Redaction that keeps the meaning

Bad redaction removes the information the AI needs. Good redaction removes the information the AI does not need and keeps everything else. The difference is knowing what the task actually depends on.

For a reply to an email: the AI needs the tone, the ask, and the relationship (customer, supplier, colleague). It does not need the name, the address or the account number. For a summary of a report: it needs the structure and the findings. It rarely needs the client name or the exact figures. For a check of a document against a standard: it needs the content. It does not need who wrote it.

The technique: replace, do not delete. "Mrs Patel" becomes "the customer." "Account 4471" becomes "the account." "£48,312" becomes "roughly £50k" if the exact figure does not matter, or "[amount]" if it does not matter at all. The sentence still reads, and the AI still understands.

In practice. A solicitor's assistant used AI to summarise long correspondence for a partner. She built a two-minute routine: find and replace the client name with "the client," the other side with "the counterparty," and any reference number with "[ref]." The summaries were as useful as before. The client's name never left the firm's systems.

At your desk. Take a real amber document. Write down what the AI needs to do the task. Then redact everything else by replacement. Paste the redacted version and check the result is as useful as you needed. Save your replacement list; it becomes a reusable routine.

Write it down.

• What the AI actually needed:
• My replacement pairs:
• Was the result still useful: