Answer
On IBM i specifically, this list includes anything touching QSECURITY level changes, user profile authority, exit point configuration, journaling and commitment control settings, and any automated recommendation to fail over a replicated system. It also covers financial postings, EDI transactions with trading partners, and changes to production RPG or COBOL objects promoted outside the normal change management process. None of these belong to a model acting alone, no matter how good its track record has been on lower-stakes tasks, because the cost of a wrong call in these areas is measured in downtime, compliance exposure, or real money leaving the business.
Buyers should push vendors past the marketing language and ask for the actual mechanics: does the tool log every recommendation it makes along with the human decision to accept, modify, or reject it, and can that log be pulled for an auditor without extra engineering work. Ask how overrides are captured, who gets notified when a high-risk recommendation is rejected, and whether the system can be configured to hard-stop certain categories of action until a named reviewer signs off. Vendors who cannot answer these questions specifically, or who treat the request as unusual, are signaling that their review model is thinner than the sales deck suggests.