A quick test
Ask three questions. If the AI gets this wrong, how bad is it? Would we notice quickly? Is there a clear right answer? High harm, slow detection and no clear answer means the task should stay with a person, with AI at most preparing information.
Examples that usually fail the test: approving payments to new suppliers, giving legal or tax advice to clients, deciding who to hire or dismiss, and replying to a serious complaint without review.
Human in the loop is a design choice
Many tasks sit in the middle. The right setup is AI drafting and a person approving, with the approval step built into the workflow rather than left to good intentions. Over time, as you see the error rate, you can decide whether some steps can run without approval.
In every ForceMX department, automations that touch money, personal data or customers go live only after you approve them, and anything unusual is routed to a person.