What not to type into ChatGPT (or any AI) at work

The most common data leak in companies today does not come from a hacking attack: it comes from someone pasting an email, a contract or a client list into a chat window to have an AI summarise it. There is no malicious intent, and that is precisely why it is the hardest risk to control. It is also, almost literally, one of the four things Article 4 requires your staff to know: what data should never go into a tool.

Why it matters where your text goes

When you type something into a free or consumer AI service, that text leaves your company and enters a third party's infrastructure, usually outside the EU. Depending on the service's terms, that text may be used to train future models, stored indefinitely, or accessible to the provider's staff during review tasks. This is not hypothetical: several companies have already had fragments of conversations their staff believed private become publicly exposed.

Five categories that must never leave your company

• Personal data of customers or employees: names together with contact details, salaries, performance reviews, health data, or any information that identifies a specific person.
• Trade secrets and intellectual property: your own source code, formulas, pricing strategy, unpublished product plans, or anything giving a competitive edge.
• Information under contractual confidentiality: clauses from customer or supplier agreements, NDA terms, agreed commercial conditions.
• Credentials and access data: passwords, API keys, access tokens for internal systems, even just to have the AI «check» or format them.
• Unpublished internal documents: contract drafts, minutes of sensitive meetings, legal communications before formal sending.

The case that confuses people most: pasting a CV or a customer complaint

Asking an AI to draft a reply to a customer complaint sounds harmless, and it can be, if done right. The correct technique is to describe the type of case, not the specific case: the model does not need the name, the customer number or contact details to draft a good reply, only the situation. Personal data goes in afterwards, in your own system, over the text already generated. The same logic applies to a CV: you can ask for help assessing the profile of «a candidate with 5 years of B2B sales experience», never by pasting the full document with name, ID number and career history.

What about paid or enterprise tools?

Enterprise plans from the leading AI tools usually offer contractual guarantees that data is not used to train models and is processed with stronger security measures. That reduces the risk, but does not remove it: you still need to assess the provider as a processor under Article 28 GDPR, sign the corresponding contract, and check where data is processed. A paid plan does not automatically make the use compliant.

Frequently asked

How we cover this in training

Module 3 of our course is dedicated entirely to this: how data travels inside a prompt, real cases of trade secret leaks through public tools, and the three-step pseudonymisation technique for using AI without exposing personal data. Not abstract theory: the same examples as here, worked through as interactive challenges with real grading.