So, where is evaluation? How often are the answers nonsensical? What fraction of common customer questions can it answer? How does LLM conpare on that?

Without those answerw, the article is meaningless.