ChatGPT, Copilot & Co. in the company: How to use AI in compliance with GDPR (status 2026)

In almost every company, employees now use ChatGPT, Microsoft Copilot or similar tools – often without the management knowing exactly what and with what data. The question “Are we allowed to do this at all?” can be answered: Yes, AI assistants can be used in compliance with data protection. But not with every tariff, not with every attitude and not without a few clear rules. This post shows what matters in 2026 – including some changes that many don’t have on their radar yet.

Note: This article provides a technical and organizational overview and does not replace legal advice. Involve your data protection officer or your data protection officer.

The short answer

An AI assistant is acceptable under data protection law in the company if four conditions are met: Contract processing contract with the provider, your input will be not for training the models used, you know, where the data is processedAnd your employees know clearly Ruleswhat data you can enter. The private free version, which many started with, usually does not meet the first two points.

Private tariff or business tariff: the decisive difference

For the large providers, there is a clear dividing line between offers for private individuals and for companies. Take the example of ChatGPT:

• Free, Plus and Pro these are private rates. OpenAI does not offer an order processing contract according to art. 28 GDPR on, and conversations can be used for training unless you turn that off in the settings. For the processing of customer, employee or business data, these tariffs are unsuitable.
• Business, Enterprise and the API are intended for companies. There is an order processing contract here, and input is not used for training by default.

For Microsoft Copilot, Google Gemini and other providers, the same principle applies: The decisive factor is not the product, but the tariff and the contract behind it.

Where is the data processed?

A lot has moved here in 2026 – in both directions.

OpenAI offers a data residence in Europe for ChatGPT Enterprise, Edu and Healthcare and has extended this to the processing itself, i.e. the calculation of the answers, in January 2026. According to reports from September 2026, the option will gradually become available for ChatGPT Business as well. Let the current status for your tariff be confirmed in writing by the provider.

Microsoft in principle, Microsoft 365 stores and processes data from the EU Data Boundary in Europe. For Microsoft 365 Copilot, however, there is an important limitation: Since the spring of 2026, Microsoft Copilot requests can also be redirected to data centers outside the EU (“flex routing”). The stored data remains in the EU, but the calculation of the response is not necessarily. The function is active by default and can be disabled by administrators. If you need exclusively European processing, you should now check this attitude – as well as the question of which additional AI models from third parties are unlocked in your own client.

What Doesn't Belong in an AI Assistant

Even with the right tariff, there is data that you should not enter or only after careful examination:

• special categories of personal data such as health data, religious beliefs or biometric data,
• complete customer or personnel data, such as exports from CRM or payroll accounting,
• Trade secrets that are specially protected by contract, such as under a non-disclosure agreement with customers,
• Access data, passwords and API keys,
• Decisions about people that AI should make alone – automated individual decisions are by type. 22 GDPR in principle inadmissible.

Seven steps to safe use

1. Inventory: Which AI tools are actually used today – officially and unofficially? The “shadow AI” with private accounts is the greatest risk in many companies.
2. Select and release tools: Choose one or two tools with Business plan and make them available to everyone. Those who have a good official offer rarely access the private account.
3. Conclusion of contracts: Examine order processing contract, document subprocessors and data locations, exclude training use.
4. Data Protection Impact Assessment: When processing personal data extensively, it is often a duty – and in any case a good opportunity to think through risks in a structured way.
5. Formulate usage policy: Brief and understandable: Which tools are allowed, which data are allowed in, how are results to be checked? One page is often enough.
6. Involve works council: AI tools that can capture employee behavior or performance are usually subject to co-determination. Early integration saves time later.
7. Employees empower: The EU AI Act requires companies to promote the AI competence of their employees. A short training on possibilities, limits and rules fulfills this and improves the results. More about this in the contribution to EU AI Act.

When sensitive data is in play: the alternatives

For many tasks, a well-configured business tariff is sufficient. But where confidential documents, customer data or construction knowledge are to be processed, there are two other ways:

• European providers and sovereign cloud: AI services operated entirely under European law. An overview of the options gives our contribution to sovereign cloud.
• Own system in the house: Open language models can now be operated with reasonable effort in your own data center. Combined with one RAG system the AI then answers questions from your own documents – without data leaving the company. What this looks like in practice, we show in the contribution to KI on-premises.

Frequent questions

Is ChatGPT GDPR compliant?

It depends on the tariff. The company tariffs (Business, Enterprise, API) offer a contract processing contract and exclude the use of training as standard and can be used in compliance with data protection. The private rates are not suitable for the processing of personal company data.

Are employees allowed to use ChatGPT privately for work?

This must be discouraged. Without an order processing contract, there is no legal basis for the processing of personal data. It is better to provide an official tool and exclude the use of private accounts for official purposes in a policy.

Is Microsoft 365 Copilot automatically compliant with data protection because we already use Microsoft 365?

Not automatically. Copilot accesses all data to which the respective user has access – excessively broad permissions in SharePoint and Teams are suddenly visible. Before introduction, check the permissions and settings for data processing, especially for flex routing.

Do we need a data protection impact assessment?

If AI processes personal data extensively or new high-risk technologies are used, in many cases yes. Your data protection officer or your data protection officer can assess this in individual cases.

Conclusion

Banning AI assistants in the company does not work in practice – the use then only shifts to private accounts. The better way is a deliberately chosen, cleanly configured tool with clear rules. This is less effort than many fear, and creates the basis for using AI productively and without a bad conscience.

If you want to introduce AI safely in your company or put the existing use on a solid footing, we support you with our AI consulting in Stuttgart – from tool selection and the directive to your own, data protection-friendly AI solution.