I like number of "great catches" as a measure of AI code review effectiveness