If we could get a highly accurate way of determining whether the AI has trouble answering a message and then logging that so we can monitor and add to the knowledge base that would be helpful