ChatGPT in business: what you need to know before using it

No AI tool is as widespread in German companies as ChatGPT, and hardly any is used incorrectly as often. The typical case: the team uses personal accounts, nobody has checked the settings, and customer data ends up in the chats. Yet with the right plan and a handful of settings, ChatGPT can be run properly. Here is the practical guide, as of July 2026.

The core problem: personal accounts used for business

On the personal plans Free, Go, Plus and Pro, OpenAI uses your inputs to train its models by default. The toggle for this is called "Improve the model for everyone" and is switched on out of the box. On top of that, there is no data processing agreement (DPA) for personal plans. As soon as personal data goes into the chat, such as customer enquiries, job applications or attendee lists, you lack the contractual basis that Article 28 GDPR requires.

The business plans work differently: on Business (until summer 2025 this plan was called Team), Enterprise and the API, OpenAI does not train on your content by default, and you can conclude a data processing agreement. Important: the agreement does not apply automatically. You have to actively conclude it using the form provided by OpenAI.

Settings you should change right away

1. Turn off training: go to Settings, then Data Controls, then switch off "Improve the model for everyone". This only applies to future chats. Anything that has already gone into training stays there.
2. Check memory: ChatGPT can permanently remember facts from your chats. In the EU, the extended memory feature is switched off by default, but you should still check it under Settings, then Personalization.
3. Use temporary chats: for sensitive one-off questions, there are temporary chats without history, training or memory. OpenAI keeps a backup copy for up to 30 days.
4. Clean up shared links: shared chats can be read by anyone with the link. In summer 2025, tens of thousands of shared ChatGPT conversations ended up in Google's search index, including names and business details. Do not share anything that should stay internal.
5. Limit integrations: connections to Google Drive or SharePoint give ChatGPT access to entire file repositories. In 2025, security researchers showed that specially crafted documents can use these connections to leak data. Only connect folders without sensitive content.
6. On the Business plan: conclude the data processing agreement and go through the admin settings, from roles to two-factor authentication.

What can go into a prompt?

The basic rule is the same as for any external service: personal data only with a data processing agreement, so only in the business versions, and only as far as the purpose requires. Trade secrets and content you have promised customers to keep confidential never belong in a personal account.

Special case: professional secrecy

Lawyers, tax advisers and doctors are subject to Section 203 of the German Criminal Code (StGB): client and patient secrets may only be disclosed to service providers that have been formally bound to confidentiality as persons involved in the professional's work. As things stand, OpenAI does not offer such an agreement, not even on the Enterprise plan. For these professions, this means working with anonymised data only. How law firms can assess their use of AI step by step under Section 43e of the Federal Lawyers' Act (BRAO) is explained in the article AI in law firms: what is allowed.

Deleted does not always mean deleted

The legal dispute between the New York Times and OpenAI shows how much the deletion promises of US services are really worth. In 2025, a US court temporarily required OpenAI to retain all chats from personal accounts, including those deleted by users. Later, OpenAI had to hand over millions of anonymised conversations to the publishers that brought the lawsuit. Business and Enterprise customers, as well as API customers with an agreed zero data retention arrangement, were exempt from the order. The lesson: if you need confidentiality, you are much better off with the business versions.

Where is your data held?

For European users, the contracting party is OpenAI's Irish company, but processing still takes place mainly in the US. OpenAI bases the transfer on its certification under the EU-US Data Privacy Framework and, in addition, on standard contractual clauses. So far, data storage in Europe is only available for new Enterprise customers and API projects, not for the Business plan, and it only covers where data is stored, not where it is processed. In short: for small businesses, the transfer to the US remains the norm. It can be resolved contractually, but it needs to be properly documented in your record of processing activities.

What obligations apply to you?

• Conclude the DPA and archive it (Business, Enterprise or API).
• Update your record of processing activities and disclose the use in your privacy policy if customer data is processed.
• Train your team: the AI literacy obligation under Article 4 of the EU AI Act has applied since February 2025.
• Label AI: since 2 August 2026, customers must be able to tell when they are chatting with an AI (Article 50 EU AI Act). Deepfakes must also be labelled.
• Check AI output before it leaves your business. Hallucinated facts about real people can violate personality rights, and unchecked texts can infringe other people's copyright.
• Set rules: who may use ChatGPT, and for what? Our guide to AI policies shows how to do this on two pages.

Latest developments as of July 2026

There are three developments you should know about. First: the Italian €15 million fine against OpenAI from late 2024 was overturned by a court in Rome in March 2026, but only on grounds of jurisdiction. The substantive data protection questions remain open and now lie with the Irish supervisory authority. Second: in November 2025, the Munich I Regional Court (Landgericht München I) ruled that OpenAI infringes copyright through its use of song lyrics. OpenAI has appealed. Third: the legal dispute with the New York Times is ongoing and remains the most important real-world test of what deletion promises are worth.

With Klaus, a one-off check becomes a system

You can implement everything in this article by hand. The problem: ChatGPT rarely comes alone. Alongside it you have Copilot, the AI feature in your CRM and three browser extensions, and for each tool you would have to answer the same questions all over again. Klaus turns this into a process: in the onboarding chat, he records each tool as a tool profile. For ChatGPT, he already knows the typical setup and suggests it, so you only need to confirm or correct it.

In the process, Klaus asks about exactly the points that make the difference in this article: does the tool run via the website, the API or a cloud provider? What data goes in, and whose data is it? Are you subject to a duty of confidentiality under Section 203 StGB? From your answers, Klaus creates the classification for each tool and your AI policy as a PDF, which sets out the conditions for everyone to read. The team training then builds its examples on the tools you actually use, ChatGPT included. And if you use ChatGPT with customers, for example as a support chatbot, Klaus systematically checks in an assessment project which risk category and which transparency obligations under Article 50 apply.

Conclusion

Using ChatGPT in your business is not a data protection problem if you take care of three things: use a Business plan with a data processing agreement, go through the settings properly once and give your team clear rules. If, on the other hand, you stick with a personal Plus account and enter customer data, you have a problem that no small print will solve. Switching costs around €20 per person per month and an afternoon of work. That is the cheapest compliance win you can get this year.