ChatGPT and GDPR: what can your business type into it?

May you use ChatGPT in your business?

Yes. The GDPR does not forbid a chatbot. It sets requirements for what you do with personal data, and those apply when you put it in a chatbot too. For text without personal data, such as a draft letter with no names, a summary of a public document or a piece of code, the GDPR plays no part. For health data, large files or if in doubt, a privacy lawyer or your data protection officer decides.

Take, for example, an accountancy firm whose staff want to use a chatbot to write emails and reports. What may go in, and what does the firm have to arrange?

What is personal data in a prompt?

Personal data is any information about an identified or identifiable natural person, says the Dutch Data Protection Authority (the AP), and that is wider than many people think (AP). In a prompt there can be, among other things:

• a name, address, phone number or photo;
• a customer’s email address, including a business address if one person is behind it, such as firstname@company.nl;
• a customer or staff number;
• a note about what someone bought or said;
• someone’s health, such as an allergy or a sick note.

Information about a company is not personal data. If it is a sole proprietorship or a small business whose data says something about the owner, it is.

What does the regulator say about chatbots?

On 6 August 2024 the AP wrote that it had received reports of data breaches in which employees had put personal data in an AI chatbot, for example medical data of patients at a GP practice and a file of customer addresses at a telecoms company (AP). The main points:

• If an employee puts personal data into a chatbot on their own initiative, against the agreements, that is a data breach.
• If your own policy allows the use, it is not a data breach, but according to the AP it is often not permitted.
• The companies behind the chatbots usually keep what you enter, on their servers, often without the person entering it or the person concerned knowing.
• You make clear agreements on whether employees may use chatbots and which data may and may not go in, and you can agree with the provider that it stores nothing.

On 13 July 2026 the AP published a tool for organisations that want to use generative AI. In it the AP advises checking first whether you can use generative AI without processing personal data (AP). If you can, the GDPR does not apply to that input.

What may go in? Three levels

Work with three levels. An employee then does not have to know the GDPR to know what they may do.

For green there is one reservation: confidential business information is a different matter, so set down what falls under it. Special categories of personal data, such as health, fall under stricter rules in the GDPR, and some data is so sensitive that a mistake does a lot of harm (AP).

What has to be arranged before you enter personal data?

If you put personal data in a chatbot, you engage another organisation to process it for you. The AP then calls you the controller and the provider the processor, and four requirements apply (AP):

1. You have a reliable provider: you must be confident that it protects the data properly.
2. You have a processing agreement, which states what the provider may and may not do with the data.
3. People can exercise their privacy rights, also for what you have outsourced.
4. There is a data-breach procedure: the provider informs you as soon as possible, and the notification to the AP remains your task.

Two more questions come on top, because with a chatbot they make the difference. What does the provider keep: is your input stored and used to improve the model? The AP names agreeing that the provider stores nothing as an option. And where is the data processed: if it is outside the EEA, a basis for that transfer is needed.

Whether a version offers a processing agreement depends on the provider and the subscription. Read the offer for it and ask if it is not there. If the provider becomes a recipient of personal data, that belongs in your privacy statement; what such a statement contains is in what must be on a website.

Would you rather keep personal data in your own system and send only text without names to a provider? Request a quote for an AI feature built that way.

A policy of one page

A policy nobody reads does not help. A one-page policy contains seven things:

1. Which tools are allowed: the name of the version and the account, not “ChatGPT”.
2. Who manages the account and who uses it.
3. The three levels, with examples from your own work.
4. What is never allowed: the red list.
5. Check what you get back: a chatbot can be convincingly wrong, and you answer for what you send.
6. What to do when it goes wrong: whom an employee tells, and within what time.
7. When you review it: with a new version, a new provider or an incident.

What do you do if an employee has already done it?

Assume a data breach and go through the AP’s steps (Datalek? Dit moet u doen, in Dutch):

1. First get an overview: what was entered, when, by whom, how many people.
2. Limit the consequences: have the provider delete the conversation if possible.
3. Decide whether you have to report it to the AP. That may have to be within 72 hours, counted from the moment you become aware of it. You sometimes also have to inform the people concerned.
4. Record it in your data-breach register, even if you do not report it.

An example: rewriting a client email

An employee of the accountancy firm wants an email about an overdue invoice to sound friendlier.

Not like this: “Rewrite this email to Jan de Vries, Dorpsstraat 12 in Zwolle, about his overdue invoice 2026-0412 of €1,250.”

Like this: “Rewrite this email to a client about an overdue invoice of €1,250, friendly and with a payment term of fourteen days.”

The employee puts the name, address and invoice number back in themselves. The result is the same email, and no personal data has gone to the provider. This takes less than a minute, and it is why the green row of the table is so large.

Five questions for a provider

Before you use a version for personal data, you want answers to these five questions:

If you get no answer to one of the five, the version belongs on your red list.

If you build AI into your own software, you largely set the answers to these questions yourself. Request a quote if you want to know what that costs for your task.

Frequently asked questions

AI in your own software, built by us

For a one-page policy and a business subscription you do not need us. If you want AI in your own software, for example to process documents or messages, we design what goes to a provider and what stays in your own system. We write that down, so that you can put it in your processing register and your privacy statement. We start with a test on your own examples, from €2,500. Request a quote and describe what the system has to do; you receive the hours, the schedule, the price and what is not included in writing.