Everybody wants to point AI at their customer data. Almost nobody does it safely. I built a scrubber that strips the sensitive parts before anything reaches the model, and can put them back only through a locked, audited door. Here is how.

Six parts, and the order is the safety.

1. Find the sensitive parts. Before any text goes to an AI, a pass finds names, account numbers, and other private details.
2. Swap them for safe tokens. Each sensitive value is replaced with a harmless placeholder, so the AI sees the shape of the data but never the secret.
3. Store the map somewhere else. The key that maps placeholders back to real values lives in a separate, locked vault, not with the scrubbed text.
4. Make restoring one-way. You can re-identify the data, but only through a separate path that logs every lookup.
5. Version the rules. What counts as sensitive changes over time, so the ruleset is versioned and can be updated without breaking old records.
6. Test that it holds. Adversarial tests try to sneak private data past the scrubber, so you know it actually works.

Anyone feeding customer text to AI, support, summaries, analysis, needs this floor first. An accounting practice, a clinic, any shop that holds records people trusted you with. The payoff is simple: use the AI, keep the private data out of it, and have a paper trail that proves it.

A browser tool that does the scrubbing part is on this workbench's build list. Until then, the blueprint above is the whole design, free to copy.

From the Lendari build. Everything described here is something I actually run; nothing on this bench is theoretical.