AI features behave differently from normal software. The same input can give different output, answers can be wrong while sounding confident, and tasks can take seconds or minutes. UX for AI is mostly about helping people work with that uncertainty.
Start by setting expectations. Tell people what the feature is for, give example prompts, and be honest about limits. An empty text box with "Ask anything" invites requests the product cannot handle.
Show what is happening. AI responses often take longer than a normal request, so use clear states: listening, thinking, working on a step, streaming the answer, done. Streaming text as it is generated makes waits feel shorter. For agents that take several steps, show the steps and let people stop the run.
Keep people in control. Output should be easy to review, edit, regenerate, undo and reject. For actions with consequences, such as sending a message, spending money or deleting data, ask for confirmation first. Where accuracy matters, show sources so people can check.
Design for failure as carefully as for success. Plan screens for wrong answers, refusals, timeouts, rate limits and "I don't know". Collect feedback in a simple way, such as thumbs up or down with an optional comment.
AI products have also created new visual elements, such as animated orbs and waveforms that show when an assistant is listening, thinking or speaking. These work best when each state looks clearly different. ExodeUI, which we make, has a free gallery of AI voice orbs built as state-driven files your code can switch.