Designing an AI product is unlike designing anything else, because two of its core materials are invisible. The first is text, the words the system says and the words it invites you to say. The second is time, the seconds a user spends waiting for a response that may or may not be worth it. Get those two right and an AI surface feels magical. Get them wrong and it feels like arguing with a slow, confident stranger.
Latency is a design material, not a bug
In most software, waiting is a flaw to be minimised. In AI products, waiting is intrinsic and constant, so it has to be designed rather than hidden. Streaming the response token by token is the single most important move, because it turns a blank pause into visible progress and lets users start reading before the system has finished thinking. Beyond that, the interface should signal what is happening, searching, reasoning, calling a tool, so the wait feels like work rather than a hang.
Show your work to earn trust
Users do not trust an AI answer that arrives from nowhere, and they are right not to. The surfaces that build trust show their reasoning: the sources a claim is based on, the steps an agent took, the confidence behind a suggestion. Citations are not a compliance feature, they are a trust feature, and they double as the user safety valve, a way to verify rather than simply believe.
The interface for an AI product is half text, half waiting. Both are materials you design, not problems you hide.
Prompts are product copy
The empty input box is the hardest screen in any AI product, because users genuinely do not know what to type. The best surfaces treat that moment as a writing problem: suggested prompts that teach what the system can do, placeholder text that models a good request, and examples that set expectations honestly. Every word the system says, and every word it suggests you say, is product copy, and it deserves the same care as a landing page headline.
Design for the wrong answer
AI systems are sometimes wrong, so the interface has to make being wrong cheap. That means easy ways to correct, regenerate, or edit a response, clear affordances to undo anything the system did, and a tone that invites the user to push back rather than accept. A surface designed only for the moments the AI is right will betray its users on the moments it is not.
These are the principles we bring to every conversational and copilot interface we ship, because the model is only half the product, the surface around it is the other half. If you are building an AI product and want the experience to feel as considered as the engineering underneath it, that is exactly the work our design and engineering teams do together.