The question comes up in almost every management meeting the moment someone mentions ChatGPT: “Are we even allowed to do this?” The honest answer is: yes, in principle. But it depends on which version you use, what you enter and whether you have done a few basic pieces of homework. AI tools are not prohibited in Switzerland, quite the opposite. But they are not a law-free zone either. This article sets out what the revised Data Protection Act actually requires, what the supervisory authority says about it and what you, as an owner or managing director, need to keep in mind in practical terms.
The revised FADP applies, to AI as well
The revised Data Protection Act (revFADP) has been in force since 1 September 2023. There was no transition period: anyone processing personal data in Switzerland had to be compliant from that date. This affects practically every business, because customer addresses, employee files, email correspondence and accounting data are all personal data.
What is decisive for the AI question: the Federal Data Protection and Information Commissioner (FDPIC) has made clear that the current data protection legislation is directly applicable to AI. No special “AI law” is needed for rules to apply. As soon as you enter personal data into an AI tool, it is data processing like any other, with the same obligations regarding transparency, data security and proportionality. The FDPIC also stresses that users have a right to know whether they are communicating with a machine and whether their inputs will be reused.
In plain terms: the technology is new, the ground rules are not. What applied to a cloud service or an external service provider also applies to the AI chatbot.
The costliest mistake: typing data into the wrong version
This is the most important and most frequently overlooked point. There is a tangible difference between the free/personal versions and the Business plans, and it determines whether your data is used to train the model.
- Free and personal versions (ChatGPT Free, Plus; Claude Free, Pro): here your inputs are used by OpenAI for training by default, unless you actively object (opt out in the settings). This means: if an employee types a customer list or a personnel case into the free version, this data can feed into model improvement.
- Business, Team, Enterprise and API plans: OpenAI states that it does not use data from ChatGPT Business, Enterprise, Edu and the API for training by default. Anthropic gives the same assurance for its commercial products (Claude Team, Enterprise, API).
Exactly the same dividing line runs through the data processing agreement (see the next section). Anyone using AI in their business should therefore enforce a simple rule: personal data belongs only in a contractually secured Business version, never in the free one.
The data processing agreement (DPA) is only available in the Business plan
If you have personal data processed by an external service provider, the revised FADP (Art. 9) requires you to ensure that this provider treats the data in just as legally compliant a manner as you would have to. This is usually governed by a data processing agreement, known internationally as a Data Processing Addendum (DPA).
Both major providers make such a DPA available, but only in the paid plans:
| Provider | DPA available in | No DPA in |
|---|---|---|
| OpenAI (ChatGPT) | Business, Enterprise, Edu, API (automatically part of the contract) | Free, Plus |
| Anthropic (Claude) | Team, Enterprise, API (via the Commercial Terms) | Free, Pro |
With both providers, the DPA includes Standard Contractual Clauses (SCCs). In addition, OpenAI now offers options for data residency in Europe (for Enterprise, Edu and the API) as well as zero-data-retention agreements; with Anthropic, a zero-data-retention option is available for Enterprise. These points need to be checked on a case-by-case basis; they are not automatically active in every plan.
A word on data transfer to the United States, which used to be the great bugbear: since 15 September 2024, Switzerland has recognised an adequate level of data protection under the Swiss-US Data Privacy Framework for US companies certified under this framework. For such certified providers, data transfers to the United States are permissible without additional safeguards. Whether a specific provider is certified should be checked in the official list at dataprivacyframework.gov, rather than relying on marketing statements.
What fines cost and who they target
The revised FADP has significantly raised the range of fines. Two particular features are important for management:
- Intentional breaches of certain obligations (for example duties to inform, to provide information, or duties of care when transferring data abroad) can be punished with a fine of up to CHF 250,000.
- Unlike under the EU GDPR, in Switzerland the fine is primarily directed at the responsible natural person, not at the company. This could be the managing director or the person responsible for data protection. If no individual can be identified, the company itself may be fined, but in that case to a maximum of CHF 50,000.
This is an essential point: personal liability cannot simply be shifted onto the firm. Anyone who is accountable for data protection within the business has a genuine self-interest in ensuring that the use of AI is properly regulated. At the same time: this concerns intentional breaches of clearly defined obligations, not every minor oversight. Those who observe the basics need not panic.
Do you need a data protection impact assessment?
The revised FADP (Art. 22) requires a data protection impact assessment where a planned processing operation is likely to result in a high risk to the personality or the fundamental rights of the data subjects. The use of new technologies, expressly including algorithmic systems in the sense of “artificial intelligence”, is cited as a typical trigger for such a high risk.
A high risk exists in particular where you process sensitive personal data on a large scale (for example health, religion or trade union data), where automated individual decisions are made, or where public areas are systematically monitored. An SME that uses AI only to draft texts or to summarise internal documents without sensitive personal data will, as a rule, not need an impact assessment. By contrast, anyone using AI to assess job applicants, prepare credit decisions or process health data should look closely.
The 7-point checklist for using AI in an SME
These seven points sum up the essentials and can be ticked off in a single meeting:
- Business plan instead of the free version. As soon as personal data is involved, use only contractually secured versions. No sensitive data in free or personal accounts.
- Conclude a DPA. Check whether a data processing agreement is in place for your plan, and activate or sign it.
- Switch off training. Make sure your inputs are not used for model training (usually the default in Business plans; in personal versions, actively disable it).
- Set internal ground rules. A short, clear policy: which data may be entered, and which may not? Who is allowed to use which tools?
- Maintain transparency. Data subjects have a right to know when their data is processed with AI support or when they are communicating with a machine.
- Assess the risk. Where sensitive data or automated decisions are involved, check whether a data protection impact assessment is needed.
- Train your staff. The most common data protection mistake is not the technology, but the careless input. A brief awareness session prevents most problems.
At Vollmer Labs, we build AI solutions that take these points into account from the outset, because we have to run them in production ourselves. Our AI agents and automations are today running, among other things, as accounting agents in a fiduciary firm, where handling sensitive financial data is part of the daily routine. These building blocks are in use, but not yet proven in many industries, and we say so plainly. Anyone wanting to work with AI in the Swiss context cannot avoid the basics described here, regardless of which solution is ultimately chosen. We show what this looks like in concrete terms in regulated industries using the example of fiduciary services.