This paper describes a study of citizens’ chats with a chatbot of a public agency. We have analyzed chat logs and identified citizens’ lack of domain knowledge as a source of inadequate or failed chatbot responses. We identify three types of lack of domain knowledge: lack of the right vocabulary, uncertainty if a regulation fits the citizen’s situation or the “shape sorting box” problem, or citizen’s misunderstanding the regulations. The most serious failure is when a misunderstanding is not detected and corrected during the chat. The chatbot we studied is not able to make sense of badly formed questions from citizens. As implications for design we suggest making the chatbot limitations visible by not presenting it as a human-like avatar with a name. We also suggest to enable domain knowledge learning through its conversations.