Most data incidents involving AI at work are not dramatic. Somebody was busy, the document was long, the chat box was right there, and they pasted it. No malice, no hacking, no headline — just a file that now exists somewhere outside the walls your organisation has spent real money building.

So it is worth having the list in your head before you are in a hurry, rather than deciding case by case at four o'clock on a Friday.

The list

  • Anything with a named person's private details. Employee records, customer contact data, health or financial information, an HR complaint. Even one row of it.
  • Credentials of any kind. Passwords, API keys, tokens, connection strings, the shared login somebody parked in a spreadsheet.
  • Confidential commercial material. Draft contracts, pricing that is not public, anything to do with an acquisition, a board pack.
  • Source code your employer owns, unless the tool is one your organisation has specifically approved for it.
  • Anything you hold under an obligation to somebody else. A client's data is not yours to paste, whatever your own policy happens to allow.
  • Anything you would not put in an email to an external address. That is the practical test, and it is the one to fall back on when the cases above do not obviously cover you.

Why the free tier is the risky one

The enterprise version of a tool your company has bought usually arrives with a contract saying what happens to the text you send it. The consumer tier you signed up for personally often does not, and its default settings may allow your input to be used to improve the service. That is a policy difference rather than a technical one — and it is precisely the difference your legal team cares about.

Once text has left, it has left. There is no recall, and deleting the chat afterwards does not undo a transfer that already happened. Treat the paste itself as the irreversible step, because it is.