Most arguments about AI decision-making start too late. They ask whether a human made the final call: signed the order, approved the hire, denied the claim, filed the brief. But decisions begin before the visible act. They begin when the problem is framed, the evidence is filtered, the options are named, the default is set, and one response becomes easier to complete than all the others.
That is where AI is already powerful. A model can "only recommend" and still decide what the human gets to inspect. It can "only summarize" and still become the evidence record. It can "only draft" and still write the reason everyone later defends. The human may remain in the loop while the meaningful control has moved upstream.
The paper will argue for a different test: not whether AI made the decision, but what the AI output did to the decision chain. Sometimes the answer is support. Sometimes it is hidden substitution. Sometimes full automation is warranted. The danger is pretending those are the same thing.