Does the nFADP apply if I use Claude, ChatGPT or another LLM via API?

Yes, as soon as personal data enters a prompt — a client's name, an e-mail address, an HR file. The model provider then becomes a processor: you need a data processing agreement, verification of where the processing takes place and, if the data leaves Switzerland, a valid basis for the transfer (an adequate country, Data Privacy Framework certification for the United States, or standard contractual clauses).

Do I need people's consent to process their data with AI?

Not by default. Unlike the GDPR, the nFADP does not require a legal basis for every private processing activity: processing is lawful as long as it does not unlawfully harm the personality of the data subjects. Consent becomes necessary in specific cases — notably the disclosure of sensitive data to third parties, or high-risk profiling when you justify it by consent. The duty to inform, however, applies to all processing: your privacy policy must say what your AI does.

Does an SME have to keep a record of processing activities?

Companies with fewer than 250 employees are exempt, unless they process sensitive data on a large scale or carry out high-risk profiling. Our recommendation: keep it anyway. It is the document that answers a client's, an auditor's or the FDPIC's questions in one page — and drawing it up forces you to do the inventory this checklist demands anyway.

What is the difference between the nFADP and the GDPR?

The nFADP is lighter: no legal-basis requirement for every processing activity, no administrative fines at 4% of turnover, an optional record for most SMEs. But it is criminal where the GDPR is administrative: the fines — up to CHF 250,000 — target the responsible natural persons, for intentional violations of specific duties. And if you serve clients in the EU, the GDPR may apply in parallel.

Who risks the fine — the company or me?

You first, if you are the person responsible. The nFADP imposes criminal sanctions on natural persons — executives or controllers — of up to CHF 250,000 for intentional violations of the duties to inform, to grant access, of due diligence or of professional secrecy. The fine can be charged to the company (up to CHF 50,000) when identifying the individual would require disproportionate effort. That is a deliberate choice by the legislator: compliance is management's business.

PlaybooksAugust 18, 20267 min read

The nFADP and AI projects: the checklist for Swiss SMEs

Switzerland's data protection law applies to your AI projects from day one — and the fines target individuals, not just companies. The 7 questions to ask before launching, the LLM-specific traps, and a checklist to copy.

Also available in: Deutsch, Français

The day a client's e-mail enters a prompt, your AI project processes personal data — and the Swiss Federal Act on Data Protection applies, in full, from that day on. No transition period: the new act (nFADP) has been in force since 1 September 2023, and the Federal Data Protection and Information Commissioner (FDPIC) has confirmed that it is technologically neutral — AI gets no special regime, neither stricter nor more lenient.

The good news: for an SME, making an AI project compliant is neither a six-month legal undertaking nor a reason not to launch. It is a series of questions to ask in the right order, before launch. This playbook asks them, explains the vital minimum of each, and ends with a checklist to copy into your project management tool.

What the nFADP changes — and doesn't change — for an AI project

Three reference points to map the terrain, without jargon:

It protects natural persons. Data about clients, employees, prospects. Purely technical or corporate data is not covered — but a named professional e-mail address is.
Private processing is lawful by default. Unlike the European General Data Protection Regulation (GDPR), the nFADP does not require a legal basis for every processing activity. In exchange, it imposes specific duties: inform, secure, govern transfers, answer access requests. Your compliance is decided on those duties.
The sanctions target individuals. Up to CHF 250,000 in criminal fines for the responsible natural person, in case of intentional violation of specific duties. The legislator wanted data protection to be a management topic — not a file you delegate and forget.

The 7 questions to ask before launching

1. What personal data enters the system?

Inventory first. List what will actually enter the prompts, the indexed documents, the logs: names, e-mails, addresses, files, photos. Flag the sensitive data (health, opinions, biometric data, origin) — it triggers reinforced requirements at every subsequent step.

Then apply the most profitable principle in the whole act: minimisation. What does not enter the system does not need to be protected. Pseudonymise before the prompt when identity adds nothing to the task — an LLM summarising a file almost never needs the client's name.

2. Where does the data go?

Every AI provider is a link in a chain: the model, its cloud, their subprocessors. The question is not "is the provider well known?" but "in which country does the processing take place, and under which guarantee?"

Switzerland and the European Union: countries recognised as adequate, no extra formality.
United States: adequate if the provider is certified under the Swiss-U.S. Data Privacy Framework (DPF), recognised by the Federal Council since September 2024. Certification can be checked in two minutes on the official registry — per company, not per product.
Elsewhere, or a non-certified provider: standard contractual clauses to be built into the contract.

3. Does your provider train on your data?

The professional offerings (API, enterprise plans) of the major providers generally exclude training on your data by default; the free consumer offerings often do not. The difference must be contractual, not assumed: demand a data processing agreement (DPA) that fixes the usage, the retention period for prompts and the subprocessing chain. That is your duty of care as the controller — the processor acts on a mandate, and mandates are written down.

4. Does the AI make automated decisions?

If the system produces an automated individual decision — a legal effect or a significant impact on the person: refusing a loan, rejecting an application, terminating a service — the act requires you to inform the person and allow them to request a human review.

Our architectural position, which we apply in our own projects: keep the human in the loop at the decision points from the design stage. That is not just compliance — it is what makes the system trustworthy.

5. Is an impact assessment (DPIA) required?

A data protection impact assessment is mandatory when the processing presents a high risk — the act expressly cites the use of new technologies, large-scale processing of sensitive data, and high-risk profiling. Many AI projects tick the first box; not all of them present a high risk for that reason alone.

The useful reflex: document the evaluation even when the conclusion is "no DPIA needed". One page is enough. If the high risk is confirmed and persists after your measures, consulting the FDPIC (or your data protection advisor) is the next step.

6. What does your privacy policy say?

The duty to inform applies to all processing: purposes, recipients or categories of recipients, destination countries. If your AI assistant sends excerpts of client conversations to an American provider, your privacy policy must reflect it — before launch, not after the first awkward question.

7. Can you answer an access request?

Anyone can ask you what data you process about them. With an AI system, an honest answer requires knowing: what sits in your prompt logs, what is indexed in your document base, what your provider retains and for how long. Define a retention policy for prompts and outputs at the design stage — it is far simpler than reconstructing the history under pressure, within the legal 30-day deadline.

The LLM-specific traps

Four points where we see projects stumble:

Prompts are personal data. Everyone protects the database; few people think of the invocation logs, the debug traces, the screenshots in tickets. Your logs are part of the perimeter.
Shadow AI is already inside your walls. If your employees paste client data into a free consumer tool, your first compliance project is not contractual: it is a clear internal directive and a professional alternative to hand them.
Document indexes (RAG) inherit access rights. An internal search engine that answers everyone with everyone's documents is a personality violation waiting to happen. The source system's access controls must survive into the index.
The subprocessing chain is longer than it looks. Model → the model's cloud → your application's host → observability tooling. Every link that sees personal data must be under contract and, if outside Switzerland, covered by a transfer guarantee.

The checklist

Copy it as-is into your project tool — each box maps to a section above:

## nFADP compliance — AI project

### Inventory

- [ ] Incoming personal data listed (prompts, indexes, logs)
- [ ] Sensitive data identified and minimised
- [ ] Pseudonymisation applied where identity adds nothing

### Providers and transfers

- [ ] Processing country identified for every link in the chain
- [ ] DPF certification checked (official registry) or standard clauses signed
- [ ] DPA signed: no training on our data, retention defined
- [ ] Provider's subprocessors reviewed

### Design

- [ ] Automated decisions identified; human review planned
- [ ] Risk evaluation documented (DPIA carried out if high risk)
- [ ] Source-system access controls mirrored in the RAG index
- [ ] Retention policy for prompts and outputs defined

### Duties

- [ ] Privacy policy updated BEFORE launch
- [ ] Access-request procedure tested (deadline: 30 days)
- [ ] Internal AI directive communicated to employees
- [ ] Record of processing activities kept (recommended even if exempt)
- [ ] Breach notification: who decides, who notifies the FDPIC, by when

What comes next?

Switzerland has no AI act today, and that is a choice: in early 2025 the Federal Council opted to ratify the Council of Europe's AI Convention and make sector-specific adjustments, rather than build an equivalent of the European AI Act. If you serve clients in the EU, the AI Act may still concern you — we will devote a Swiss Radar article to it.

Until then, the nFADP is the entire playing field, and this playbook covers the essentials. The texts are authoritative: the act on Fedlex and the FDPIC's guides. And if you would rather put these questions to someone who has already worked through them while building AI systems — that is precisely the first workshop of our strategy engagement.

Frequently asked questions

Does the nFADP apply if I use Claude, ChatGPT or another LLM via API?: Yes, as soon as personal data enters a prompt — a client's name, an e-mail address, an HR file. The model provider then becomes a processor: you need a data processing agreement, verification of where the processing takes place and, if the data leaves Switzerland, a valid basis for the transfer (an adequate country, Data Privacy Framework certification for the United States, or standard contractual clauses).
Do I need people's consent to process their data with AI?: Not by default. Unlike the GDPR, the nFADP does not require a legal basis for every private processing activity: processing is lawful as long as it does not unlawfully harm the personality of the data subjects. Consent becomes necessary in specific cases — notably the disclosure of sensitive data to third parties, or high-risk profiling when you justify it by consent. The duty to inform, however, applies to all processing: your privacy policy must say what your AI does.
Does an SME have to keep a record of processing activities?: Companies with fewer than 250 employees are exempt, unless they process sensitive data on a large scale or carry out high-risk profiling. Our recommendation: keep it anyway. It is the document that answers a client's, an auditor's or the FDPIC's questions in one page — and drawing it up forces you to do the inventory this checklist demands anyway.
What is the difference between the nFADP and the GDPR?: The nFADP is lighter: no legal-basis requirement for every processing activity, no administrative fines at 4% of turnover, an optional record for most SMEs. But it is criminal where the GDPR is administrative: the fines — up to CHF 250,000 — target the responsible natural persons, for intentional violations of specific duties. And if you serve clients in the EU, the GDPR may apply in parallel.
Who risks the fine — the company or me?: You first, if you are the person responsible. The nFADP imposes criminal sanctions on natural persons — executives or controllers — of up to CHF 250,000 for intentional violations of the duties to inform, to grant access, of due diligence or of professional secrecy. The fine can be charged to the company (up to CHF 50,000) when identifying the individual would require disproportionate effort. That is a deliberate choice by the legislator: compliance is management's business.

#nlpd #conformité #llm #pme #sécurité #suisse

← All articles