What you may enter into ChatGPT: AI tools and personal data

The most common question in training is not how to write a better prompt, but whether you may paste a customer email into an AI tool at all. The answer is neither a ban nor “no problem” — it depends on the type of data and the version of the tool you use.

The starting point: this is processing of personal data

When you enter a name, address, email, phone number, employee data or the content of correspondence with a customer into an AI tool, you are processing personal data and the General Data Protection Regulation applies. That means you need a legal basis for the processing, the tool provider becomes a processor, and an appropriate agreement has to be in place with them.

Free and business versions are not the same thing

With consumer, free versions of AI tools the data can as a rule be used for further model development, and there is no processing agreement. Business and team versions offer a data processing agreement, exclusion from model training, control over data retention and access logging.

In practice: if there is a single case in the company where client data goes into AI, a business version is not a luxury but a precondition. The cost is on the order of a monthly subscription per user.

A rule that fits on one page

Freely. Publicly available texts, product descriptions, draft quotes without buyer details, internal notes without names, translations of technical documentation.

With a business version and a clear purpose. Client correspondence, contracts, quotes containing buyer details, the content of web form enquiries.

Never into public tools. Health data, national ID numbers and data from identity documents, salary and performance data, credentials, and documentation marked confidential under contract.

Anonymisation solves most cases

Most tasks do not need real data. Strip the name, address and contract number out of an enquiry and what remains is technical content the model handles just as well. Build a simple habit: before entering anything, replace names with labels like “CUSTOMER A”, then put them back in the final document.

When data must not leave the company

For documentation that must never go to an external provider, there are setups where the model runs inside your own infrastructure or in a European environment under contracted terms. That is more expensive and only makes sense when the volume of such material is large — in most companies that means a few processes, not the whole business.

What to put in the internal rules

Five points are enough: the list of approved tools, the three data categories above, mandatory review of output before it leaves the company, a ban on using private accounts for business tasks, and who to ask. Publish the rules, walk through them once with the team and record the date — the same document also serves as evidence of measures under the AI literacy obligation.

In practice, risk rarely comes from bad intent. It comes because nobody said what was allowed, so everybody decided for themselves.