Your team is probably already using ChatGPT, often on personal accounts and without shared rules. So the real question is not whether AI comes into your business, but on what terms. The Swiss Federal Act on Data Protection (FADP, in force since 1 September 2023) does not ban ChatGPT. It governs what you put into it, the contract with the provider and the countries your data goes to.
This guide sets out the position as of 30 October 2026: what OpenAI’s plans promise, what the law and the Federal Data Protection and Information Commissioner (FDPIC) say, the alternatives SMEs ask about most, and a checklist. Throughout, we separate the law, the authorities’ guidance and our advice.
General information, not legal advice. Providers change their terms often: check them before you sign.
ChatGPT plans and what they do with your data
Individual plans (Free, Go, Plus, Pro): training on by default
On the plans for individuals, OpenAI may use your conversations to train its models. You can opt out in two ways:
- in ChatGPT, go to Settings > Data controls and switch off ‘Improve the model for everyone’;
- or select ‘Do not train on my content’ in OpenAI’s Privacy Portal.
Temporary chats are not used for training. One catch: if you rate an answer with a thumbs up or down, OpenAI says the whole conversation may be used, even after you have opted out.
These plans come with no data processing agreement. Our advice: no client or staff data on personal accounts.
ChatGPT Business (formerly Team)
ChatGPT Team was renamed ChatGPT Business on 29 August 2025. According to OpenAI:
- workspace data is not used for training by default, and rating an answer does not change that;
- a data processing agreement (DPA) is available through an online form;
- admins control how long data is kept, and deleted conversations are removed within 30 days unless the law or the protection of OpenAI’s services or third parties requires longer;
- the plan has passed a SOC 2 Type 2 audit and offers SAML single sign-on and multi-factor authentication.
OpenAI’s help article on data residency does not list Business among the eligible plans.
ChatGPT Enterprise
The same commitments, plus custom retention policies and data residency at no extra cost. New Enterprise workspaces can store their content in the ‘Europe (EEA + Switzerland)’ region and, optionally, run model inference there too. Pricing is agreed with OpenAI.
The OpenAI API
This is what developers and integrators use, for example to automate email triage. Data sent to the API since 1 March 2023 is not used for training by default. It may be kept for up to 30 days for abuse monitoring, and zero data retention can be requested for eligible uses. A DPA and a Europe region are available.
Where does the data go?
For customers based in Switzerland, OpenAI’s contracting party is OpenAI Ireland Ltd, under its Services Agreement. Even with data residency, OpenAI says some processing can happen outside the chosen region: sign-in, routing, metadata, web search or connected apps. There is no Switzerland-only option.
ChatGPT prices in Switzerland, as of 30 October 2026
Prices shown on OpenAI’s pricing page when viewed from Switzerland:
| Plan | Price per user per month | Trains on your data by default |
|---|---|---|
| Free | CHF 0 | Yes, opt-out available |
| Go | CHF 7 | Yes, opt-out available |
| Plus | CHF 20 | Yes, opt-out available |
| Pro | From CHF 83 | Yes, opt-out available |
| Business, Standard seat | CHF 15 (annual) or CHF 19 (monthly) | No |
| Business, Premium seat | CHF 81.10 (annual) or CHF 102.70 (monthly) | No |
| Enterprise | On request from OpenAI | No |
Business needs at least two seats and is aimed at teams of 2 to 200 people. OpenAI’s help centre notes that prices vary by country and currency, so check the final amount and VAT when you order.
What Swiss law says
The FADP applies directly to AI
This is the FDPIC’s position, published on 9 November 2023 and updated on 8 May 2025: the Act is technology-neutral, so it already applies to AI-supported data processing, with no need to wait for an AI law. The Commissioner expects in particular that:
- the purpose, functioning and data sources of AI processing are transparent;
- people know whether they are communicating with a machine and whether their data is used to improve the system;
- a data protection impact assessment is carried out where the risk is high.
What the NCSC and the FDPIC recommend
On 16 October 2026, the National Cyber Security Centre (NCSC) and the FDPIC published recommendations on AI at work:
- use the Business or Enterprise versions your employer provides rather than personal accounts;
- before each prompt, look for personal data, client data and trade secrets, and anonymise them;
- know when your data is used for training and use opt-outs where they exist;
- for employers: clear internal rules, staff training, contracts that exclude training and, where needed, data loss prevention tools.
The articles that matter
| Article | What the law says | In practice with ChatGPT |
|---|---|---|
| Art. 6 FADP | Lawful, in good faith, proportionate, for a recognisable purpose, accurate | Send only what the task needs and check the answers |
| Art. 7 FADP | Data protection by design and by default | Restrictive defaults: training off, few connectors |
| Art. 8 FADP | Data security appropriate to the risk | Strong sign-in, limited access |
| Art. 9 FADP | Processors only under a contract, if no duty of confidentiality prevents it and security is ensured | Sign the provider’s DPA |
| Art. 16–17 FADP | Disclosure abroad: adequate country, safeguards (such as standard clauses) or exceptions | Check where data is processed, sub-processors included |
| Art. 19 FADP | Duty to inform, including recipients and countries abroad | Update your privacy notice |
| Art. 21 FADP | Automated individual decisions: inform and allow human review | Do not let AI decide a case on its own |
| Art. 22 FADP | Impact assessment where the risk is high | Sensitive data on a large scale, profiling |
| Art. 61–62 FADP | Fines up to CHF 250,000 for wilful breaches, on complaint | Aimed in principle at the responsible individuals |
Worth knowing: businesses with fewer than 250 employees are generally exempt from keeping a record of processing activities (Art. 24 DPO), unless they process sensitive data on a large scale or carry out high-risk profiling.
Transfers to the United States
Since 15 September 2024, the Federal Council has recognised an adequate level of protection for US companies certified under the Swiss-US Data Privacy Framework (Annex 1 of the DPO, which also lists the EU and EEA states). Without certification, you need other safeguards, such as standard contractual clauses recognised by the FDPIC (Art. 16(2) FADP). So check every sub-processor listed in the DPA, not just the company you sign with.
Professional secrecy: extra care
Art. 321 of the Swiss Criminal Code punishes, on complaint, the disclosure of secrets entrusted to lawyers, notaries, doctors, dentists, pharmacists, psychologists, auditors bound by secrecy under the Code of Obligations and their assistants, among others. The penalty is up to three years’ imprisonment or a monetary penalty. Disclosure is not an offence if the person concerned consents.
Other professions are not listed but are still affected. For an accounting firm that keeps its clients’ books, or a property manager, Art. 62 FADP may apply: it fines the wilful disclosure of confidential personal data learned while practising a profession, up to CHF 250,000. And Art. 9 FADP rules out using a processor where a legal or contractual duty of confidentiality prevents it.
Our advice: anonymise such data before entering it, get the client’s consent or use a Swiss-hosted model under contract. Ask your professional association too.
The alternatives and where they stand on data
Microsoft 365 Copilot and Copilot Chat
With a work account, prompts and responses are covered by Microsoft’s DPA and Product Terms, with Microsoft acting as processor, and are not used to train foundation models. For customers signed up in Switzerland, they fall under the EU Data Boundary, which covers the EU and EFTA, so Switzerland too. In November 2025 Microsoft announced in-country processing for Switzerland in 2026. On 3 April 2026 it said that for EU and EFTA countries this would be delivered at regional level, under the EU Data Boundary. Documented exceptions: web searches sent to Bing and the Anthropic models offered in Copilot.
Gemini in Google Workspace
Google says Workspace does not use customer data to train models without permission, that prompts count as customer data under its Cloud Data Processing Addendum, and that they are not reviewed by humans. Since June 2025, Enterprise Plus customers, or those with the Assured Controls add-on, can require Gemini processing in the EU or the US. There is no Swiss region. The consumer Gemini app follows different rules.
Mistral Vibe (formerly Le Chat)
The French company Mistral renamed Le Chat to Vibe in 2026. Data is hosted in the EU by default, with possible temporary transfers to listed sub-processors. According to its help centre (September 2026), Enterprise customers are opted out of training by default; individual users are not and must opt out themselves. On Team, the admin can switch training off for the whole organisation (‘Allow your interactions to be used to train our models’), so check that setting.
Claude (Anthropic) Team and Enterprise
Anthropic’s commercial plans (Team, Enterprise and the API) are not used for training by default, and its DPA is built into its Commercial Terms. Its individual plans (Free, Pro, Max) have had a training setting since 2025; if it is on, data is kept for up to five years. For processing in Europe, Anthropic points to AWS Bedrock and Google Cloud Vertex AI. Check the current terms for Team and Enterprise.
Apertus, the open Swiss language model
Apertus is built by EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS). The current version, Apertus 1.5, was released on 24 July 2026. It adds image and audio understanding and is published under the Apache 2.0 licence, with its weights, code and training data documentation. Where to use it:
- at Infomaniak, in its Euria assistant: the model runs on Infomaniak’s infrastructure in Switzerland and content is not used for training;
- at Swisscom, which opened it to business customers on its sovereign AI platform when Apertus launched in September 2025;
- on your own infrastructure, by downloading the weights from Hugging Face (8 and 70 billion parameter versions; the larger one needs far more computing power).
Open source does not mean compliant: hosting and the contract are what count.
Comparison table, as of 30 October 2026
| Tool | Trains on your data | Contract (DPA) | Where data is processed |
|---|---|---|---|
| ChatGPT Free, Go, Plus, Pro | Yes by default, opt-out available | No | No choice |
| ChatGPT Business | No by default | Yes | No data residency |
| ChatGPT Enterprise | No by default | Yes | Europe (EEA + Switzerland) available |
| OpenAI API | No by default | Yes | Europe (EEA + Switzerland) available |
| Microsoft 365 Copilot (work account) | No | Yes | EU Data Boundary, Switzerland included, with exceptions |
| Gemini in Workspace | Not without permission | Yes | EU or US, depending on edition |
| Mistral Vibe | Yes by default, except Enterprise; can be switched off | Yes | EU by default |
| Claude Team, Enterprise | No by default | Yes, in the terms | Europe via AWS or Google Cloud |
| Apertus hosted in Switzerland | Depends on the host | With the host | Switzerland possible |
Checklist for your SME
1. Choose the plan
- Ban personal accounts for work and give staff a business account.
- If you already use Microsoft 365 or Google Workspace, compare with the AI included there first.
- For automations, use an API with a DPA and, where possible, a European region.
2. Set it up
- Turn on single sign-on and multi-factor authentication.
- Enable only the apps and connectors you need: each one opens access to more data.
- Set a suitable retention period where the plan allows it.
- Mistral Team: switch off training in the admin console.
- Copilot: decide whether web search and Anthropic models should stay on.
- Any remaining individual accounts: training off, and temporary chats for tests.
3. What never to paste
Our rule, even on a business plan:
- data covered by professional secrecy, unless anonymised or consented to;
- sensitive data under Art. 5 FADP, such as data on health, religious or political views, ethnic origin, genetic or biometric data, criminal proceedings or social assistance;
- passwords, access keys or codes;
- HR files: appraisals, salaries, medical certificates;
- bank details or AHV numbers linked to a person;
- documents a client or partner gave you in confidence.
4. When to sign a DPA
As soon as personal data about clients, applicants or staff goes through the tool (Art. 9 FADP). Check the list of sub-processors, the countries involved, the transfer safeguards (Art. 16 FADP), breach notification and deletion at the end of the contract.
5. When to inform clients
- Always in your privacy notice: recipients such as the AI provider, and countries abroad with their safeguards (Art. 19 FADP).
- When a client talks to a chatbot: tell them it is a machine.
- When a decision about them is made automatically: tell them and offer human review (Art. 21 FADP).
- For data covered by professional secrecy: get their consent.
6. When to prefer a Swiss-hosted or open-source model
- Data covered by professional secrecy.
- Sensitive data or large volumes of personal data.
- Clients or contracts that require processing in Switzerland.
- A need to keep the same model version for a long time.
In these cases, a model such as Apertus, hosted in Switzerland or on your own servers, keeps transfers to a minimum. It takes more setting up and its results differ from the leading models, so test it on your real cases.
What about EU law?
If the output of your AI tools is used in the European Union, the EU AI Act can apply to your business even though it is based in Switzerland. Our guide to Swiss AI law explains this, with an up-to-date timeline.
Next steps
A written rule works better than a ban nobody follows. Our AI usage policy template helps you set out which tools are allowed and with what data. In our half-day team workshop, at your office, we apply these rules to your own examples and you leave with your policy.
For automations, we choose the model for each task and for how sensitive the data is: leading models (GPT, Claude, Gemini, Mistral…) or open-source models such as Apertus, hosted in Switzerland or on your own infrastructure for sensitive data. You decide where your data is processed. To get started, book a free AI audit.