The phrase gets used loosely enough to be meaningless, so it is worth being specific about what it commits us to in the design.
Take any operational decision — approving something, flagging something, deciding whether an exception is real. Almost all of the elapsed time around that decision is not decision-making. It is finding the relevant documents, reconciling two systems that disagree, chasing the person who has not replied, and assembling the context into a form where the judgment can actually be exercised. That surrounding work is what we automate.
What reaches a person, then, is not a queue. It is a small number of items that genuinely need a human, each arriving with the reasoning and the evidence already attached. The reviewer is not reconstructing the situation; they are ruling on one.
The corollary is that the system must be able to say it does not know. A system that resolves everything is not exercising judgment on your behalf — it is concealing the cases where it should have asked. Refusing, holding, and escalating are features, and a system with no held items is a system that is guessing somewhere.
The other half of the commitment is that every decision the system does make is recorded and explainable. When an auditor, a board, or your own controller asks why something happened, there is an answer, and the answer is not that the model decided.