Privacy, AI and the rules that apply

A Swiss company that uses AI works under its own data protection law from the first prompt that contains a name. If it serves people in the EU, the GDPR and the EU AI Act come into view too. This year several of those rules moved.

The bigger risk is often habit, not law. IBM’s 2025 study of data breaches found that one in five organisations had a breach linked to shadow AI, the tools staff use without approval, and that a high level of it added an average of 670'000 dollars to the cost of a breach. Of the organisations with a breach of an AI model or application, 97 percent lacked proper access controls for AI.

Here is where the rules stand in October 2026, and what a careful company does. It is an overview, not legal advice.

Switzerland: the law already applies

Switzerland has no AI law yet, but AI is not unregulated. The revised Federal Act on Data Protection has applied since 1 September 2023, and the Federal Data Protection and Information Commissioner, the FDPIC, has said since November 2023, and again in May 2025, that it applies directly to AI supported processing of personal data.

The FDPIC expects transparency about the purpose, the functioning and the data sources of AI processing. People have a right to know whether they are dealing with a machine, and whether what they enter is used to improve the system. High risk processing needs a data protection impact assessment, and a manipulated face, image or voice of an identifiable person must always be clearly marked. The commissioner names comprehensive facial recognition in real time and social scoring as prohibited.

It applies this to the large platforms too. In March 2025 it closed a preliminary investigation into the use of posts on X to train Grok, and pointed out that users can object.

A dedicated framework is coming. The Federal Department of Justice and Police is drafting a bill for consultation by the end of 2026 to implement the Council of Europe’s convention on AI, focused on transparency, data protection, non discrimination and supervision.

The EU AI Act, after the omnibus

The EU AI Act entered into force on 1 August 2024 and applies in stages. Prohibited practices and the duty of AI literacy have applied since 2 February 2025, the obligations for general purpose models since 2 August 2025. Most of the act applies from 2 August 2026, including the transparency rules: people must be told when they talk to a chatbot, and deepfakes must be labelled.

The digital omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026. It moves the rules for high risk systems: those used in areas such as biometrics, education and employment face their obligations from 2 December 2027, and those built into regulated products, such as lifts or toys, from 2 August 2028. It also prohibits AI that generates sexual images of people without consent or child sexual abuse material from December 2026, and extends simpler documentation to small mid cap companies.

For most companies that use AI rather than build it, the delay changes little. The transparency and literacy duties apply now. The extra time is for systems that decide about people, such as who gets hired.

The GDPR, and a reform that is not law

The GDPR itself is unchanged. In December 2024 the European Data Protection Board set out how it reads it for AI models. A model trained on personal data counts as anonymous only if identifying people from it, directly or through questions put to the model, is very unlikely, judged case by case. Legitimate interest works as a legal basis only where the processing is strictly necessary and the balance of rights holds. And a model built on unlawfully processed data can affect whether its later use is lawful, unless it has been properly anonymised.

A separate omnibus would ease parts of the GDPR and the Data Act. On 7 October 2026 the ambassadors of the member states postponed their vote on the Council’s position after Germany and France asked for more time, and at their meeting on 11 October the text was neither discussed nor voted on. Berlin wants a category of low risk controllers that would free small firms and non profit organisations from many duties. None of it applies yet, and we plan with the GDPR as it is written today.

Contracts and transfers

Most AI at work runs through a provider that processes personal data on the company’s behalf, and the contract is where the company’s rules live. It should say what the provider may do with prompts and files, how long it keeps them and its logs, which subprocessors it uses and where, and what happens to the data at the end. If a provider says it does not train on your data, find that sentence in the contract you signed.

Data that leaves Switzerland for the United States needs a legal basis. Since 15 September 2024, the Swiss-US Data Privacy Framework allows transfers to certified US companies without further guarantees, so check each provider and each subprocessor for the certification. The EU’s equivalent survived its first challenge at the EU General Court on 3 September 2025. The appeal, filed on 31 October 2025, has as far as we know not been decided. Know which of your tools depend on a transfer to the US, and what you would do instead.

What a careful company does now

None of this needs a large programme. It needs a list, a few decisions and someone who owns them.

Make the list: every AI tool in use, the data that goes into it, the provider, the country and the contract. Ask teams what they actually use, not what was approved.

Sort the data into public, internal and confidential, and say on one page which class may go into which tool.

Offer something good. Shadow AI grows where the approved tool is worse than the free one. An approved assistant that works removes most of the reason to paste client data into a private account.

Control access. AI tools get the same roles, sign in and offboarding as every other system. An assistant that can read everything will one day show it to the wrong person.

Tell people and assess the risky cases. Say on the website when a customer is talking to a machine, and do the impact assessment before AI touches sensitive data or decides about people.

Give it an owner. In Cisco’s 2026 study, three in four organisations had a body for AI governance, and only 12 percent called theirs mature. And keep a person on the send button: what may be done with someone’s data is a judgement, and judgement stays with people.

How we build for it

Our products follow the same rules. Kontor writes offers and invoices in bexio and stores nothing itself: the data stays in bexio, and sending, payments and reminders stay with a person.

Twin gives each role only the part of the company brain it needs, and every write is signed in per user. It runs in Switzerland or the EU, or on the client’s own hardware, and the data never trains a public model.

When the data must not leave the company at all, Eigna, built by our partner Eigna Labs, runs as a private model in the company’s own cloud or on its own servers, designed for the GDPR and the EU AI Act.

The rules will keep moving. A company that knows where its data goes can follow them.