What data should you never put into an AI tool?

The framing that causes trouble is whether a tool is safe. That is a supplier question and it is partly answerable. The prior question is whether the data was yours to send, and for a great deal of what a small business handles the answer is no — not because the tool is untrustworthy but because sending it anywhere is a disclosure the business was not authorised to make.

Information held on someone else's behalf is the main category and it is broader than people assume. A client's documents. Patient information. Employee records. A supplier's or customer's confidential material received under an agreement. In each case the business is a custodian, and confidentiality obligations do not distinguish between a disclosure to a person and a disclosure to a service. In regulated professions the obligation is stricter still, and in some the fact of the relationship is itself protected.

Credentials and keys are the second category and the rule is simply never. Passwords, API keys, access tokens, anything that grants entry. They frequently arrive by accident — pasted inside a configuration file or a log while asking for help with something else — which is why the practical control is a habit of stripping them before pasting, rather than an intention not to include them.

The third is anything that would be embarrassing or damaging in a context you do not control: an unresolved dispute, a personnel matter, commercially sensitive plans, a draft of something whose existence is confidential. The test is not whether the supplier is likely to misuse it. It is whether you could explain, to the person concerned, having sent it to a third party — and if the answer requires a paragraph of justification, it belongs outside the tool.

What determines how much of this a business can safely do is the account arrangement rather than the model. Consumer accounts and business accounts typically differ in whether inputs may be used to improve a model, how long data is retained, and who at the supplier can access it — and those terms are published, change over time, and are the actual control surface. Establishing which arrangement you are on is a five-minute task most businesses have never performed, and it can move a substantial amount of work from prohibited to permitted.

The workable policy is short enough that people follow it. Never credentials. Nothing held on a client's or patient's behalf unless the agreement in place explicitly permits it. Anonymise before asking about a real situation — the tool rarely needs the name to answer the question. And when in doubt, ask the question without the document, which is usually possible and almost always sufficient.