Most AI decisions get made backwards. Someone builds a thing that works in a demo, and the question of how much it should be trusted arrives afterwards, when the answer is inconvenient and there is already a sunk cost arguing for it.
This asks the question first. Seven answers, three gates, one verdict: whether the use case has earned the right to act on its own, whether it should propose and let a person decide, or whether it needs redesigning before anyone builds it.
The use case
Verdict
How the gates work
Three gates. None of them is about how good the model is, because that is not what decides this. They are about what happens when it is wrong, which it will be.
Data and lawful basis
Personal or special category data with no audit trail cannot be defended, whatever the model does. Special category data acting with no human review is worse. This gate blocks first because it is the one that is not yours to trade away.
Consequence and reversibility
Material harm that cannot be undone, or that nobody checks, does not get autonomy. The pairing matters more than either half: a reversible mistake somebody notices is a normal operational risk, while an irreversible one nobody notices is how an incident becomes a discovery six months later.
Evidence and grounding
Answers from the model's own knowledge are fine for drafting and unacceptable for anything that matters, because there is nothing to check them against. Anything reaching customers or a regulator without a reconstructable trace fails here too.
The three verdicts
- Redesign
- A gate blocks. This is not a scoring problem and no amount of prompt work fixes it. Change the design: narrow the data, add the review step, ground the answers.
- Assist
- The model proposes and a person decides. This is where most genuinely useful work lives, and it is a destination rather than a waiting room. Plenty of use cases should never leave it.
- Automate, with controls
- Every gate is clear and being wrong is survivable. Autonomy still comes with the controls listed in the verdict, because unattended and unmonitored are not the same thing.
What this deliberately does not do
It does not evaluate the model, and it does not care which one you picked. Model choice is the most discussed and least decisive variable here: swapping to a better model does not make an irreversible action reversible, and it does not create an audit trail that was never built.
It also will not tell you whether the use case is worth doing. That is a prioritisation question, and there is a separate tool for that.