AI UX

AI UX is product design for features whose output varies and is sometimes wrong. Traditional interfaces assume the system does the same thing every time; an AI feature might produce a perfect draft, a mediocre one, or a confidently wrong one from the same input. Good AI UX designs for that whole range, not just the demo case.

It shows up in concrete decisions. Should this be a chat box, an inline suggestion, a button that fills a field, or a background job that tags records? Does output apply automatically, or appear as a proposal the user can accept, edit, or reject? How do you show where an answer came from, what the user sees while the model works, and what happens when it fails? Streaming, citations, undo, regenerate, suggested prompts, and feedback controls are the usual building blocks.

A frequent mistake is defaulting to chat because it's easy to build, which pushes the work of prompting onto users. Another is showing raw confidence percentages, which people can't act on. Match the interface to the task, make correcting the AI cheaper than doing the work by hand, and keep people in charge of consequential outcomes.