The law already applies. Since before AI.

“Understanding AI” series, note 9 of 12. The previous eight explained the technology and how it is brought into the company. This one explains what the law says, with a warning first: I am not a lawyer, this note is not legal advice, and what it does is put in order what your adviser will confirm. Read it to arrive at that meeting with the questions already written.

Monday, 11:20 am, an estate agency in Marbella. A sales agent has to prepare a deposit contract for a buyer from Zurich and, to go faster, pastes into a free account with an AI assistant the buyer’s scanned passport, their account number and another client’s previous contract as a template. In forty seconds she has an impeccable draft. Three weeks later, in a data protection audit requested by the Swiss group that has just bought the agency, someone asks where the data from that passport is. Nobody knows. The assistant was free, and in the free accounts of some providers what is written can be used to train the model.

The scene is imaginary. The audit’s question is not, and it is the first one you have to be able to answer. In this note I give you the three layers of law that affect a company using AI, in Switzerland and in Spain, with what each requires and the dates that really concern you. You will see that most of it is not new: it is the data protection law of always, applied to a new tool. And you will see where there are new things, with dates.

AI did not create new obligations about your customers’ data. It created a very fast way of breaking the ones you already had.

Layer 1. Data protection, already in force.

In Spain the European Union’s General Data Protection Regulation, the GDPR, has applied since May 2018. In Switzerland the revised Federal Act on Data Protection has been in force since 1 September 2023, and it was drafted to stay close to the GDPR. Neither mentions language models. And both apply to them in full. The Swiss Federal Data Protection and Information Commissioner said so in writing in May 2025: the current law is directly applicable to data processing with AI, regardless of any future regulation.

What those laws require, translated to an AI system, fits in six points. I list them with the obligation and with the estate agency example.

• Purpose and legal basis. Each item of personal data is processed for a specific purpose and with a reason the law admits. The buyer’s data was collected to sign a sale, not to feed an external system for unknown purposes. Pasting it into a free account is a new processing with no basis.
• Transparency. People have the right to know what is done with their data and, according to the Swiss Commissioner, to know that they are talking to a machine and whether what they write can be used to improve the system. An assistant on your website that passes itself off as a person breaks this in Switzerland today, and in the European Union from August 2026 by a second route, as you will see in layer 2.
• Automated decisions. If a decision with legal or significant effects on a person is taken by a machine alone, that person has the right to have someone review it. The GDPR has it in Article 22 and the Swiss law in a rule of its own. A system that rejects a rental application, a loan or a job application on its own falls squarely here.
• Impact assessment. When a processing involves a high risk for people, it has to be analysed in writing before starting; the GDPR calls it an impact assessment, Article 35, and the Swiss law requires the same for high-risk AI. Health data, minors, profiles, surveillance: before switching on, the paper.
• Processors and contracts. The model provider processes data on behalf of your company, so it is a processor and a contract is needed that says what it does with the data, where, for how long and how it deletes it. Article 28 of the GDPR spells it out. A free account has no such contract.
• Transfers outside Europe and Switzerland. Most of the large model providers are in the United States. Sending data there requires a recognised safeguard, which serious providers offer in their business terms and which has to be checked. Ask where your data is processed and under which safeguard. If the provider cannot answer, that is its answer.

There is one point worth separating clearly: a consumer account is not the same as a business account. Anthropic, for example, documents that by default it does not use what companies send through its API or to its commercial products to train its models, unless the customer authorises it or sends feedback, and that personal accounts are governed by other terms. Other providers have similar distinctions. The practical rule is simple: customer data only in business accounts with a contract. Everything else is the estate agency in Marbella.

Layer 2. The EU AI Act, with its dates.

Regulation (EU) 2024/1689, known as the AI Act, entered into force on 1 August 2024 and applies in phases. In July 2026 a second regulation, the so-called Digital Omnibus, postponed some of those phases. I give you the calendar as it now stands, with an important clarification: it applies to whoever uses or places AI systems on the market in the European Union, whether based inside or outside, so a Swiss company that serves customers in Spain or Germany with an AI system is inside.

• Since 2 February 2025: the prohibited practices, such as social scoring or subliminal manipulation, and Article 4, the obligation for the people who use AI systems on behalf of the company to have a sufficient level of AI literacy. It is already in force. It is the training from step 8 of note 6.
• Since 2 August 2025: the obligations of the providers of the large general-purpose models. They affect you indirectly: they are what allow you to demand documentation from your provider.
• From 2 August 2026: Article 50, transparency. An AI system that talks to people must be designed so that they know they are talking to a machine, unless it is obvious. Images, audio and video generated with AI that could pass for real must be marked. This is new and it does concern you: the assistant on your website, the voice that answers the phone, the generated photo of your hotel.
• From 2 December 2027 (previously August 2026): the high-risk systems of Annex III, such as those used in recruitment, in creditworthiness assessment or in access to essential services, with their obligations of human oversight, logging and documentation. The Omnibus postponed it; it did not remove it. If your company uses AI to screen applications, you have until then to do it properly.
• From 2 August 2028: the high-risk systems built into regulated products under Annex I, such as medical devices or machinery.

Notice the detail most often overlooked, which I already dealt with in an earlier note on this calendar: the law distinguishes between the provider, who develops the system, and the deployer, who uses it. Almost all companies see themselves as the latter. But if a company takes a model, configures it, puts its brand on it and offers it to its customers as its own, it may be turning into a provider, with greater obligations. It is a question for your adviser, and it is better asked before putting the logo on the assistant.

Layer 3. Switzerland: no AI law, and a decision taken.

Switzerland has no specific artificial intelligence law today, and has decided not to copy the European one. On 12 February 2025 the Federal Council opted for a sector-based approach: ratifying the Council of Europe’s convention on AI, which it signed in March 2025, and adapting existing laws where needed, with a preliminary draft for consultation planned before the end of 2026, focused on transparency, data protection, non-discrimination and oversight. Meanwhile, what applies is layer 1: the data protection law, which the Federal Commissioner considers sufficient for the essentials, and the rules of each sector, such as those of the financial market or of healthcare.

For a Swiss company this has two practical consequences. The first: do not wait for a Swiss AI law to put your data in order, because the law that binds you already exists. The second: if you sell in the European Union, the calendar of layer 2 is yours, even if your head office is in Lugano or Zurich. Both are solved by the same work.

What regulators look at, in one case.

On 18 December 2024 the European Data Protection Board, which brings together the authorities of all the Union’s countries, published an opinion on AI models at the request of the Irish authority. Three ideas from that text are useful to you even if you will never train a model. First: a model trained on personal data is not anonymous by virtue of being a model; it has to be shown, case by case, that nobody can be identified and that their data cannot be extracted with questions. Second: legitimate interest can serve as a legal basis for developing and using models, but only after a three-step test and taking people’s reasonable expectations into account. Third: if a model was built with unlawfully processed data, that can affect the lawfulness of using it. Translated: the provenance of your provider’s data is also your problem, and it is legitimate to ask about it.

What to do before calling your lawyer.

None of these things requires legal knowledge, and all of them make the consultation with your adviser shorter and cheaper.

• The inventory. Which AI systems the company uses, including those people use on their own, with which data and from which provider. Without an inventory there is no compliance possible, only luck.
• The contracts. For each provider, the processing contract: where it processes, whether it trains on your data, how long it keeps it, how it deletes it, which safeguard it uses to take data out of Europe or Switzerland. If it does not exist, a business account with a contract or nothing.
• The two-line rule. Which data is never pasted into any assistant other than the company’s: customer data, employee data, health, financial, third-party contracts. Written, signed, repeated.
• The machine says it is a machine. In every assistant that talks to customers, from the first sentence. In Switzerland it is already an expectation of the Commissioner; in the European Union it is an obligation from August 2026.
• Nobody decides alone. Every decision with effects on a person, a rejection, a rate, a selection, goes through someone with a name who can change it. It is the law and it is note 2.
• The paper before the risk. Before using AI with health data, with minors, with profiles or with job applications, a written impact assessment. And for job applications, the December 2027 calendar already in the diary.
• Images, marked. From August 2026, anything generated with AI that could pass for real is flagged. A one-line policy for marketing solves 90% of cases.
• Training, recorded. Who has received which training on the systems they use. Article 4 does not ask for a certificate; it asks to be able to show it was done.

1. This week

   Do the inventory with an amnesty

   Ask each team which assistants it uses and with which data, with no consequences. Note provider and type of account.

2. This week

   Close the free accounts holding customer data

   And replace them with business accounts with a contract. It is the only urgent action in this note.

3. This month

   Ask for the contracts and the three answers

   Where they process, whether they train, how they delete. Keep the answers with a date.

4. Before August 2026

   Review the assistant on your website and the generated images

   That it says it is a machine and that generated material is marked. Two small changes with a fixed date.

5. With all of the above

   Go to your adviser with the inventory in hand

   The consultation goes from “what do we have to do about AI?” to “is this right?”. It is a different invoice.

How I approach it.

The systems I build use models through business accounts with a contract, customer data does not leave the agreed flow, the assistant says it is a machine in its first sentence and in all six languages, and generated images are declared in the caption. The inventory and the two-line rule are part of the delivery. And the legal review is done by your lawyer, with my documentation in front of them, because each of us does our own job.

A test with no lawyer. Tomorrow, ask three people on your team, from different departments, whether they have ever pasted a client’s document into an AI assistant. Ask without reproach in your voice, or they will not tell you the truth.

The answers will tell you more about your real situation than any reading of the regulation. And they open a less comfortable question: did anyone in your business decide this was allowed, or did nobody simply say no?

Next note: security. The ten risks that an organisation of security experts has catalogued for systems with language models, explained with a letter from a fraudster, and the three public cases that teach more than any manual.