How does a data team prevent poor data from poisoning AI when they have piles of raw and imperfect data?