Identify what can identify a person
Direct identifiers include names, addresses and account IDs. Indirect clues can include rare job roles, events, locations and combinations of dates. Free text and attachments often contain details that a column-level transform misses.
Preserve utility deliberately
Decide whether relationships between records need to survive. Consistent placeholders can preserve sequence, but they also allow linking records. Keep any re-identification key separate and tightly controlled; replacing names does not necessarily make information anonymous.
Test the output, not just the script
Review structured fields, free text, metadata, filenames, images and quoted replies. Test rare cases, long documents and examples that span multiple systems. Document remaining risks and access restrictions.
Use the local demo cautiously
The text-redaction demo on this site detects a narrow set of obvious email and phone patterns. It does not detect all personal data, guarantee anonymisation or provide legal clearance. Use synthetic text for learning, not confidential production archives.