Write the data rules before the first prompt

Most AI projects start with a demo. Mine start with a page of rules: what may reach a model, where it is processed, who approves the result, and what happens when the system doesn’t know.

The demo is the easy part. A model will happily summarise a tax notice, draft a reply to a client or sort a pile of claim photos. The hard part comes later, when someone asks where the client’s ID number went, who saw it, and who signed off on the reply. If nobody wrote that down before the first prompt, the answer is a shrug.

One hard line, for an accounting practice

For an accounting practice that meant one hard line: no name, CNP or IBAN goes to the model. The toolkit swaps them for placeholders first and puts them back only on the practice’s own computer.

To [Person-1], ID no. [ID-1]. You have not filed your 2025 income return.

The model still classifies the notice and finds the deadline. It never needs the person’s name to do that.

The four questions

1. What may reach a model, and in what form? Raw, redacted, or not at all.
2. Where is it processed, by whom, and for how long is it kept? Named in writing, per project.
3. Who approves each output before it leaves? A person, with a name.
4. What happens to the cases the system can’t handle? They go back to a person, with a reason.

The answers differ by project. A household app can send speech to an EU vendor and discard the audio; an accounting practice keeps its digital certificate on its own machine. What doesn’t change is that the answers exist before the code does.

Coming

A small assessment tool that asks you these questions about your own process before we talk, so the first call starts with answers.