Most of what professionals ask an AI model to do with a document, summarising, drafting a reply, comparing two versions, does not need the real names. It needs to know that the same person appears in paragraph 2 and paragraph 9.
Deleting names leaves a text the model cannot follow. Replacing each person, company, address and identifier with a stable label keeps the structure: [PERSON_1] signed the lease, [PERSON_2] is the guarantor, the deposit was paid from [IBAN_1]. The model can reason about who did what, and its answer comes back with the same labels.
The step that makes this usable is the return trip: the labels in the answer are swapped back to the real values on the user's side, so the result reads naturally and the model never saw the originals.
Replacing identifiers with labels while keeping a separate key that maps them back is what the GDPR calls pseudonymisation (Article 4(5)): data that can no longer be attributed to a person without additional information kept separately and protected. Pseudonymised data is still personal data, but it reduces what each recipient sees, which is the point of data minimisation and of security measures under Article 32.
In 2026 the European Data Protection Board put new guidelines on pseudonymisation and anonymisation out for public consultation; a summary is in this note on pseudonymisation and anonymisation after the EDPB guidelines.
Redactic does this automatically for documents and for chat with AI models, with processing in the EU. The same questions apply to any approach you build yourself.