Anonymizing is not the same as redacting
A redacted PDF is meant for people who should read the document without seeing certain details. For an AI model, an analysis or a training example, black bars are a problem: from “█████ owes █████ €2,400” nobody can tell who owes whom.
When you anonymize, each detected person gets a fixed number that stays the same throughout the document. “Mr Becker owes Ms Meier €2,400” becomes “[PERSON_1] owes [PERSON_2] €2,400”. A language model can still summarise or check the case, while real names never leave your organisation.