Imagine an employee pasting the transcript of a customer meeting into ChatGPT and having it generate a summary. It’s quick. Whether this is permissible under data protection laws, however, depends less on the provider than on three factors: the version and account she’s using, the data she enters, and the task for which she’s doing it.
Can a Swiss company use ChatGPT?
Yes. Swiss companies are permitted to use ChatGPT, Claude, Gemini, Copilot, and other cloud-based AI tools, including the free versions. In principle, data must not be stored in Switzerland.
No, that does not mean that any customer, employee, or confidential data can be entered. Whether an entry is permitted depends on the purpose, the provider’s terms and conditions, the security measures in place, and any transfers of data abroad. The responsibility lies with the company, not with the provider.
There is no need for a new AI law to address this. In May 2025, the EDÖB reaffirmed that the current Data Protection Act applies directly to AI. The following guidance applies to private companies. Government agencies and professions with special confidentiality obligations—such as doctors, lawyers, or trustees—require additional review. Our article “Using AI Safely: Approving Cloud AI in Compliance with the nDSG and GDPR” explains how to approach governance, licensing tiers, and in-house solutions.
What exactly does "data location" mean?
Many people ask right away, “Is the data stored in Switzerland?” That’s a valid question, but it’s only part of the answer. Four factors are important for the legal review:
- Storage: Where are chats, uploaded files, and backup copies stored?
- Processing: Where does the model read your input and calculate the response? Even processing that does not involve permanent storage is relevant under data protection law.
- Access: From which countries can support staff or service providers access the content? Access from abroad may also constitute a disclosure to a foreign country.
- Recipients: Which companies receive the data, including subcontractors, and what laws do they are subject to?
Think of it like a package that’s stored in a Swiss warehouse but is briefly shipped to a workshop in the U.S. for processing. In that case, you have to secure the transfer to the U.S., even if the warehouse is located in Zurich. It’s exactly the same with a file that remains stored in Switzerland but is sent to a model in the U.S. That’s why “server in Switzerland” or “data in Europe” alone isn’t enough.
Here's what that means depending on the country:
- Switzerland: No notification is sent abroad regarding this step.
- EU and EEA: The Federal Council recognizes an adequate level of protection. You just need to check whether the data is transferred from there to other countries.
- United States: Permitted if the specific recipient company is certified under the Swiss-U.S. Data Privacy Framework and the transfer is covered by it (official list), or using standard contractual clauses that have been adapted for Switzerland and whose validity has been verified in the specific case, with additional measures if necessary (notification abroad, Art. 16 DSG).
In short: The data location determines how much protection you need. It does not, on its own, determine whether an AI service is permitted. Storing data in Switzerland is not a general requirement. However, it may become necessary because customer contracts, confidentiality obligations, industry standards, or your own policies require it.
Two more factors that determine the traffic light system
- Training: The provider uses your input to improve future models. Imagine a translator who copies your letters into his workbook: He gets better, but your content ends up in his possession as well. The commercial versions from OpenAI, Anthropic, Google, and Microsoft do not train on your data by default.
- Data Processing Agreement (DPA): The service provider processes your data solely on your behalf and in accordance with your instructions. You’re familiar with this concept from the accounting firm that handles your books but is not permitted to use the figures for its own purposes. The DSG requires a contract for this, and subcontractors need your prior approval. Be sure to review the contract’s provisions regarding compliance with instructions, confidentiality, subcontractors, security, data deletion, and assistance with requests for access to data.
The traffic light system serves as a guide for selecting AI services. It does not replace a data protection review or approval. The specific data used, the intended purpose, the contracts with the provider, and the actual configuration are the decisive factors.
This results in the following traffic light system for handling customer, employee, and confidential company data:
- Green: Generally considered safe for typical business applications and generally well-suited, provided the necessary contracts, settings, and protective measures are in place. Typical example: Business version with a General Terms and Conditions (AVV) and without training; data storage can be selected in Switzerland or Europe; the risk associated with disclosure abroad is manageable.
- Yellow: Suitable for business use if additional requirements are reviewed and met, particularly regarding data processing or disclosure abroad. Typical example: Business version with AVV and no training, but with storage or processing in the U.S. or without a selectable location; you ensure the security of the transfer as described above.
- Red: Suitable for general tasks that do not involve sensitive, personal, or confidential information. Not suitable for customer or employee data without appropriate contractual and technical safeguards. Typical example: A personal account without a General Terms and Conditions agreement.
Even "green" doesn't mean that every step takes place in Switzerland or Europe: Depending on the settings and function, some processing may occur outside those regions. The entries below list these exceptions, such as Flex Routing at Microsoft.
Red does not indicate a legal ban on the product. It means: Without proven safeguards, the relevant data should not be included. A paid personal subscription such as ChatGPT Plus or Claude Max does not change this, and even disabling the training model is no substitute for an appropriate contract. The traffic light system evaluates versions, not providers; swissAI does not recommend any specific product.
The Traffic Light System: Which Version Is Consistent with Data Protection?
The information is taken from the providers' privacy policies and contract documents, as of October 2026. Each entry follows the same order: Traffic Light, Training, AVV, Data Location, Classification.
ChatGPT (OpenAI)
- ChatGPT Free, Plus, and Pro: Red
- Training: Yes, by default; can be disabled in the settings under “Improve the model for everyone.” According to the Data Controls FAQ, OpenAI does not use temporary chats for training and deletes them after 30 days.
- AVV: No.
- Data location: cannot be selected.
- Classification: Good for general ideas, phrasing, and programming questions; not for customer data.
- ChatGPT Business (formerly Team): Yellow
- Training: No, excluded by default according to OpenAI.
- AVV: Yes, you sign it using OpenAI's DPA form; this also applies to Enterprise and API.
- Data location: The name “Business” does not imply any commitment to Europe or Switzerland. OpenAI documents the choice of region for eligible Enterprise, Edu, and API customers (OpenAI), not for Business.
- Classification: Permitted if the transfer to the U.S. is secure; review the applicable General Terms and Conditions and the site-specific conditions.
- ChatGPT Enterprise: Green
- Training: No, by default.
- AVV: Yes, via OpenAI's DPA.
- Data Location: Authorized customers should select the “Europe” region —that is, the EEA and Switzerland—for new workspaces. Calculating responses in this region is an optional feature, and certain additional processing may take place outside of it. “Europe” does not refer exclusively to Switzerland.
- Classification: The version with the most control options at OpenAI.
- OpenAI API: Yellow
- Training: No, by default.
- AVV: Yes, using the DPA form.
- Data location: Europe (selectable by eligible customers). OpenAI retains logs for abuse monitoring for up to 30 days; data is only deleted after review by OpenAI.
- Classification: yellow, because you're building this yourself and are responsible for its use with the API.
Microsoft Copilot
Microsoft is gradually merging the personal Copilot and Microsoft 365 Copilot into a single app. Therefore, the account you use to sign in is crucial; according to Microsoft, personal and business accounts will remain separate.
- Copilot with a personal Microsoft account: Red
- Training: Yes, in principle, according to the Privacy FAQ, with exceptions for certain countries and user groups; can be disabled in the privacy settings.
- AVV: No.
- Data location: beyond your company's control.
- Classification: a personal device, even if it runs on a company laptop.
- Microsoft 365 Copilot and Copilot Chat with a work account (Entra ID): Green
- Training: No. According to the documentation, Microsoft does not use prompts, responses, or data from Microsoft Graph to train the base models.
- AVV: Yes, through Microsoft’s Data Protection Addendum and Product Terms (Enterprise Data Protection). This also applies to Copilot Chat with a corporate account without an additional Copilot license: “Free” does not mean “unsuitable” in this context; what matters is corporate access.
- Data location: If Switzerland is your tenant's default region, Microsoft states in the Product Terms to store Copilot interactions in Switzerland—for Copilot starting in March 2024 and for Copilot Chat starting in September 2025. This commitment now applies only as long as a setting in the Admin Center is enabled. Processing takes place within the EU data boundary, not in Switzerland: Microsoft had announced in November 2025 that processing would also take place within Switzerland; starting in April 2026, Microsoft plans to process data for the EU and EFTA exclusively within the region (Microsoft). There are three settings you can check in the Admin Center:
- "Store Microsoft 365 customer data in-country " determines whether data from Copilot, Exchange, SharePoint, OneDrive, and Teams remains in Switzerland. For tenants created after September 14, 2026, this setting is turned off by default; for older tenants, the News Center indicates how it is configured. If it is turned off, Microsoft may move the data within the EU data boundary starting December 14, 2026: The data will remain in Europe, but no longer necessarily in Switzerland. If you need Switzerland as the storage location—for example, due to customer contracts—you should turn the setting on.
- Flex Routing allows processing in the U.S., Canada, or Australia during peak loads, while the content remains stored within the EU data border. If you need to prevent this, simply turn it off.
- The Anthropic models in Copilot are exempt from the EU data boundary. Simply disabling Flex Routing is not enough for them; in the EU and EFTA, they are disabled by default.
- Note: Copilot finds everything that the signed-in user is authorized to view; make sure to review and adjust any overly broad permissions in SharePoint and Teams beforehand. The web search feature in Copilot has its own privacy policy.
Google Gemini
- Gemini with a private Google account: Red
- Training: Yes, as long as the “Keep Activity” setting is turned on; this is the default setting for users 18 and older; it can be turned off. According to the Gemini Apps Privacy Hub, some chats are reviewed by humans and retained for up to three years. Google itself advises against entering confidential information there.
- AVV: No.
- Data location: cannot be controlled.
- Conclusion: Manual review and long-term storage are not suitable for company data.
- Gemini in Google Workspace Business (Starter, Standard, Plus): Yellow
- Training: No. According to the Workspace Privacy Hub, prompts and workspace content are neither used for training nor reviewed by people outside your organization without your permission.
- AVV: Yes, through Google's Cloud Data Processing Addendum. Gemini must run under the enterprise terms, meaning with a corporate account.
- Data Location: A Workspace account alone does not guarantee a location in Europe.
- Classification: Permitted if the transmission is secure; the administrator will determine which location-based controls your edition offers.
- Gemini in Google Workspace Enterprise Plus with data region: Green
- Training: No, as with the Business editions, according to Workspace Privacy Hub.
- AVV: Yes, through Google's Cloud Data Processing Addendum, using the business account.
- Data Location: For the Gemini app, Google states that data is stored and processed in the selected region, such as Europe (Workspace Updates, June 2026). There is no separate Swiss region. The administrator must select the region and verify the available features.
- Note: This is the Workspace version with location-based controls; “Europe” does not refer exclusively to Switzerland.
Anthropic Claude
- Claude Free, Pro, and Max: Red
- Training: Can be disabled. You can decide in the privacy settings whether Anthropic is allowed to use your chats. If you allow it, Anthropic will retain the data for up to five years; otherwise, for 30 days (Anthropic).
- AVV: No. Pro and Max are also consumer plans.
- Data location: cannot be controlled.
- Note: Even with training disabled, a private version without AVV is available.
- Claude Team, Enterprise, and API: Yellow
- Training: No, by default; unless someone actively submits feedback, such as via the rating buttons (Privacy Center). Admins can disable this feature.
- AVV: Yes. Anthropic processes the company data as a data processor; the Data Processing Addendum, which includes standard contractual clauses, is part of the commercial terms.
- Data location: According to Anthropic, data is stored in the United States; requests may be processed in the United States, Europe, Asia, and Australia.
- Note: Permitted if these data transfers are secured for your use. If you need a specific region, use Claude via a cloud platform, such as Amazon Bedrock (starting with the Zurich region ) or Google Vertex AI. This is not the same as the Claude web app.
Mistral Vibe (formerly Le Chat)
- Mistral Vibe Free: Red
- Training: Yes, by default; can be disabled (Mistral Help Center).
- AVV: No.
- Data location: By default, within the EU.
- Classification: EU location, but still a private version.
- Mistral Vibe Team and Enterprise: Green
- Training: No, by default.
- AVV: Yes, through the Data Processing Addendum for business customers.
- Data location: By default, within the EU; depending on the function, data may also be temporarily stored outside the EU, according to Mistral, with contractual safeguards.
- Classification: A European provider that stores data in the EU.
Swiss Offers
There are also a number of Swiss AI chatbots. It’s important to note that “Swiss-hosted” does not automatically mean that the model runs in Switzerland. The same four questions regarding storage, processing, access, and recipients apply. You can find a comparison of twenty options in the overview of Swiss AI chatbots.
Under what circumstances is the purpose permitted?
The service is only half the story. The other half is the purpose. It is permitted if you are legally authorized to process the relevant data for that specific purpose. “We want to work more efficiently” is too vague for this. A useful description of the purpose might be: “We respond to customer inquiries; the AI generates draft responses based on the necessary details from each inquiry.”
It’s important to note that using an AI tool does not automatically constitute a new purpose. If a customer sent you an email so you could process her request, an AI-generated draft response serves the same purpose. The situation is different if the content is used to train third-party models, for advertising, or to create new personality profiles. You must assess this additional use on your own.
Under the Swiss Data Protection Act (DSG), private companies do not need consent for every routine data processing activity. However, they must comply with the following principles: purpose limitation, proportionality, transparency, and data security. If data processing results in an infringement of personal rights, it must be justified—for example, by an overriding interest, a legal basis, or valid consent.
If the GDPR applies, every instance of data processing requires a legal basis, such as a documented legitimate interest; the performance of a contract serves as a basis only to the extent that the processing is truly necessary for that purpose. “It’s in our privacy policy” is not a legal basis in and of itself. Incidentally, the GDPR applies to a Swiss company not simply because it uses a European server, but rather, for example, if it has a branch in the EU, if it makes targeted offers to individuals in the EU, or if it monitors their behavior (Art. 3 GDPR). The article “The EU AI Act for Swiss Companies” explains what additional requirements the European framework for AI imposes.
Typical Office Tasks: What's Possible?
Many tasks in everyday Swiss office life can, in principle, be performed using AI. The following examples illustrate how these principles play out in everyday practice. As soon as actual personal data is involved, any approval requires a verified business version.
This is also possible with a personal or free account, as long as it does not contain any personal data or confidential content:
- Presentation on a general subject: using factual information from public sources.
- Excel formulas or a short script: A table structure and fictitious sample data are sufficient; no real customer records or login credentials are required.
- Sales trends using truly anonymized totals: no conclusions can be drawn about individual persons; trade secrets must be protected.
Permitted with the business version, subject to the applicable limit:
- Responding to customer emails: Use the relevant inquiry and omit unnecessary personal details.
- Create a quote based on a customer inquiry: Use the details provided for the requested service.
- Import invoices, suggest journal entries, prepare reminders: only the necessary invoice and payment data; review results before sending.
- Translate or summarize a supplier contract: Pay attention to confidentiality agreements and authorized subcontractors.
- Summarize project documents and identify tasks and deadlines: Only authorized individuals receive the documents and results.
- Clean up CRM duplicates and standardize addresses: maintain existing contacts; review changes; do not create additional profiles.
- Summarize a customer meeting based on your own notes: omit personal asides and unnecessary details.
- Search internal documents using an AI knowledge assistant: existing access rights remain in place; HR documents must not suddenly become visible to everyone.
With regard to human resources matters , Article 328b of the Swiss Code of Obligations (OR) also applies: Employers may, in principle, only process data that pertains to a candidate’s suitability for the position or is necessary for the employment relationship. Consent does not simply override this restriction. For job applications, this means: include only information relevant to the position; do not draw conclusions about health, religion, or personality based on a photo or writing style; and do not issue automatic rejections. For employment references: include only documented, work-related achievements, and the evaluation must be reviewed by a human.
In meetings, criminal law also applies. Anyone who records a non-public conversation and has it transcribed by AI must first obtain the consent of all participants (Art. 179ter of the Swiss Criminal Code) and explain the purpose, the AI service, and the retention period. A visible record button or an AI bot in the meeting does not replace this consent.
I wouldn't make the following publicly available:
- Upload the entire mailbox to respond to a single complaint: Select the relevant correspondence.
- Secretly assessing employees' motivation or intent to resign based on their messages: This is an additional, highly intrusive form of analysis—not part of everyday office life.
- Sharing customer emails for the provider's training purposes: This is not covered by the original purpose of customer service.
- Analyzing health-related information from customer emails for sales profiles: data requiring special protection and a new purpose; a business subscription is not sufficient for this.
Here’s an example of the “purpose limitation” principle: Extracting the amount, due date, and payee from an invoice for accounting purposes serves the original business purpose. Deriving personal circumstances from the same invoice and building a sales profile constitutes a different use. This distinction applies regardless of whether the server is located in Zurich, Frankfurt, or Virginia.
When is it okay to send the first real customer email?
Once you've clarified and briefly documented the use case:
- Task and Data: The task is permissible, and the AI receives only the necessary information. Data is truly anonymous only when no one can be identified; simply removing names is not enough, and personally identifiable information that is publicly accessible remains personally identifiable information.
- Provider and Contract: The correct business version requires a corporate account, and the General Terms and Conditions apply.
- Countries and Data Transfers: Storage, processing, and access are documented and legally safeguarded. Record the recipients, countries, and legal basis for the transfer.
- Operations: Training and feedback functions are regulated, as are apps, integrations, and sharing links that may trigger additional data sharing. Access rights and retention periods are defined, as are the procedures for handling requests for information and deletion. The privacy policy specifies the recipient countries and the safeguards in place. Employees are familiar with clear rules regarding what information may be included.
This review covers an entire category of tasks, such as “draft responses to common customer inquiries.” It does not need to be repeated for every email; new, sensitive content or new features require a new assessment. No authorization from a government agency is required for this. If an application could pose a high risk to data subjects—for example, in the case of health data or decisions affecting individuals—the law requires a data protection impact assessment to be conducted beforehand.
What should employees never enter?
Even the best business version is no substitute for thinking things through. Here’s a rule of thumb: Would you send this information via email to an external service provider without a contract? If not, it doesn’t belong in a private version. In addition:
- Login credentials: Passwords and API keys do not belong in any AI tool, not even in the business version.
- Personal data in private versions: no names, addresses, or meeting notes regarding customers, employees, or suppliers.
- Personal data requiring special protection: health data, religious or political views, and information regarding criminal proceedings—only after prior clarification, including in business contexts.
- Professional and business secrets: patient records, client documents, or information covered by a confidentiality agreement—only once it has been determined whether you are permitted to share them.
Rules like these only work if everyone knows them. You can find certified AI trainers to provide training.
Frequently Asked Questions
Is ChatGPT Free prohibited in the workplace?
No. For general tasks that do not involve personal data or confidential information, you may use free versions—for example, for ideas, wording, or programming questions. However, customer and employee data should not be included unless there is a contract in place and appropriate safeguards are in effect.
Does the data have to be stored in Switzerland?
No, the nDSG does not require data to be stored in Switzerland. Data may be transferred abroad if an adequate level of protection exists there or if appropriate safeguards are in place. The entire chain—including storage, processing, access, and recipients—is decisive. Customer contracts or professional confidentiality obligations may impose stricter requirements.
Do I need my customers' consent to reply to their emails using AI?
Under the Swiss Data Protection Act (DSG), generally not, as long as the AI serves the same purpose for which you received the email, you use only the necessary information, and you provide transparent information in your privacy policy. This is not sufficient for new purposes, such as profiling or training.
What is the difference between ChatGPT Business and ChatGPT Enterprise?
When it comes to data protection, the location of the data is key. By default, neither service trains on your data, and both are subject to OpenAI’s Terms of Service. Enterprise also offers eligible customers data storage in Europe—including Switzerland—and a retention period of their own choosing; for Business, OpenAI does not guarantee a location in Europe.
Does Microsoft Copilot comply with data protection regulations?
With a business account, yes—as long as you set it up correctly. Microsoft 365 Copilot and Copilot Chat operate under Microsoft’s Data Protection Addendum, without training using your data. Before proceeding, check the permissions in SharePoint and Teams, and review three settings in the Admin Center: in-country storage, Flex Routing, and the Anthropic models. Storage in Switzerland applies only as long as “Store Microsoft 365 customer data in-country” is enabled; processing takes place within the EU data boundary.
Conclusion
The question “Can we use ChatGPT?” is actually too broad. The right question is: Which version, with which account, with which data, and for what task? Personal versions are suitable for general tasks, not for customer data. The business versions from OpenAI, Microsoft, Google, Anthropic, and Mistral can be used in compliance with data protection regulations if the contract, data transfers, settings, and rules are all aligned. The data location determines how much safeguards are needed, not whether you’re allowed to use AI.
Where does your company stand today? If you want to address this in a structured way with your executive team, the free “AI Workshop and AI Strategy for SMEs” package can help.
In a nutshell:
- AI services are permitted, including free versions; however, customer and employee data should only be stored in secure business versions.
- The data location encompasses storage, processing, access, and recipients; it determines security measures, not authorization.
- Data transfers to the United States are permitted with DPF certification from the recipient or approved standard contractual clauses.
- Business versions do not train on your data by default and come with a standard service agreement; with OpenAI, you sign it via a form.
- Using AI for the same task is nothing new; training, advertising, and profiling, on the other hand, are.
- Before sending your first actual customer email, go over the data, contract, data transfers, and operations for each task category.
- HR-related tasks and meeting recordings are subject to additional restrictions under the German Commercial Code (OR) and criminal law.
As of October 14, 2026. This article is intended as a general overview and does not constitute legal advice; the assessment depends on the specific circumstances of each case.
Sources: FDPIC on the DSG and AI · Data Protection Act on Fedlex · Federal Council on the Swiss-U.S. Data Privacy Framework · Art. 328b of the Swiss Code of Obligations (OR) · Art. 179ter of the Swiss Criminal Code (StGB) · OpenAI Enterprise Privacy · OpenAI Data Residency · Microsoft Copilot Privacy · Microsoft Product Terms on Data Location · Microsoft Flex Routing · Google Workspace Privacy Hub · Gemini Apps Privacy Hub · Anthropic Privacy Center · Mistral Data Processing Addendum

