An enterprise is concerned about the cost of using a large LLM for a high-volume customer support chatbot. They want to reduce token consumption while maintaining response quality. Which strategy would be MOST effective?