Tip: test prompts with a made-up person
Work out whether a tool is any good using an invented case, not a real one. You learn the same thing without putting anyone's information anywhere.
Before you try an AI tool on real work, invent someone.
Write a realistic but entirely fictional case: a name that belongs to nobody on your books, an invented address, a plausible mix of needs, medication and events. Use that to test prompts, compare tools and train staff.
Why it works
Everything you actually want to learn during a trial — is the structure right, does it keep the important details, how often is it wrong, does it sound like us — you can learn from an invented case. The tool cannot tell the difference.
What you avoid is putting a real person’s information into a tool that has not been approved, may not have a data processing agreement, and on consumer plans may use what you type or let a human read it.
Doing it properly
- Invent, do not anonymise. Removing a name from a real record is not the same as inventing one. Rare conditions, unusual combinations of needs and small locations can identify someone with the name already gone.
- Make it realistic. A case that is too tidy will not surface the failures you care about. Include the awkward bits: a medication that sounds like another one, an unclear account of an incident, a detail that does not fit.
- Keep it. Reuse the same invented case across tools and over time, so you are comparing like with like.
- Label it. Mark it clearly as a test case so it can never be mistaken for a record.
Only once the tool has passed on invented cases, and only once you have a DPIA and an approved account, does it go anywhere near a real person’s information.