On this page (9 sections)
Switzerland has no dedicated AI act in 2026. AI is regulated through the revised Federal Act on Data Protection (FADP, in force since 1 September 2023), through sector supervisors such as FINMA and Swissmedic, and — for any system whose output reaches the EU market — through the EU AI Act, whose high-risk obligations started applying on 2 August 2026. The detail that catches legal teams off guard: FADP fines are personal, up to CHF 250,000 against the individual who made the call, not against the company.
Key takeaways
- Switzerland's approach is sector-specific, not framework-first. A federal AI bill has been in preparation with a consultation draft expected around the end of 2026.
- The FADP, not an AI act, is the operative text for most AI deployments — and its Art. 21 automated-decision rights are the clause that matters most.
- The EU AI Act reaches Swiss providers whose system output is used in the EU. Establishment is not the test; use is.
- Swiss residency plus zero retention removes a class of vendor risk. It does not remove your DPIA obligation.
- If your model is wrong often enough to matter, no amount of data residency fixes that. Pick the model on measured accuracy first.
What is the current state of AI regulation in Switzerland in 2026?
As of November 2026, Switzerland has no sector-agnostic AI act. The Federal Council chose to adapt existing law rather than build a new horizontal regime, so AI obligations arrive through data protection law, sector regulation, and the Council of Europe Framework Convention on AI, which Switzerland signed in 2024. A federal AI bill has been in preparation, with a consultation draft expected around the end of 2026.
In practice that means four regulators do most of the work:
- FDPIC — data protection, including automated decision-making, DPIAs, and cross-border transfers.
- FINMA — governance and risk management for AI in supervised financial institutions, plus existing outsourcing and cloud rules.
- Swissmedic — software as a medical device, aligned with the EU MDR via mutual recognition.
- SECO and the FOITT — product safety and federal IT procurement respectively.
The FDPIC's guidance page on artificial intelligence and data protection is the page to bookmark, though it was returning an error when I checked it in November 2026.
A note on sourcing. Two of the references I was given for this piece — the CASUS explainer on AI and Switzerland's Data Protection Act and SIX's story on AI regulation in Switzerland — were also returning 404s in November 2026. I link them because they are the references I was handed, but treat the FDPIC as the authoritative source and check any secondary summary against the statute.
How does the Swiss Federal Data Protection Act (FADP) apply to AI?

The FADP applies to AI whenever personal data is processed — training, inference input, logging, or evaluation. It has no AI-specific chapter; instead, four provisions do the heavy lifting: Art. 6 (principles), Art. 7 (privacy by design and default), Art. 21 (automated individual decisions), and Art. 22 (data protection impact assessment).
Art. 21 is the one compliance teams under-read. Where a decision is taken exclusively by automated means and has legal consequences or significantly affects the person, that person can demand to be informed, express their point of view, and request that a natural person review the decision. A confidence score in a log file is not a review.
Art. 22 requires a DPIA where processing is likely to result in a high risk — and profiling, mass processing of sensitive categories, or scoring decisions generally qualify. The register of processing activities (Art. 12) is only mandatory above 250 employees or for high-risk, high-volume processing, which is narrower than the GDPR's equivalent.
For teams running classification rather than generation, explainable AI with calibrated probabilities is what makes Art. 21 workable: a calibrated probability is one where a 0.8 actually means the model is right about 80% of the time, which is something you can show a reviewer.
GDPR vs Swiss DPA for AI: what actually differs?

The two regimes are close enough that a GDPR programme covers most of the FADP, but the differences are real and they bite in specific places. The FADP has no one-stop-shop, no mandatory DPO, a narrower register requirement, and — most importantly — criminal rather than administrative fines aimed at individuals.
| Attribute | Swiss FADP | EU GDPR |
|---|---|---|
| Automated decisions | Art. 21: inform, express view, request human review | Art. 22: right not to be subject to the decision |
| DPIA trigger | High risk; consult FDPIC if residual risk stays high | Art. 35; consult supervisory authority |
| Breach notification | To FDPIC "as soon as possible" | 72 hours to supervisory authority |
| Maximum penalty | CHF 250,000, personal liability | €20m or 4% of turnover, corporate |
| DPO | Not required; advisor recommended | Required in defined cases |
| One-stop-shop | None | Lead supervisory authority |
| Register of processing | Only 250+ staff or high-risk/high-volume | Broadly required |
The practical consequence of the penalty row: under the FADP your engineering lead and your head of legal can be personally exposed for a decision they signed off. That changes how you document model choices — and it is the single strongest argument for keeping an auditable decision trail rather than a prompt string.
Navigating the EU AI Act: when does it apply to Swiss-based AI?
The EU AI Act applies to Swiss providers when the output of the system is used in the EU — establishment is not the test. Art. 2 reaches providers and deployers in third countries where the output produced by the system is used in the Union. If a Swiss-hosted model classifies EU customer messages, you are in scope.
The timeline as it stands:
| Date | What applies |
|---|---|
| 1 August 2024 | AI Act enters into force |
| 2 February 2025 | Prohibited practices |
| 2 August 2025 | General-purpose AI model obligations |
| 2 August 2026 | High-risk obligations (Annex III) |
| 2 August 2027 | High-risk obligations for Annex I products |
The Commission has proposed postponing parts of the high-risk timetable. I would not build a compliance plan on a date you have not re-checked against the consolidated text this quarter.
One structural point favours you: a classification model that returns a label and a probability is a much simpler thing to document than a generative system. There is no sampling, no prompt drift, no free-text output to explain. If you are weighing that trade-off, Laya Studio versus OpenAI classification via LLM prompts walks through where each one wins.
AI data residency in Switzerland: what does the "Swiss advantage" actually buy you?
Swiss residency buys you a specific legal position: personal data sits under FADP jurisdiction, and — if the provider has no US parent exercising control — it is outside the reach of the US CLOUD Act. It does not exempt you from the EU AI Act, and it does not exempt you from your own DPIA.
What it does change is vendor risk. Switzerland has long been recognised by the EU as providing an adequate level of data protection, so transfers into Switzerland are not a Chapter V problem in the way transfers to the US are. For a health-tech team, that removes an entire annex from the transfer impact assessment.
What does zero data retention mean for Swiss AI in practice?
Zero retention means the provider does not store your input text or your output after the response is returned, and does not train on it. That is a contractual and architectural claim, and you should verify it rather than accept it — ask where the logs live, how long they live, and who can query them.
For Laya specifically: no text is generated at all, so there is no completion to store, and the service runs in Switzerland with zero data retention. The honest limits of that are worth stating plainly:
| Claim | What it gets you | What it does not |
|---|---|---|
| Data in Switzerland | FADP jurisdiction; no US CLOUD Act exposure absent US control | EU AI Act still applies if output is used in the EU |
| Zero retention | No training on your data; no stored corpus to breach | Does not remove your DPIA obligation |
| No text generation | Cannot hallucinate a label | Does not make the model accurate |
AI liability in Switzerland: who is accountable when a model gets it wrong?
Switzerland has no AI-specific liability statute. Claims run through the general tort provision (Art. 41 of the Code of Obligations), the Product Liability Act for defective products, and contract — which means liability questions are answered by ordinary negligence reasoning: what did you know, what did you document, and what would a reasonable operator have done.
That makes documentation the actual defence. Three things I would keep for every automated decision in a regulated process:
- The model version and the date it was evaluated.
- The measured accuracy on a held-out set that resembles your traffic.
- The threshold you chose and the false-positive/false-negative cost that justified it.
And the accuracy point deserves honesty. On Banking77, Laya's model scores 0.425 against TypeSafe Jev's 0.870. If your use case is banking intent classification, that gap is decisive — use Jev. Residency and retention do not compensate for a model that is wrong more often.
Sector-specific AI rules: finance and medical data in Switzerland
Finance and healthcare are the two sectors where Swiss AI regulation is genuinely more than FADP plus good intentions. FINMA published guidance on governance and risk management when using AI for supervised institutions in late 2024, layered on top of its existing outsourcing and cloud circulars; Swissmedic regulates AI as software within the medical device framework, aligned with the EU MDR.
Finance. FINMA's position is that using AI does not change the underlying obligations: you remain accountable for the outcome, model risk must be governed, and outsourcing to a cloud provider requires notification and a documented ability to audit. A credit-scoring or fraud model is a governance object, not a feature. There is a working pattern for real-time AI for compliance monitoring beyond quarterly audits that fits this model.
Medical data. Health data is a sensitive category under the FADP, so the DPIA threshold is met almost by default. Add Swissmedic device classification if the software informs diagnosis or treatment, and the Human Research Act if you are processing data for research. The compliance considerations for real-time AI decisioning in healthcare are a reasonable starting checklist.
Best practices for AI compliance in Switzerland
Treat the FADP as your primary text, the EU AI Act as a scope question, and sector guidance as the tiebreaker. The checklist below is what I would actually run before a Swiss AI deployment goes live.
- Confirm whether output is used in the EU. If yes, classify against Annex III before anything else.
- Run a DPIA if the processing is high-risk — profiling, sensitive categories, or scoring.
- Implement Art. 21: a real path to human review, not a support email address.
- Measure accuracy on a held-out set that resembles production traffic, and record the number.
- Document the threshold and its cost trade-off.
- Verify the retention claim contractually: where logs live, how long, who can query them.
- Check the sub-processor chain for US control if CLOUD Act exposure matters to you.
- Re-check the EU AI Act timetable each quarter; the high-risk dates have been in motion.
If you want to test the residency and retention claims against your own traffic rather than take my word for it, Laya Studio's pricing and free tier start at 100,000 free questions and then $5 per million — enough to run a real evaluation before you commit to anything.
FAQ
Does the EU AI Act apply to Swiss companies?
Yes, when the output of the AI system is used in the EU. Art. 2 of the AI Act reaches providers and deployers established in third countries on that basis, so a Swiss-hosted model classifying EU customer data is in scope regardless of where the servers sit.
Is the Swiss FADP equivalent to the GDPR?
Close but not equivalent. The FADP has no one-stop-shop, no mandatory DPO, a narrower register requirement, and criminal fines capped at CHF 250,000 that fall on individuals rather than the company. A GDPR programme covers most of it, but the automated-decision rights in Art. 21 and the personal penalty exposure need separate attention.
Do I need a DPIA for an AI classification pipeline in Switzerland?
Usually yes, if the processing is high-risk — which includes profiling, large-scale processing of sensitive data such as health information, or automated decisions with significant effects. Art. 22 of the FADP requires it, and if residual risk stays high after mitigation you must consult the FDPIC before proceeding.
What does zero data retention mean for a Swiss AI vendor?
It means your input and output are not stored after the response is returned, and are not used for training. Verify it contractually: ask where logs are written, how long they persist, and who can query them. A provider that cannot answer those three questions has not implemented zero retention, whatever the marketing page says.
Can an AI decision be made without human review under the FADP?
Yes, but the person affected can demand review. Art. 21 gives them the right to be informed, to express their point of view, and to have the decision reviewed by a natural person, unless an exception such as explicit consent or a statutory requirement applies. Build the review path before launch, not after the first complaint.
Topics
- Swiss AI data regulations 2026
- Switzerland AI legal framework
- Swiss Federal Data Protection Act AI
- FADP AI compliance
- AI data residency Switzerland
- GDPR vs Swiss DPA for AI
- AI liability Switzerland
- Swiss financial sector AI regulations
