Insights · August 2026 · 5 minute read
AI in regulated work: a trust checklist before you buy.
If your work is regulated, the risk from a wrong AI output is not embarrassment. It is liability. A clinician, a solicitor, an accountant and a safety manager all carry a duty that a confident-sounding tool cannot discharge for them. Before you let AI near client-facing or compliance work, these are the six questions to ask.
1. What does it do when it is not sure?
A tool that always answers has been built to look useful, not to be safe. Ask to see it decline, hedge, or flag a low-confidence output. If it never does, it will present a wrong answer with the same certainty as a right one.
2. Who is the competent person, and where is the sign-off?
The tool should draft and prepare. A named, competent human should approve before anything reaches a client or a regulator. Ask where that gate sits in the workflow. If the output can go out without a sign-off, the tool has moved from augmenting judgement to replacing it, and the duty has not moved with it.
3. Can it show the source for each claim?
In regulated work, a claim without a source is a claim you cannot defend. A good tool cites what it used and marks what it could not verify. Be wary of anything that produces fluent references you cannot click through to a real document.
4. Where does our data go, and on what basis?
Ask where the data is processed, on what lawful basis, and whether special-category data is handled separately. Ask for a DPIA. If the answer is vague, assume the worst and check before you upload a single client record.
5. What happens when the model changes underneath us?
The model behind a tool is updated without your say-so. Ask how the vendor tests that an update did not change the answers you rely on. A serious tool has an evaluation set it re-runs. A weaker one finds out from a complaint.
6. Can we get our data and outputs out?
Ownership is a compliance question, not just a commercial one. Ask for export of your data and the tool's outputs, in a format you can keep. If leaving is expensive, you do not own your record.
The short version
A tool that answers these six plainly is one you can put in front of a regulator. One that cannot, is not. The competent person stays in charge, and the tool earns its place by making that person faster, not by pretending to be them.