When measuring the success of an AI feature in a software product, how do you choose a primary metric that reflects user value rather than model accuracy? Can you share one case where changing the metric changed a roadmap decision?