The default interface for LLM features has become a chat window, and for most product jobs it is the wrong one. Chat puts the burden of knowing what to ask on the user, hides capabilities behind a blank box, and turns two-click tasks into negotiations. Here are the patterns that consistently work better, and where chat genuinely fits.
Inline actions beat destinations
Users want help where the task already is. “Summarize this thread” as a button on the thread outperforms a separate assistant tab that must be told which thread. The rule: attach intelligence to objects (this document, this transaction list, this form), not to a general-purpose oracle. Discoverability comes free, and prompt-writing skill stops being a requirement, which matters because most users never learn to prompt.
Structured inputs beat blank boxes
A blank text box is a usability dead end dressed as freedom. Chips, dropdowns, and templates (“Draft a follow-up: friendly / firm / short”) constrain requests to what the feature does well, communicate scope, and prevent the disappointing first attempt that kills adoption. Constraint here is the same kindness as good form design: the interface carries the cognitive load, not the user.
Design the generating state honestly
LLMs take seconds, and seconds feel long. Streaming output word by word beats a spinner because progress is visible and interruptible; skeleton previews set shape expectations; and a visible stop button respects the user who can already see the answer drifting. Never fake instant: motion and state exist to tell the truth about what the system is doing.
Output is a review surface, not an endpoint
The interaction is not ask-and-receive; it is ask, receive, adjust. Good LLM UX makes the adjustment loop first-class: edit the draft in place, regenerate a selected section rather than the whole, compare versions, and accept explicitly, because nothing should become record without consent. The accept action is also where quality signals live; instrument it.
Where chat actually fits
Conversation earns its place when the task is genuinely exploratory: debugging a question the user cannot name yet, multi-step refinement, support triage across a wide surface. Even then, seed it: suggested prompts grounded in the user’s context, answers that cite their sources, and a visible route to a human. Chat as the last resort surface, not the front door, and never as a patch on bad navigation.
The pattern behind the patterns
Every choice above moves effort from the user to the design: naming the capability, constraining the input, showing the state, structuring the review. That is ordinary product design discipline applied to a new material, which is why teams that design AI as a product surface rather than a bolt-on ship features people actually adopt.
Building an LLM feature and unsure which surface it deserves? hello@beconfidency.agency, we design and ship these end to end.
Designing AI this way is the whole point of our AI design and integration service.