Chatbots are increasingly used for customer support, promising quick and efficient assistance. But how can we ensure these AI helpers are providing accurate information, not just friendly chatter? Researchers at Dutch software company AFAS tackled this very question, developing a method to assess the 'truthfulness' of their LLM-powered chatbot. Since the chatbot assists with complex software questions, defining a 'right' answer isn’t straightforward. The team’s research focused on creating a system that mimics how human support staff evaluate chatbot responses. They analyzed hundreds of chatbot answers, observing how support staff identified and corrected inaccuracies. This led to the development of a decision tree – a visual representation of the human thought process when assessing a chatbot's truthfulness. This tree then informed the creation of specific metrics that could automatically score responses. Interestingly, the research revealed that the type of question dramatically influences how mistakes manifest. For example, instructions and yes/no questions showed different error patterns. This insight allowed the team to tailor their automated scoring system for different query types. While still under development, this research offers a promising path toward more reliable chatbot interactions. By automating the evaluation process, companies could significantly reduce the time spent by human staff verifying chatbot responses, leading to faster customer support and improved experiences. However, challenges remain. Nuances like context-specific jargon and understanding subtle inaccuracies still pose difficulties for automated systems. The next stage of research will explore how to incorporate a knowledge base and machine learning to tackle these complexities, paving the way for truly trustworthy AI assistants.