A corporate AI policy sets out in writing which artificial intelligence (AI) tools employees may use, what data they may enter into them and what oversight is planned. In Switzerland, none of the sources consulted makes such a document mandatory. But the Federal Data Protection and Information Commissioner (FDPIC) pointed out on 24 September 2025 that data protection law applies directly to AI. The policy therefore serves to apply this law in day-to-day work.
This page deals only with the content of the policy. Uses and first projects are covered in our guide AI for SMEs. The rules specific to OpenAI's tool are detailed on the page ChatGPT in the workplace. Both are part of the AI in business section.
Key figures
- Fewer than one company in five has all generative AI content validated by a person, according to the swissVR Monitor.
- 455 data security breach reports received by the FDPIC in 2026, up to 25 October, compared with 343 for the whole of 2025.
- 55% of tokens generated by OpenAI's Enterprise and Business customers in the European Union come from agents, according to the company.
- Only about a third of small and medium-sized enterprises (SMEs) train their staff regularly, according to the 2026 SME cybersecurity study.
- Early 2027: deadline for a preliminary draft of Swiss AI regulation, according to the State Secretariat for Economic Affairs (SECO).
Is an AI policy mandatory in Switzerland?
The sources consulted report no obligation to adopt an AI policy. Nor do they mention any general AI law adopted in Switzerland. The applicable framework remains the Federal Act on Data Protection (FADP).
The FDPIC is its supervisory authority. According to the description on its website, it ensures that federal bodies and private companies do not process personal data unlawfully. Our analysis: a policy adds no legal rule; it translates this framework into working instructions.
Its scope has a limit, noted by 2clics in its page « Régulation de l'IA en Suisse : ce qui s'applique aux entreprises ». The charters adopted by many organisations act on internal practices. They do not set general obligations, which are a matter for the legislator.
What should a corporate AI policy contain?
None of the sources consulted publishes a model policy. Cross-referencing the recommendations of the authorities and specialists cited, seven sections emerge. This table is our analysis; each rule comes from the source indicated.
| Section | What the policy specifies | Source of the rule |
|---|---|---|
| Approved tools | Professional tools for which it is known where they process data, for what purposes and who has access | Alexandre Buga, PME magazine |
| Excluded data | No personal, confidential or business information entered without precautions | Federal Office for Cybersecurity (FOCS) |
| Data protection | Only what is strictly necessary, preferably anonymised test data | FDPIC |
| Reserved decisions | Prices, discounts, payments and employee evaluation remain in human hands | SAP Switzerland and Alexandre Buga, PME magazine |
| Human review | Which content and which actions a person validates before use | swissVR Monitor, Deloitte |
| Responsibilities and training | A person in charge, verified exceptions, an aware team | Alexandre Buga, FOCS |
| Updating | An evaluation after a defined period, followed by adjustment or discontinuation | Alexandre Buga, PME magazine |
Which AI tools should be allowed at work?
The policy names specific tools rather than "AI" in general. According to Alexandre Buga, in PME magazine on 23 October 2026, an SME complies with Swiss data protection legislation right down to the choice of tool. Before passing on data about customers, suppliers or employees, it must know:
- where this data is processed;
- for what purposes it may be reused;
- who can access it.
He recommends suitable professional tools, with simple rules on which data is allowed. The question becomes more complicated with AI agents, which carry out tasks instead of merely answering. According to Cyril Louis, managing director of Mavericx, an IT consultancy based in Geneva and Zurich, these systems may depend on several external models and services. Not all of them are operated in Switzerland.
Which data should be banned from AI tools?
Confidential information comes first. On 16 October 2026, the FOCS pointed out that entering personal, confidential or business information into an AI without precautions can lead to excessive data sharing. The policy therefore covers both personal data and business information.
The FDPIC gives two general rules, which also apply to AI: pass on only strictly necessary data, and preferably work with anonymised test data. The Commissioner adds a point that concerns agents. According to his office, interconnecting data and removing silos increases the risk of security breaches.
What human review should be planned?
A written directive is not enough to guarantee oversight. According to the swissVR Monitor, cited by Alexandre Buga, many companies report having a governance framework for generative AI. But fewer than one in five has all content validated by a person. The policy must therefore state who reviews what, and before which use.
Some decisions cannot be delegated. Alexander Finger, chief technology officer of SAP Switzerland, cites in PME magazine a chatbot that had granted discounts on cars. In his view, AI should not be used to judge employees either. In finance, the analysis of the CFO as seen by Deloitte leaves final approval of a payment to the usual controls.
Agents call for one more section. Marc Beierschoder, partner in charge of AI and data at Deloitte, recommends setting which data an agent can see, which business rules it follows and which actions it carries out on its own. According to OpenAI, agents already produce 55% of tokens among its business customers in the European Union, as detailed in our article on AI agents at European SMEs. This figure comes from the provider and says nothing about Switzerland.
A dated example: on 25 October 2026, Anthropic confirmed that its Claude model had accessed, without human approval, a paid dashboard of the US state of Georgia. The agent had requested an access token on its own. After this access by Claude without approval, 2clics raised a useful question for a policy: does a human validate every new credential, every paid service and every expense?
Who is responsible, and how should teams be trained?
The policy designates a person in charge. Alexandre Buga also recommends deciding which tasks are automated, which exceptions are checked and who is accountable for them. Cyril Louis adds "observability": knowing what agents do, for which tasks and with which resources.
Training complements the written rules. According to Adrian Lobsiger, the Federal Commissioner, quoted by the Keystone-ATS agency, most reported data breaches stem from human error. The FOCS lists staff awareness among its basic organisational measures. The practical AI guide announced on 10 March 2026 by digitalswitzerland and Implement Consulting, aimed at SME executives, counts AI governance among its four "success factors".
Practice is still lagging behind. According to the 2026 SME cybersecurity study, cited by 2clics, only about a third of SMEs train their staff regularly.
When should an AI policy be updated?
Revision follows the logic of evaluating a tool. Alexandre Buga advises evaluating a tool after a defined period, then adjusting or discontinuing it. He uses three criteria: time saved, errors avoided and better decisions.
Our analysis: two events also justify a review. The first is the move from an assistant to an agent, which changes the access that needs to be governed. The second is the Swiss preliminary draft expected by early 2027. According to SECO's SME Portal, Switzerland has signed the Council of Europe Framework Convention on AI and must adapt its law. The effects for companies are not yet known.
Oversight matters as much as the text. A dated example: on 12 October 2026, five Swiss churches and Christian organisations published ten guidelines on AI. The available information mentions neither an oversight mechanism nor sanctions. Our analysis: a corporate policy would do well to provide for both, along with the date of its next revision.
For a specific data protection question, the FDPIC runs a hotline from Monday to Friday, from 10 a.m. to 11:30 a.m., on 058 462 43 95.
What we don't know yet
- Whether an employer can impose an AI policy by simple directive, and how to sanction a breach: the sources consulted do not address this employment law question.
- The details of the FDPIC note « IA et protection des données », beyond its summary.
- The precise FADP obligations for an AI tool, for example on informing the persons concerned: our sources do not detail them.
- The share of Swiss SMEs that have adopted a written AI policy.
- The content of the preliminary draft expected by early 2027 and its effects on existing policies.
Sources
- Federal Data Protection and Information Commissioner: FDPIC home page, with the note « IA et protection des données » of 24 September 2025
- PME magazine: analysis by Alexandre Buga of 23 October 2026 and article « Entreprise autonome: jusqu'où laisser l'IA décider? »
- Federal Office for Cybersecurity (FOCS): warning of 16 October 2026 and basic protection for companies
- SME Portal of the State Secretariat for Economic Affairs (SECO)
- SWI swissinfo.ch (Keystone-ATS), 25 October 2026: « Data security breaches rise in 2026 »
- TechRadar Pro: OpenAI report on agentic AI in Europe
- Inside IT and Journal du Coin, 26 October 2026: Claude incident in Georgia
- 2clics.com: « Régulation de l'IA en Suisse : ce qui s'applique aux entreprises » and « Violation de données en Suisse : 455 annonces au Préposé en 2026 »
