Generic metrics, such as “hallucination” or “toxicity,” often miss domain-specific issues. Here's what the best AI teams track instead and how they leverage error analysis.