Blog · Data & AI · · 1 min read

Anonymize an Excel export before AI

It's 5 p.m. on a Friday. Someone asks for "a quick look" at the client book. So you export the database to .xlsx, open ChatGPT, and paste. The idea isn't unreasonable. The volume is. An export isn't a document at all, it's three thousand real people leaving in one go for a third-party server, names, emails and IBAN included.

Client
Email
IBAN
Status
1
[ORGANISATION_1]
[EMAIL_1]
[IBAN_1]
Active
2
[ORGANISATION_2]
[EMAIL_2]
[IBAN_2]
Active
3
[ORGANISATION_3]
[EMAIL_3]
[IBAN_3]
Closed

Anonymized export: identifying columns become tokens, the values useful for analysis stay.

One export, hundreds of people at once

Paste a document, you expose one person. Paste an export, you expose a whole column. It's a matter of scale: where a single letter holds one name, your file lines up a thousand of them, row after row, each one escorted by its matching email and IBAN. Nothing accidental here. The leak runs into the thousands from the very first copy-paste. Here are the columns to neutralize before the AI sees them:

  • Name columns: client, contact, employee.
  • Emails, phone numbers and postal addresses.
  • IBAN and identifiers (SIREN, staff number).
  • Amounts that single someone out, which you can also bucket into ranges.

The real trap is what the spreadsheet hides

An .xlsx is never just the grid in front of you. Far from it. A forgotten second tab, rows filtered out but still sitting in the file, a hidden column, a formula quietly pointing at a "HR Sources" sheet: all of it travels with the workbook. You think you're sending three columns. You're sending fifteen. Hence the rule: you neutralize the whole workbook, not only the active view.

Mask the column, keep the number

The value of an export? The trends. Not the names. So you can pseudonymize the identifying columns while leaving fully intact the statuses, categories and amounts the AI actually needs to crunch the numbers. The table stays analyzable row by row. And every pseudonym is reversible: once the analysis comes back, you line your real clients up against the results in a single move.

Safe-Doc handles spreadsheets. Excel, CSV: just like a PDF or a Word file. It walks through every tab, masks the sensitive columns and keeps the original structure. Details on the Safe-Doc for accounting page.

Test it on a real export. Reversible pseudonymization. Every tab checked, the structure left fully intact.

Part of the guide : Use cases ↗