Evaluating AI system performance based on actual customer success metrics rather than task completion rates.