How would you approach evaluating the performance of an AI agent conducting voice interviews? What metrics would you consider important?