The default move when you bolt an LLM onto a product is to add a chat box. Sometimes that's right. Usually it isn't.
The chat-box trap
A chat box is a blank page. Blank pages are intimidating, and they push the discovery cost onto the user. If your product had a clear set of jobs before the LLM showed up, those jobs are still there — the model is just a new tool for accomplishing them.
Three patterns I keep coming back to
- Inline transforms. The user has selected something. Offer an action. The model runs in place; the result lands where they were already looking.
- Drafts on demand. Pre-fill a form, a layout, a starting point. The user edits rather than authoring from scratch.
- Quiet reviewers. The model watches a workflow and surfaces a single, well-timed suggestion. No interruption, no chat history.
Each of these has the same shape: the model amplifies an existing intent rather than asking the user to invent a new one.
When chat is right
When the task genuinely is open-ended exploration. When the user is the expert and the model is the interlocutor. When the alternative is no UI at all. Otherwise, ask whether the chat box is doing real work or just signaling that you shipped something AI-shaped.