What happens to what you type?

An AI chatbot such as ChatGPT, Copilot or Gemini does not run on your own computer. Your question, together with anything you paste or upload, goes to the supplier's servers, where it is processed and often stored for some time. Depending on the version and the settings, what you enter may also be used to train the model further.

Business versions usually promise that your input will not be used for training and that data is better protected. That is an important difference, but it is not a free pass: you are still giving information to an outside party. So check which tool your organisation has approved, and sign in with your work account. A business version and a personal version often have the same name.

Rule of thumb: if in doubt, do not type anything into a chatbot that you would not also put in an email to a supplier you do not know.

The traffic light: what can go in?

Red: never enter, not even in a business version

  • Personal data of customers, job applicants or colleagues, such as names, addresses, email addresses, phone numbers, customer or staff numbers, salaries, BSN (citizen service number) and medical or financial information
  • Passwords, access codes and API keys (codes that give software access to a system)
  • Confidential contracts, supplier quotes or legal documents
  • Financial figures that are not yet public
  • Source code that contains passwords or customer data

Exception for personal data and figures: AI built into software your organisation already uses for that data, such as an assistant in your email program or AI in the customer system or accounting package. It may work with that data, but only if your organisation has approved that AI feature, and only as agreed. A new AI button in software you already know is not automatically approved. A separate chat window or app does not count, even from the same supplier.

Amber: only if your organisation has agreed rules for this

  • Internal documents without personal data, such as a process description or your own code without passwords
  • Customer questions that you have properly anonymised first, so that nobody can be recognised (see below)
  • Drafts containing information that must not get out yet, such as a press release about a new product

Only do this in the tool your organisation has approved, signed in with your work account. Never in a personal account, not even one you pay for yourself.

Green: fine

  • Public information, such as text from your own website
  • General questions ("How do I write a friendly reminder for an unpaid invoice?")
  • Your own texts without confidential or personal information, to have them improved

Green information, too, only goes into a tool your organisation allows.

Anonymising is harder than it looks

Removing a name is not enough. "A 52-year-old cashier from our Zwolle branch who has been off sick since March" is instantly identifiable to anyone in your company. Combinations of details make someone recognisable. Replacing a name with "employee A" does not help either if the rest of the story is still recognisable.

So also remove job titles, locations, dates and specific details, or describe the situation in general terms: "How do I approach a conversation with an employee who is on long-term sick leave?"

Examples from everyday work

Customer service. You want help replying to an angry customer email. Do not copy the whole email with the name, order number and address. Summarise the complaint ("customer is angry about a late delivery and wants a refund") and ask for a reply to that. You then fill in the personal details yourself.

HR. Never paste a CV or a performance appraisal into a chatbot. Instead, ask for a general structure: "What questions should I ask in a performance review with a team leader?"

Marketing. Having text written for your website or social media is fine. But watch out for figures and claims: check every fact before you publish. That is what the next lesson is about.

Finance. Do not put real figures from your accounts into a chatbot. If you want a calculation explained, work with an example that uses made-up amounts.

Checklist: before you press Enter

  1. Am I working in the tool my organisation has approved, signed in with my work account?
  2. Does my question, or the file, screenshot or open page the AI can read, contain names or other identifiable details?
  3. Does it contain confidential business information?
  4. Can I ask my question in more general terms and still get what I need?
  5. Do I know what I am going to do with the answer, and who will check it?

In doubt? Then do not enter it, or make your question more general first. Has something gone wrong? Report it straight away to your organisation's privacy contact (if you don't know who that is, your manager).

Test yourself

Choose the best answer for each question. You will see straight away whether you are right.

1. A customer emails a complaint with their name, address and order number. What do you do?

Correct. This way you get a useful answer without customer data leaving your organisation.

Not right. The name, address and order number can all be traced back to the customer. Summarise the complaint without those details and fill them in yourself afterwards.

2. Your organisation has a business version of a chatbot. Does that mean you can enter anything?

That is right. A business version reduces the risk, but red information such as passwords should never go in, and your organisation's rules still apply.

Unfortunately not. A business version offers better protection, but it is not a free pass: red information never goes in. For everything else, you follow your organisation's rules; you do not need to ask permission each time.

3. Which question can you ask without worrying?

Good. A general question without any data is green. Do ask it in a tool your organisation allows.

No, that contains personal data or even an entire customer database. Ask the question in general terms.

4. "A 52-year-old employee from our Zwolle branch who has been off sick since March." Is that anonymous?

Exactly. Combinations of details make someone identifiable, even without a name.

Unfortunately not. It is the combination of details that makes someone recognisable, even without a name or age: colleagues will know straight away who it is about. Describe the situation in general terms.

5. You want to use an AI tool that your organisation has not (yet) approved. What do you do?

Good. That way your organisation knows which tools are being used, and that is part of AI literacy too.

Not right. Even a general question only goes into a tool your organisation allows, and never into a personal account. For a tool that has not been approved, your organisation has no agreements in place about your data. So ask first, not afterwards.

Summary

  • What you type, paste or upload into a chatbot leaves your organisation.
  • Red information (personal data, passwords, confidential documents, unpublished figures) never goes in, not even in a business version.
  • Removing a name is not the same as anonymising.
  • Only use the tools your organisation has approved, signed in with your work account.
  • Ask your question in the most general terms you can; fill in the details yourself.

Want to know more about the rules? Read what the AI Act requires or see our AI use policy template.

This information has been compiled with care, but it is not legal advice.