People get so impressed by the things we can do with AI these days, they forget that there's a basically irreducible error rate.

So you can have a system which flags cases for human review and that's fine, but it's insanity to have fully automated decision making.