It can feel arbitrary to leave AI out of a decision it could technically handle. It is not arbitrary — it is about who bears the consequence, and whether a person can be held accountable for it.
Never let AI decide
Whether someone gets access, gets paid, gets a job, or gets refused a service. Anything involving a vulnerable person's safety or wellbeing. Anything with a legal or compliance consequence attached to being wrong.
Fine for AI to draft, never to finalise
A first pass at a policy, a reply, a summary or a plan. It can generate the draft. A person reads it, is accountable for it, and decides whether it goes out as written.
The plain test
If someone asked "who decided this, and can they explain why?" — the answer needs to be a name, not a model. If it can't be, the decision was made too far upstream of a person.