Designing interfaces for AI agents
Chat is the default interface for a language model, and almost always the wrong starting point. Here's how we design surfaces where an AI agent can act, not just talk.
The default interface for a language model is a chat box, and it is almost always the wrong starting point for a product built around one. Chat is the interface every team reaches for first because it is the interface every team has seen — ChatGPT made it familiar, and familiarity feels like a safe design choice. It is safe for the team building the product and expensive for the user who has to use it.
Why chat is usually the wrong default
Chat externalizes effort onto the user in a way most product teams underestimate. A blank text box asks the user to know what to ask, how to phrase the request so the model understands it, and what to do with whatever answer comes back. That is three separate cognitive tasks stacked on top of whatever task the user actually came to accomplish, and it is why so many 'AI features' that are just a chat panel bolted onto an existing product see disappointing adoption despite genuinely capable models underneath.
None of this is the model's fault. A capable model wrapped in a blank chat box is being asked to do its job through the worst possible interface — one that requires the user to already have some idea what a good prompt looks like, which is precisely the expertise most users of a product do not have and should not need to develop.
Meet the user in the flow they're already in
Good AI interfaces do the opposite of what a chat box does: they meet the user inside the flow they were already in, rather than asking them to step outside it and describe their problem from scratch. A user editing a document does not want to open a side panel and explain what kind of rewrite they want in prose — they want to select a sentence and see a rewrite suggestion appear where the sentence already is. A user reviewing a support ticket does not want to ask a chatbot to summarize it — they want the summary already sitting above the ticket when they open it.
This is a design constraint as much as it is a technical one. It means the AI feature has to be scoped to a specific moment in an existing user journey, with a specific, narrow job to do, rather than positioned as a general-purpose assistant floating alongside the product. Narrow and well-placed beats general and bolted-on almost every time.
Find the smallest surface where the model can act
The practical design exercise is to look for the smallest possible surface where the model can act directly, rather than defaulting to the largest one. An inline suggestion the user can accept or dismiss with one keystroke. A rewrite button that shows a diff instead of replacing text silently. A smart default pre-filled into a form field, with an obvious way to override it. A one-click categorization that the user corrects rather than types from scratch.
Each of these surfaces has something a chat box does not: a built-in mechanism for the user to correct the model quickly, without composing a follow-up message explaining what went wrong. That correction loop is often more valuable to product quality over time than the model's raw accuracy, because it turns every use of the feature into a small piece of feedback the team can learn from.
When chat is genuinely the right answer
None of this means chat is always wrong. Chat earns its place when the space of things a user might want is genuinely open-ended — a research assistant fielding questions nobody could have anticipated, a support bot handling the long tail of issues too varied to enumerate as buttons, a coding agent whose next task depends entirely on what the previous one uncovered. In those cases, the openness of chat is the feature, not a limitation to design around.
The discipline is in recognizing which category a given feature actually falls into before defaulting to a chat box out of habit. Most AI features inside an existing product are narrow, repeatable tasks wearing a general-purpose interface that does not fit them. Design for the narrow task first, reach for chat only when the intent space is truly open, and the AI feature stops feeling like a novelty bolted onto the product and starts feeling like part of it.