The moment a bot doesn't understand
No matter how well a conversational system is built, it will eventually misunderstand the user. Speech might be transcribed wrong, an intent might be misclassified, or a user might phrase something in a way the system just doesn't handle. What separates a chatbot people tolerate from one they actually trust often isn't whether this happens, it's how the system responds when it does.
A bad fallback repeats the same unhelpful 'I don't understand, please try again' regardless of how many times the user has already tried, or worse, confidently answers a question the system actually misread, giving the user wrong information with no signal that anything went wrong. A good fallback does two things: it clarifies specifically what the system didn't understand, rather than a generic error, and it offers a concrete way forward, like rephrasing, choosing from a short list of likely options, or escalating to a human.