Conversations aren't linear scripts
A naive conversational flow assumes the user will answer exactly what was asked, in order, without deviation. Real conversations don't work that way. A user might change the subject mid-task, correct something they said earlier, or ask a clarifying question before continuing. A good multi-turn design has to handle all of these without the whole conversation falling apart, and doing that well is a real design discipline in its own right, not simply a matter of the underlying model being smart enough.
Consider a bot in the middle of booking a flight that asks for a departure date, and the user responds with 'actually, can you tell me if you support international flights first?' A rigid flow that only expects a date in that slot will either misparse the question as a date or ignore it entirely. A well-designed flow recognizes the detour, answers the side question, and then returns to where it left off, resuming the original task instead of starting over or getting stuck.