ok so every single team i've worked with reaches for a chat box the SECOND "ai" gets mentioned in a brief. like immediately. and look i get it, its familiar, theres a million tutorials for it, but it lowkey pushes every single problem into the same shape whether it fits or not 😅
chat is genuinely good when the task is actually conversational, like the user doesnt even know what to ask for yet and the whole point is exploring back and forth. but its a REALLY bad fit when the task is basically a form or a lookup or transforming some data. in those cases a structured result you can just look at gets people to their answer way faster with less confusion tbh.
honestly the best interfaces i've used (and built ngl) combine both. conversational entry point to frame the problem, then a structured editable output once the model actually has context. streaming text is nice for showing the thing is "thinking" but its a bad substitute for something you can scan or sort or actually DO something with.
one more thing i keep telling people and they dont beleive me until they see it themselves: latency is a design material. not just an engineering problem you throw at the backend team. where the model pauses, what shows up while its pausing, how fast a partial answer starts feeling trustworthy... all of that shapes how the tool feels, sometimes more than how accurate the model actually is lol
