A user interface is the layer where a person meets the software. Buttons, forms, menus, a voice command flow and the message box in a chat window all belong to it. However good the system underneath is, this layer is what the user sees and touches.
In AI products the interface has a slightly different job than it used to. In conventional software the interface enumerated what the system could do, one button at a time. In a product built on a language model the input is free text, so the interface mostly has to hint at what you can even ask for. Example prompts, suggested follow-ups, editable answers and visible sources became part of the design for that reason.
The second difference is making uncertainty visible. The model is sometimes wrong, so a good interface stops presenting its answer as the single truth. It shows what the answer rests on, leaves a path to correct it, and keeps the action reversible.
A worked example: an assistant added to an accounting app sorts invoice lines into categories automatically. Rather than saving those categories outright, the interface marks them as suggestions the user confirms or changes with one click. That small difference keeps a bad classification from slipping quietly into the books.
Accessibility sits in this same layer. Without keyboard navigation, screen reader support and adequate colour contrast, the interface simply does not work for part of the audience.
