Swiss FADP (nLPD) and AI: 2026 Compliance Guide for SMEs

You use ChatGPT to draft client emails. Your accountant runs PDFs through Claude. Your marketing team is testing Midjourney with your product photos. And somewhere in the back of your mind, a question lingers: is any of this actually legal in Switzerland?

Short answer: not always. The revised Federal Act on Data Protection (FADP / nLPD), in force since September 1, 2023, modernised Switzerland’s framework to match the EU’s GDPR. It applies fully to AI projects — even if you assumed you were too small to care.

This guide is written for SME leaders, IT managers and DPOs in Switzerland who want to understand concretely what they can and cannot do with AI, without the jargon.

Context: what is the revised FADP?

The new FADP replaced the 1992 version. It’s supervised by the Federal Data Protection and Information Commissioner (FDPIC / PFPDT).

Main novelties:

Privacy by design and by default — protection must be built in from the start.
Register of processing activities mandatory for companies with 250+ employees, strongly recommended otherwise when risk is high.
Data breach notification to the FDPIC “as soon as possible”.
Data Protection Impact Assessment (DPIA) mandatory for high-risk processing.
Criminal sanctions up to CHF 250,000 for the natural person responsible (not the company).
Right to data portability and reinforced rights for data subjects.

Unlike GDPR, FADP fines target individual decision-makers (director, DPO, IT lead), not the company. That changes how leadership perceives the risk.

What this means for your AI projects

AI isn’t named explicitly in the FADP. But three provisions directly concern it.

1. Automated individual decisions (art. 21 FADP)

When an automated system makes a decision with “legal effect” or that “significantly affects” a person, you must:

Inform them
Allow them to request human review
Document the decision logic

Concretely: AI credit scoring, automated CV filtering, dynamic pricing based on customer profile — all fall in scope.

2. High-risk profiling (art. 5 FADP)

Profiling means automated evaluation of personal aspects (behaviour, health, economic situation). Once profiling is “high-risk”, you enter the strictest regime: explicit consent required, reinforced documentation.

3. Transparency (art. 19 FADP)

Every individual must be able to know:

What data you collect about them
Why
Who can access it (including AI sub-processors)
Whether data leaves Switzerland

This last point is widely overlooked, and where most SMEs unknowingly breach the law.

The 5 most common pitfalls

What we consistently see in SMEs rolling out AI without a proper framework.

Pitfall 1: Using public ChatGPT with client data

Consumer ChatGPT (Free or Plus) stores your conversations and can use them to train future models. If your accountant copy-pastes an Excel sheet with client names, addresses, and turnover figures, you are non-compliant.

Rule: never send identifiable personal data to a consumer LLM. Use Enterprise tiers (OpenAI Enterprise, Anthropic Claude for Work) that provide a DPA and guarantee no reuse of data.

Pitfall 2: Data leaving Switzerland unframed

OpenAI servers sit mostly in the US. Anthropic’s too. The FADP requires transfers to countries without equivalent protection to be framed (standard contractual clauses, explicit consent, or adequacy decision).

Switzerland recognises adequacy with the US via the Swiss-US Data Privacy Framework for certified companies. Check your AI vendor is on the list.

Pitfall 3: SaaS LLM without a signed DPA

Many SMEs sign up to Make or n8n Cloud without reading the terms. If these tools process personal data on your behalf, you are the controller and they are processors. The FADP requires a sub-processing contract (art. 9).

Pitfall 4: Logs that never get purged

Chatbot conversations, sent prompts, generated outputs — all of it gets logged. Three years on, you have a full historical record of client or employee queries. Without a purge policy, you breach the storage limitation principle (art. 6 FADP).

Pitfall 5: No AI processing register

Even under 250 employees, once you profile or process sensitive data via AI, a register becomes strongly recommended. It’s also your first line of defence during an FDPIC audit.

10-point AI compliance checklist

What we use with clients before deploying any AI project:

Data mapping: which personal data feeds the AI system?
Legal basis: consent, contract, legitimate interest? Documented.
Sub-processors identified: every AI vendor (LLM, hosting, vector DB) listed.
DPAs signed with every processor handling personal data.
Data location: where it’s stored, processed, transits.
Security measures: encryption in transit and at rest, access control, audit logs.
Retention period defined with automatic purge.
Subject information updated in your privacy policy.
Rights exercise mechanism: how does a user request deletion?
DPIA performed if high-risk processing.

Swiss hosting: where and how?

To cut international transfer risk, Swiss hosting remains the best path. Solid providers:

Infomaniak (Geneva): cloud, VPS, object storage — all in Switzerland.
Exoscale (Switzerland + EU): AWS-equivalent IaaS.
Swisscom Cloud: for heavy regulatory constraints.
Self-hosting on a dedicated server in a Swiss datacenter (Interxion, Green IT).

For AI itself, several options keep data in Switzerland:

Open-source LLMs (Llama 4, Mistral, Qwen 3.5, GLM-5) on your own infra.
Ollama + vLLM to serve models locally.
Mistral AI (France, EU): GDPR-adequate, culturally close.

For no-code automation: self-hosted n8n on a Swiss VPS beats Zapier or Make Cloud on compliance.

# Sample docker-compose for self-hosted n8n (Infomaniak VPS)
services:
  n8n:
    image: n8nio/n8n:latest
    environment:
      - N8N_HOST=automation.your-sme.ch
      - N8N_PROTOCOL=https
      - GENERIC_TIMEZONE=Europe/Zurich
      - DB_TYPE=postgresdb
    volumes:
      - ./n8n_data:/home/node/.n8n

DPIA: when is it mandatory?

A Data Protection Impact Assessment is mandatory when processing is high-risk. For AI that means:

Systematic profiling of individuals (scoring, automated evaluation)
Large-scale sensitive data processing (health, political opinions, minors)
Systematic surveillance (behavioural analysis)
New technology with strong impact (biometric generation, facial recognition)

The FDPIC publishes a DPIA template you can use directly. Count 2–3 days of work for a solid first DPIA.

What you actually risk

CHF 250,000 maximum, on the individual decision-maker.
Possible publication of the decision (reputational risk).
Civil damages claims from data subjects.
Cease-and-desist order (forced shutdown of the AI project).

The FDPIC favours education over enforcement, but investigations have accelerated since 2024. Three public decisions sanctioned non-compliant AI uses in 2025.

How we handle it at skifo/

Our defaults across every project:

Swiss hosting by default (mainly Infomaniak).
LLM via Mistral or self-hosted models when sensitive data shouldn’t leave the EU/CH.
DPAs provided for every project with third-party processors.
AI processing register delivered with each project.
Automatic log purge configured from the POC stage.
DPIA included when the use case warrants it.

We never send client data to US LLMs without a signed DPA and, when technically feasible, prior anonymisation.

FAQ: 5 common questions

My employees use ChatGPT on personal accounts. What’s my exposure?

If company data (clients, contracts, internal docs) flows through it, you’re liable as employer. The fix: clear internal policy plus ideally a framed alternative (Claude for Work, Mistral Le Chat Enterprise, or an internal instance).

We’re a 15-person SME, are we really in scope?

Yes. The FADP doesn’t scale obligations by size in most areas. Only the processing register can be lightened under 250 employees, and not always.

Are embeddings stored in a vector DB “personal data”?

If a person can be reconstructed or re-identified from those vectors (technically feasible), yes. Treat them as such.

Is anonymisation enough to exit FADP scope?

Real anonymisation (irreversible) takes data out of scope. Pseudonymisation (replacing a name with an ID) does not — it remains personal data.

Can we use Microsoft 365 Copilot?

Yes, with the Enterprise plan that bundles a DPA and EU Data Boundary. Make sure your tenant is set to “Switzerland North” or “Switzerland West” for maximum compliance.

The FADP isn’t designed to block your AI projects. It’s designed to frame them. A well-prepared SME can deploy powerful AI solutions while staying compliant — and that’s a commercial advantage in regulated sectors (health, finance, legal, public).

Want a quick diagnostic? We offer a free 30-minute AI compliance audit that identifies the three biggest risks in your current setup.