GOVERNANCE & SECURITY
NUMEZISADVISORY &
ENGINEERING
Contents
DECISION AID · SWISS AI COMPLIANCE
Three areas, one record you can defend
The legal analysis becomes practical when each area ends with a person responsible, a set of records and a decision. The question is not whether the company uses AI, but whether it can explain each use when asked.
inventory, data flows, contracts
clauses, controls, assessments
training, single entry point, quarterly review
A defensible record answers who, what, where, why and on what terms within one working day.
A Swiss company using generative AI in 2026 is subject to two regimes at once. The first is explicit: the revised Federal Act on Data Protection (nFADP), in force since 1 September 2023, applies to every prompt, every integration and every data flow that contains personal data. The second is less visible: the EU AI Act, which came into force on 1 August 2024, has effects outside the Union that many Swiss organisations discover late, often in a customer questionnaire or a call for tenders.
The gap we see is not a lack of goodwill but a lack of structure: legal thinks in terms of texts, IT in terms of systems, management in terms of risks, and nobody thinks in terms of proof. Yet proof is exactly what will be asked for one day, by a customer, an auditor or an authority. This article offers a reading in three areas, the dated deadlines, then a 90-day plan to move from good intentions to a record you can actually show.
Two regimes already in force, and a Swiss blind spot
The nFADP is technology-neutral: it does not mention AI, and that is precisely why it already applies. The FDPIC has made the point: the law applies directly to AI-based data processing. As soon as an employee pastes a customer email into an assistant, or a connected system indexes staff files, the obligations kick in: purpose and proportionality, information, control of processors, conditions for disclosure abroad. The law also provides for criminal fines aimed at the individuals responsible, not just the company.
The AI Act follows a different logic: it is product regulation, graded by risk, with fines proportional to turnover and phased application (see the timeline below).
The blind spot fits in one sentence: Switzerland has no general AI law. In February 2025 the Federal Council chose to ratify the Council of Europe convention on artificial intelligence, which Switzerland signed in March 2025, and to adapt existing law sector by sector, with a draft expected for consultation. Many conclude that there is a legal vacuum. That is wrong on two counts: the nFADP already covers the raw material of generative AI, which is data, and the AI Act does not stop at the border.
The three areas
Rather than piling up analyses one text at a time, we frame each situation through three areas. Each answers a different question, involves different people and produces its own records.
Area 1: personal data. Which personal data goes into our AI use, and do we comply with the nFADP along the whole chain: purpose, information, processors, disclosure abroad, impact assessment?
Area 2: the European market. Through our customers, our subsidiaries or the output of our systems, are we within the scope of the AI Act, and in what role: provider or deployer?
Area 3: sector and contracts. What obligations come on top of the law: professional secrecy, supervisory expectations, customer clauses, audit requirements?
A single use can trigger one, two or all three areas. An internal assistant for technical documentation often only triggers the first, and lightly. Assisted screening of applications for jobs in Geneva and Lyon triggers all three. It is this combination, not the tool, that determines how much compliance effort is justified.
YOUR AI USE
PERSONAL DATA
nFADP · FDPIC
EU MARKET
AI ACT REACH
SECTOR AND CONTRACTS
SECRECY · AUDITS
Area 1: personal data (nFADP)
Purpose and proportionality. Unlike the GDPR, the nFADP does not require a legal basis for every processing operation; it requires compliance with principles (a defined purpose, proportionality, accuracy, security) and a justification where processing departs from them. For each use, the company must be able to say why the data is processed and why that is proportionate. "Being more productive" does not justify feeding in the entire customer history.
Information. The privacy notice must reflect actual use: categories of data, recipients (including model providers), countries of processing. Automated individual decisions must be disclosed when they have legal or significant effects on the person concerned (Art. 21).
Processors. An AI provider that processes personal data on your behalf is a processor (Art. 9): a contract is required, along with security guarantees and control over its own sub-processors. The consumer terms of many tools do not meet these requirements; business plans generally do, but check with the contract in hand.
Disclosure abroad. Processing takes place somewhere. EU countries offer protection recognised as adequate. For the United States, companies certified under the Swiss-US Data Privacy Framework have been recognised since 15 September 2024; for an uncertified provider or another country, safeguards are needed, typically standard contractual clauses, and you must check that they are effective. The place of processing becomes a criterion for choosing a provider, not a technical detail.
Impact assessment. Where processing is likely to pose a high risk to personality or fundamental rights (profiling, sensitive data, new technologies on a large scale), an impact assessment is required before deployment (Art. 22). Generative AI applied to HR, health or financial data often ticks these boxes. If the residual risk remains high, the FDPIC must be consulted.
Area 2: the European market (AI Act)
"We're Swiss, the AI Act doesn't concern us" is one of the most expensive sentences of 2026. The regulation also covers organisations established outside the EU, and a Swiss company can come within its scope in three ways.
The market. Anyone placing on the European market a product or service that includes an AI system is concerned, wherever their head office is. That includes a Swiss software company whose AI feature is sold to European customers.
The output. The regulation also covers cases where the output of a system operated outside the EU is used in the EU. A Geneva company screening applications for its Lyon branch, or delivering analyses produced by its system to a German customer, can fall within scope without "selling AI".
The group. A European subsidiary using the group's AI tools is a deployer, with its own obligations, and passes those requirements back up to the parent company.
Role matters as much as scope. Most Swiss companies are deployers: their obligations concern proper use, human oversight, logs and monitoring. You become a provider, with far heavier obligations, by marketing under your own brand a system built on a third-party model, or by substantially modifying an existing system.
Two obligations apply to almost every company within scope, whatever the level of risk:
- AI literacy (Art. 4). Providers and deployers take measures to ensure that the people using AI systems on their behalf have a sufficient level of competence, given their role and the context. In practice: training suited to how the tools are used, and a record of who completed it.
- Transparency (Art. 50). People must know they are dealing with an AI system (a chatbot on a website, for example); synthetic content must be marked by its providers; deepfakes and certain texts published to inform the public must be flagged as generated or altered by AI.
The AI Act timeline
| Date | What applies |
|---|---|
| 1 August 2024 | The regulation enters into force |
| 2 February 2025 | Prohibited practices and the AI literacy obligation (Art. 4) |
| 2 August 2025 | Obligations for providers of general-purpose AI models, governance, penalties |
| 2 August 2026 | Most other obligations, including transparency (Art. 50) and the high-risk systems in Annex III (recruitment, credit, access to essential services…) |
| 2 August 2027 | High-risk systems built into regulated products (Annex I) |
In November 2025, the European Commission proposed, in a simplification package known as the "digital omnibus", to postpone some of the obligations for high-risk systems. Check whether this text has been adopted before setting your own timetable. Waiting for the last deadline is still a losing strategy: European customers are already demanding proof in their contracts today.
Area 3: sector and contracts
This area is the one most often forgotten, yet it bites fastest: no authority acts more quickly than a customer who cancels a contract.
Professional secrecy. Lawyers, notaries, doctors, auditors and their assistants (Art. 321 of the Swiss Criminal Code), banks (banking secrecy), fiduciaries (duty of discretion towards their clients): passing information covered by secrecy to an external AI service without a suitable framework is a risk in its own right, separate from the nFADP. The right question is not "is this provider reputable?" but "is this channel compatible with my duty of secrecy?"
Sector supervision. Financial institutions bring their AI use within operational risk management and outsourcing rules; healthcare and the public sector have their own requirements. A supervisory authority will not wait for an AI law before asking questions about a critical process handed to a model.
Customer clauses. Confidentiality agreements, data processing contracts, bans on subcontracting without consent, data location requirements: a use of generative AI can breach a contract before it breaches a law.
Audit requirements. ISO 27001, SOC 2, customer audits and due diligence now cover AI. An unmanaged use becomes an audit finding, with a direct effect on sales.
What to keep on record
| Area | Questions to ask | Records to keep |
|---|---|---|
| Personal data (nFADP) | What data goes into which uses? For what purpose, with which processors, to which countries? | Record of processing activities covering AI, data processing agreements, the basis for each transfer abroad, impact assessments for high-risk uses |
| European market (AI Act) | Is our output used in the EU? Are we a provider or a deployer? Which uses are high-risk or fall under Art. 50? | Map of EU exposure, role for each system, training completed (Art. 4), transparency notices, tracking of the timeline |
| Sector and contracts | What secrecy obligations, customer clauses and supervisory expectations apply? | Review of key contracts, register of commitments, rules of use by category of data and customer |
What this means in practice
An inventory of uses. Not just the tools you have bought, but actual use, including undeclared use. For each one: data processed, users, provider, channel (consumer account, business plan, programming interface) and output produced.
Classification. Each use is given a level according to the data involved and the consequence of a mistake. For data, we use three levels: internal, customer and protected, the last covering sensitive data within the meaning of the nFADP, information subject to secrecy and trade secrets. Classification decides everything else: which tool is allowed, whether an impact assessment is needed, whether a person must check the output.
Supplier clauses. Content retention, exclusion from training (by default and by contract), place of processing and storage, list of sub-processors, how you get your data back at the end of the contract.
Technical controls. Single sign-on and role-based access first; then limiting the data that connected systems can reach, separating environments and addressing the risks specific to language models (prompt injection, data leakage, poisoned context), documented in particular by the OWASP Top 10 for LLM Applications.
Logging. Who used which system, when and for what. It must stay proportionate: monitoring staff would create a new problem under the nFADP.
Training. A short policy on use, examples for each role, a common foundation for everyone and additional material for exposed functions (HR, legal, finance, customer service). This is also the answer to Art. 4 of the AI Act.
Common mistakes
The phantom inventory. Management believes there are three uses; the survey reveals twenty-five, several of them on personal accounts.
The default contract. The tool was adopted on consumer terms: data that can be reused, uncontrolled retention, unknown sub-processors.
The decorative impact assessment. Done after deployment and never updated.
The "we're Swiss" reflex. Exposure to the AI Act is discovered in a European customer's questionnaire, at the end of the sales cycle.
A rule with no tool. A policy forbids what the infrastructure allows, without offering a business tool: the ban produces undeclared use, not compliance.
A decision with no record. Reasonable calls were made, but by nobody in particular, on no particular date, in no particular document.
The 90-day plan
Days 1 to 30: establish the facts. Inventory of actual uses; data per use; uses whose output reaches the EU; current supplier contracts. Deliverable: a list of uses that are permitted, tolerated under conditions or prohibited, signed off by management.
Days 31 to 60: close the gaps. Move the relevant uses to business plans with checked clauses; update the privacy notice; carry out impact assessments for the two or three most exposed uses; single sign-on, access rights and logging; transparency notices where Art. 50 applies.
Days 61 to 90: make it stick. A published and taught policy on use, with training recorded; a single entry point for any new use, with short response times; a register of decisions (who approved what, when and on what conditions); a quarterly review that also tracks the AI Act timeline and the Swiss draft legislation.
Success is not measured by the volume of documents. It is the ability to answer within a day when a customer, an auditor or an authority asks: which uses, which data, which clauses, which decisions?
The starter checklist
- List every use of generative AI, including undeclared ones, with data, users and providers.
- Classify each use by data (internal, customer, protected) and by the consequence of a mistake.
- Flag the uses whose output is used in the EU and determine your role.
- Check for each provider: retention, training, place of processing, basis for the transfer, sub-processors, how you get your data back.
- Update the privacy notice and the information on automated decisions.
- Carry out an impact assessment before deployment for high-risk uses.
- Review professional secrecy and customer clauses before approving any new use.
- Switch on single sign-on, role-based access and proportionate logging on approved tools.
- Train users according to their role and keep a record of the training.
- Record each approval decision with its owner, date and conditions, and schedule the quarterly review.
How we can help
The AI Workshop trains a team of up to 20 people in half a day, with the one-page sheet on permitted data: a concrete starting point for the AI literacy required by Art. 4. AI Adoption Support takes care of the inventory, choosing and configuring tools with your IT team, the rules of use and the regular review. For an SME, the guide on using AI without putting your data at risk sums up the method in six steps.
This article is an operational analysis for management teams; it is not legal advice on your particular situation. Texts and drafts as at 26 October 2026.
Discuss this decision
hello@numezis.com ↗