Using ChatGPT safely in your business

Public information and your own texts without personal data can go into any version. Internal documents and customer data only in a business environment with agreements, and preferably without names. Passwords, citizen service numbers (BSN) and medical data don't belong in any chat.

A handy classification is the traffic light:

Green: fine in any AI tool

• Public information, such as texts from your own website
• General questions: 'how do I write a clear quote?'
• Your own drafts without names, figures or customer details

Orange: only in the business tool, according to the agreements

• Internal documents, procedures and handbooks
• Customer emails and files, preferably with the names taken out
• Confidential business information, such as quotes, contracts and figures

Red: never in a chat tool

• Passwords, API keys and login details
• Citizen service numbers (BSN), copies of identity documents and bank details of individuals
• Special categories of personal data, such as health, religion or criminal records
• Information covered by a duty of confidentiality, unless you have separate agreements on it

Pseudonymizing: a habit that pays off

Often you don't need the name at all. 'Jan de Vries from Utrecht is complaining about his delivery of 12 March' becomes 'Customer A is complaining about a delivery'. The answer is just as good. Teach your team that habit and most of the risk disappears.

Note: just removing the name isn't always enough. A combination of job title, place of residence and a rare event can still make someone identifiable.

With the business versions (Business and Enterprise), OpenAI doesn't use your data to train models by default, you can sign a data processing agreement and the organization manages the accounts. With the free version and personal plans, none of that is the case.

OpenAI writes that by default it doesn't use data from ChatGPT Enterprise, ChatGPT Business, ChatGPT Edu and the API platform to train or improve its models (OpenAI, 2026). With free and personal accounts, OpenAI can use your conversations to improve models, unless you turn that off in the settings. Even if you do, you still don't have a data processing agreement or any oversight as an organization.

The main differences:

• Training on your data. Off by default for business accounts; for personal accounts only off if the user turns it off themselves.
• Data processing agreement. Only with the business versions.
• Management. With a business account, the organization decides who has access, and you take back an account when someone leaves. With a private account, the knowledge, and the data, leave with the employee.
• Retention periods and extra security. More extensive with Enterprise.

ChatGPT Business costs $ 20 per user per month billed annually or $ 25 per month, from two users, excluding VAT (OpenAI, checked on 26 October 2026; check the current price with the provider). For most SMEs, that's the starting point. Enterprise is on request and intended for larger organizations with additional requirements.

The same logic applies to Claude and Microsoft Copilot: the business versions don't train on your data by default.

Yes, as soon as personal data goes into ChatGPT. OpenAI then processes that data on your behalf, and the GDPR (Article 28) asks you to make written agreements about it. OpenAI offers business customers a data processing agreement.

In that relationship, you are the controller: you decide why and how the data is used. OpenAI is the processor. The data processing agreement sets out what the processor may and may not do, how the data is secured, which sub-processors are involved and what happens in the event of a data breach.

Three things to arrange alongside the data processing agreement:

• Your record of processing activities. Note in it that you use an AI service, for what purpose and with which data.
• A DPIA if the risk is high. If you process sensitive data on a large scale with AI, for example in healthcare or when assessing people, a data protection impact assessment is often mandatory. Ask your privacy officer.
• Transfers outside the EU. If data ends up in the US, that transfer has to be properly arranged. That belongs in the agreements with the provider. Check it, or have it checked.

This sounds like a lot of paperwork. In practice, it's an afternoon's work for someone who has done it before, and then it's in place.

That depends on your plan. With new ChatGPT Enterprise and Edu workspaces, you can choose storage in Europe, and since January 2026 also processing in Europe. For the API, you can create projects with European data residency. With Business and the personal versions, check for yourself where your data is stored.

OpenAI introduced data residency in Europe in February 2025. For ChatGPT Enterprise and Edu, that means: conversations, uploaded files and custom GPTs are stored in Europe, only for new workspaces. For the API, you can create a project in which requests are handled in Europe and not retained (OpenAI, 2025). In January 2026, processing in Europe was added for eligible Enterprise and Edu customers (OpenAI, 2026).

If data in the EU is a hard requirement, for example in healthcare or government, you have roughly three routes:

1. 01ChatGPT Enterprise with European storage and processing.
2. 02Azure OpenAI in a European region, within your own Microsoft environment. That's not a chat app for everyone, but the basis for an application you have built.
3. 03Microsoft 365 Copilot, which works within the EU Data Boundary for European customers. Watch out for the flex routing setting, which lets Copilot process outside the EU at peak times.

What NinA builds runs in the EU: on the client's own infrastructure or on servers with a European provider. We work to ISO 27001 and NEN 7510.

Give your team a short list that fits on one A4 page. You can adopt and adapt these ten rules as they are. They work better when you explain them in a team meeting or training, not just in an email.

1. 01Only use the business account for work. No private accounts, not even 'just quickly'.
2. 02No passwords, citizen service numbers (BSN), bank details or medical data. Never, in any AI tool.
3. 03Take names out where you can. 'Customer A' works just as well as 'Jan de Vries'.
4. 04Don't paste whole files when a summary is enough. The less you share, the less can go wrong.
5. 05Check everything that goes out. AI can write convincing nonsense. You're responsible for what you send.
6. 06Check facts, figures and sources. Especially on legal, financial and medical topics.
7. 07Don't let AI make decisions about people. Rejecting an applicant or turning down a customer stays human work.
8. 08Don't use unknown AI tools or plug-ins. If you want a new tool, request it from the owner of the AI policy.
9. 09Report it when it goes wrong. Put something in an AI tool that doesn't belong there? Report it right away. Reporting is good, keeping quiet isn't.
10. 10Ask if you're in doubt. There's a fixed point of contact. Doubt is a perfectly good reason to pick up the phone.

Put these rules in your AI policy, together with the list of approved tools and who is responsible for what.

If someone has entered personal data into an AI tool that doesn't belong there, that may be a data breach. Stop, delete the conversation, report it internally to the privacy officer and assess whether you need to notify the Dutch Data Protection Authority and the people concerned.

If the breach poses a risk to the people concerned, the GDPR requires you to notify the AP within 72 hours of discovery (Article 33 GDPR). Even if you don't have to notify, you record the breach in your own register of data breaches.

This is how to tackle it:

1. 01Stop and record. Which data, whose, in which tool, with which account, when.
2. 02Delete the conversation. But don't assume that the data is immediately gone everywhere as a result.
3. 03Report it internally. To the privacy officer or whoever handles data breaches.
4. 04Assess the risk. Was it a business account with a data processing agreement, or a free private account? That makes a big difference.
5. 05Notify where necessary. The AP, and if the risk is high, the people concerned as well.
6. 06Learn from it. Why did it happen? Was a business tool missing, was the rule unknown, or was there time pressure? Adjust your agreements or training.

A culture in which people dare to report a mistake is worth more than any ban, however strict. Anyone who's afraid of punishment reports nothing.

If customer data runs through AI every day, for example when answering customer emails, don't let that depend on copying and pasting into a chat. Build a workflow for it that shields the data, runs in the EU, logs everything and has a human approve.

Copying and pasting into a chat is fine now and then. But when a team runs a hundred customer emails through ChatGPT every day, each of those hundred moments is a chance of a mistake. Then it's safer, and faster, to set up the process itself:

• Shielding data before it goes to the model. Names and numbers out, then back into the draft. As in the example at the top of this page.
• A model that runs in the EU, with a business agreement and without training on your data.
• Logging. What went in, what came out, who approved.
• A human in the loop. The AI prepares, the employee checks and sends.

At one of our enterprise clients, 80 percent of the 6,000 customer questions a month now have a prepared answer before the ticket opens. An employee checks and sends, always.

We build workflows like this within the AI Partnership. If you first want to teach your team to work safely with ChatGPT, a ChatGPT training on your own work is a quick first step. A half-day workshop costs € 4,500, excluding VAT.

This article is an explanation, not legal advice. If you work with special categories of personal data, have your approach reviewed by your privacy officer or lawyer.