How should we measure whether an AI product is actually good?

This piece is being written. The question is one I'm actively working through in practice, and I'd rather publish it late than publish it thin.