What should I remove before pasting into AI?

Watch the briefing

Watch the full video on YouTube

Full briefing transcript

You want help writing a reply to a complaint, but the message includes a customer's name, home address, and order number. Do those details help with the wording? Usually the drafting task needs the problem and requested remedy, not the person's identity. Start by narrowing what the draft must do, then remove information that does not serve that purpose. Less information is not automatically anonymous information.

Here is our entirely fictional complaint. Mira Ellis, at eighteen Willow Lane, unit four, reports that order Q X eight four seven two nine arrived on September fourteenth with a cracked desk lamp shade. She asks for a replacement. The message also says she is the only overnight baker in Ash Ferry and can collect after a three forty shift. Every person, place, and identifier here is invented.

First remove the name, address, and unique order identifier. Now inspect what remains. The rare occupation, small location, and exact shift time may still point toward one person when combined with other information. Simply replacing a name with Customer A does not erase those clues. Our example therefore removes them too. We are minimizing a drafting input, not certifying that a real complaint has been fully anonymized.

The revised input is short: a desk lamp arrived with a cracked shade; the customer requests a replacement shade. Ask for a courteous acknowledgment, with no invented collection arrangement, eligibility decision, or replacement promise. Add do not send anything. That preserves what the writer needs while excluding the details that belong in an authorized customer service system. A real service team can match the reviewed draft to its records later.

Our illustrative output says: thank you for reporting the cracked shade on your desk lamp. We have noted your request for a replacement shade. Our team will review the request and confirm the available next steps. We wrote this output for teaching; it is not a recorded model result. Check that it acknowledges the complaint without approving a replacement. The person handling the case still owns the factual decision.

Before pasting your own material, ask whether each detail is necessary, whether combinations could identify someone, and whether the destination is authorized for that information. Inspect attached filenames, comments, headers, and hidden metadata as well as visible text. NIST identifies privacy and de-anonymization risks; our checklist is an editorial adaptation, not a certified procedure. If safe removal is uncertain, use a fully invented example or an approved internal workflow.

Share only necessary information; anonymization can be incomplete. In our complaint, the damaged shade and requested remedy survive, but identity and unusual personal clues do not. That makes the draft easier to inspect without treating redaction as permission to upload anything. The next briefing separates a training setting from retention and workplace authorization, because those are different questions.