How would you answer an interview scenario involving LLM observability and cost control?

For a rapidly growing AI feature, explain how you would detect rising cost per request and identify which prompts or user flows are responsible.