How do you evaluate whether a customer problem is best solved with AI or with a simpler non-AI approach?