Generative UI is a way of building software in which an AI model puts together a live, interactive screen at the moment a user asks something, instead of answering in a chat window with text. UI stands for user interface: the screens, buttons, tables and forms through which a person uses an app. In an ordinary app, designers and developers build every screen in advance. With generative UI, the screen for a question is put together only when the question is asked.
Here is how it works. The app sends the user's question, and the data that this user is allowed to see, to a language model. The model does not draw the screen itself. It writes a short, structured description of a screen, usually in JSON, a standard text format for data. The description names the parts to use and what goes in each: for example, "a bar chart of these numbers, then a table, then a button labelled Approve". The app's own code reads this description and builds the screen from its UI component library, the set of ready-made charts, tables, forms and buttons that its developers already use. So the generated screen looks like the rest of the app.
Consider an HR software company in Hyderabad whose product is used by managers in hundreds of companies. A manager types into the product's assistant: "Whose leave requests are waiting for my approval this month?" A chatbot would reply with a paragraph. With generative UI, the manager sees a small table of three employees, with their dates and an Approve button beside each. When she presses Approve, the button calls the app's own API, the fixed way in which one program asks another to do something. This is the same request that the normal leave screen sends. So the usual permission checks run, exactly as they would on that screen.
What can go wrong? A model can describe a screen with wrong numbers or a misleading chart, the same kind of error as an AI hallucination. So well-built apps take the numbers from their own database, not from the model's text, and let the model choose only the layout. The model must never decide who may see what; the app's code decides that. Testing is also harder, because the same question can produce a slightly different screen each time. This problem is described in testing non-deterministic software.
How is generative UI different from an AI app builder? An app builder writes code once, and the app's screens then stay fixed until someone changes them. Generative UI writes no code that is kept. It puts together a new screen for each question and each user, and the screen is discarded afterwards. The companies that sell it give developers a toolkit that connects their app to a model and to their component library. The models used are general foundation models.