool by tool, because the terms change.
For the nFADP this point is decisive. Using an AI tool means processing data: if the prompt contains personal data of customers or employees, you remain the data controller, and the responsibility does not transfer to the tool. The nFADP does not ban AI: it asks you to know what you enter, where it ends up and with what safeguards.
What you should NEVER enter
Four categories that should never go into the consumer version of an AI tool.
- Personal data of customers and employees. Names, emails, phone numbers, addresses. The classic example: pasting the customer list from your ERP to "tidy it up", or a customer's complaint email, full name included, to have the reply written for you.
- Data requiring special protection. Health, religious or political views, candidate data and CVs. An employee's medical certificate in ChatGPT is probably the most serious breach an SME can commit with AI.
- Trade secrets. Confidential price lists, recipes and production processes, commercial strategies, proprietary code that contains your business logic.
- Financial data tied to names. Payslips, invoices with names on them, the debt situation of specific customers, confidential contract terms.
The practical rule I use in my training sessions: if you wouldn't email it to a stranger, don't paste it into a consumer AI tool.
What you can safely do
The list of what is safe is much longer than the list of prohibitions, and that is where the value lies.
- Generic texts and drafts. Product descriptions, posts for your website or LinkedIn, internal communications without names, outlines for meetings and presentations.
- Rewording and translations of non-confidential texts: making a standard offer clearer, translating a page of your website, adjusting the tone of a reply (without the recipient's name).
- Code and formulas without data. An Excel formula, a script, a query: as long as they contain no real data or credentials, no problem.
- Analysis on anonymised or made-up data. "I have 200 customers split into these three tiers, how would I segment them?" works perfectly well without a single real name.
In practice: almost all writing, summarising and reasoning work can be done safely, if you strip identifiable people and confidential data out of your prompts.
The right settings and versions
Beyond the "what", the "where" matters too. Three concrete moves, in order of effort.
- Turn off training wherever possible. The consumer versions of many tools have a setting to exclude your conversations from model training. It is the bare minimum, to be done right away on every account used for work.
- Move to a business version. ChatGPT Team or Enterprise provide that your data is not used for training and offer admin controls. This is not a detail: it changes the contractual framework with the provider.
- Consider alternatives within your perimeter. If you work in Microsoft 365, Copilot used within the company tenant keeps the data inside the existing Microsoft perimeter, with the permissions you have already configured.
An honest caveat: no version makes you compliant "automatically". Compliance depends on what you enter, on how you configure the tool and on how you regulate the relationship with the provider when you process personal data. This matters all the more once AI stops being an occasional assistant and becomes part of your processes, for example in an AI automation project: there, the choice of tool and data perimeter has to be made upfront, not after the fact.
A simple 5-point AI policy
Without written rules, everyone decides for themselves, and you end up with Shadow AI: unauthorised tools used in secret, with data leaving the company without your knowledge. The solution is not a thirty-page rulebook. One page is enough, with five points.
- 1. Approved tools. Which AI tools can be used for work and with which accounts. Everything else is out.
- 2. What never to enter. The four categories above, with two or three concrete examples taken from your own company.
- 3. Who to ask. One point of contact for doubts. When in doubt, ask before, not after.
- 4. Training. At least one hands-on session for the whole team: what is safe, what is not and why. It is the point that makes the other four work, and it is the heart of my AI training for SMEs.
- 5. Periodic review. Tools and terms of use change fast: reread the policy every six months and update the list of approved tools.
Don't ban: train
After seeing dozens of different situations, my conclusion is always the same: the best way to manage the risk is not to ban, it is to train. A ban pushes usage into the shadows and makes you lose the benefits; a trained team knows what it can enter, uses the right versions and turns AI into an advantage instead of a risk. That is exactly what I do with my AI training for companies and SMEs: half a day on your concrete case, with your tools and your rules. If you want to talk it through, write to me.
This article is for information purposes and does not constitute legal advice. For specific cases, consult a legal advisor or the competent authority (the FDPIC).
Want to use ChatGPT at work without unnecessary risks? I help you choose the right tools and versions, write the policy and train your team, starting from your real processes.
Frequently asked questions
Is ChatGPT banned under the nFADP?
No. The nFADP does not ban ChatGPT or other AI tools: it regulates the processing of personal data. It depends on how you use it: if you don't enter personal or confidential data, or if you do so with adequate safeguards, you can use it. The problem arises when customer or employee data ends up in a tool without any control.
Can I enter customer data if it is anonymised?
Yes, if it is truly anonymised: the person must not be identifiable even by cross-referencing the remaining information. Removing just the name is often not enough, because role, company and context can re-identify someone. When in doubt, replace the data with generic placeholders before pasting.
Is Copilot safer than ChatGPT for a company?
It is not a binary comparison: it depends on the configuration. Copilot used within the company's Microsoft 365 tenant, with the data staying inside that perimeter and no training on your data, offers guarantees that the free version of ChatGPT does not. But ChatGPT in its Team or Enterprise version also changes the picture considerably. What matters is how the tool is configured, not the logo.
Does a micro-business also need a written policy?
Yes. One page is enough: which tools are approved, what must never be entered, who to ask when in doubt. Even with three people, without a written rule everyone decides for themselves, and the risk is the same as in a large company.