«InsightsData protection & AI
Privacy-compliant AI in Switzerland: what the nDSG requires
The revised Data Protection Act (revDSG, often called nDSG) has applied since 2023, and many companies hesitate on AI for fear of breaching it. The good news: AI and data protection are not a contradiction. What matters is how data is processed, where the models run and what gets logged.
AI and data protection aren't at odds
Many companies hesitate on AI for fear of breaching the revised Data Protection Act. The concern is understandable, but it often leads to the most expensive reaction of all: doing nothing and leaving the value on the table entirely.
Yet AI and data protection are not a contradiction. What matters is how data is processed, where the models run and what gets logged. Answer those three questions cleanly and you can use AI productively while staying compliant.
Worth knowing: the law does not ban AI. It requires diligence, and diligence can be planned, especially when you build it in from the start instead of retrofitting it at the end.
What the revised FADP actually requires
The revised Federal Act on Data Protection, revFADP or often nDSG, has applied since 1 September 2023 and is closely aligned with the European GDPR. It requires a legal basis for processing, transparency toward the people concerned, data minimization, appropriate security, plus privacy by design and privacy by default.
On top of that come concrete duties: a register of processing activities, in some cases a data protection impact assessment for high-risk processing, and reporting of data breaches to the FDPIC, the Federal Data Protection and Information Commissioner. Individuals have rights to information, correction and deletion.
The framework has teeth: for intentional violations, such as of information or diligence duties, fines of up to CHF 250,000 apply, and against the responsible individual, not just the company. The law is also extraterritorial, covering foreign providers whose processing has an effect in Switzerland.
The critical point: where do the models run?
Many AI services send data to servers in other countries, often to the US. That isn't automatically prohibited, but it needs a clean legal basis. Since 15 September 2024 there is an important easing here: the Federal Council granted the US an adequacy decision through the Swiss-US Data Privacy Framework.
In practice that means: personal data may be transferred to US companies certified under this framework without additional safeguards. For non-certified providers you still need standard contractual clauses and a transfer risk assessment.
The safest path for sensitive data still remains EU or Swiss hosting. Then processing stays within a legal framework whose rules you know, and many tricky transfer questions never even come up.
Pseudonymization as the most effective lever
The single most effective technical lever is to pseudonymize personal data before a model ever sees it. Names, addresses and other identifying details are replaced with placeholders; the model works with the content, not the identity.
That maps directly onto two core requirements of the law: data minimization and appropriate technical security. You get the value of the AI without real personal data flowing needlessly through third-party systems.
After processing, the placeholders can be mapped back in a controlled way, where it is actually needed. That keeps the value while making sure raw data never leaves the company uncontrolled, and data protection turns from a brake into a property of the system.
What it looks like in practice, and what to clarify first
Put together, privacy-compliant AI looks like this: personal data is pseudonymized before processing, models run in EU or Swiss data centres, access is logged, and there are clear rules on which data is kept and for how long. Gardeo, our own AI SaaS, is built on exactly these principles: nDSG- and GDPR-compliant, with automatic pseudonymization, EU hosting and access to more than 20 models.
So we know the questions from regulators and IT security not from slides, but from day-to-day production. What we recommend, we run ourselves, every day, with real users and real data.
Before you start an AI project, clarify three things: which personal data is involved? Where may it be processed? And who has access to the results? A short AI-potential and data-protection audit answers that early, and clarifying data protection early is far cheaper than retrofitting compliance onto a finished solution.
Let's plan your project.
Every successful project starts with clarity. Book a short call: no sales talk, just an honest read on your idea.