They overlap, and one depends on the other, but they answer different questions and fail in different ways. What follows: where the line actually sits, what a good AI policy does, the mistakes I see most often, the pre-work worth doing before you write a word, and a model policy you can adapt.