How do you test an AI chatbot feature when its answers are different every time?

Our product team added an AI assistant that answers customer questions using a large language model. My usual approach of an exact expected result doesn't work, because the same question gets a slightly different answer each time. How do QA teams test features like this?