Generative UI, one year later

Eleven months ago we wrote about generative UI, the idea that AI should answer with interfaces instead of paragraphs. Google had just launched experiments in the Gemini app that built a custom page for a prompt, and OpenAI and Anthropic had proposed a shared standard for putting interactive tools inside chat. Our advice was to separate two things: models designing interfaces from scratch, which was impressive but not ready, and models choosing from components you had already built, which was ready to use.

A year on, both halves have moved, and the second has moved further.

The year in short

In December, Google released A2UI, an open format that lets an AI agent describe an interface as structured data and leaves the drawing to the app's own components. In January, MCP Apps became the first official extension to the Model Context Protocol, and Claude started showing interactive versions of Slack, Asana, Figma and others inside the conversation. OpenAI's ChatGPT became fully compatible with the same standard in February. In July, a major update to the MCP specification made extensions such as MCP Apps a formal part of the protocol.

At its developer day on 29 September, OpenAI went further: plug-ins in ChatGPT can now have a home in the sidebar and panels users work in during a chat, and OpenAI is adding automations triggered by events in a connected app, through a proposed extension to MCP. Then, last week, ChatGPT launched what OpenAI calls Intelligent UI. Answers now come with tappable buttons, interactive charts, editable graphs and calculators built for the question, for every user from Free upward.

For ChatGPT's users, interactive answers are now the default rather than an experiment.

What the year settled

The two patterns we described have both survived, and they have found different jobs.

For an AI assistant's own answers, such as a savings calculator, a diagram or a chart of what you asked about, the model now generates the visual itself. ChatGPT now does this for everyone, Google has been building the same into Gemini and Search, and the visuals are mostly about explaining things.

For actions that touch a business, such as booking, buying, approving or changing an account, the interface comes from the business. MCP Apps, A2UI and ChatGPT's plug-ins all work this way: the company provides the component, the assistant decides when to show it, and anything that changes data goes through the company's own checks and the user's confirmation.

That split is sensible, and it is likely to hold. Nobody wants a model improvising the screen where they confirm a payment.

What it did not settle

Accuracy is still the open question. A beautiful interactive chart with a wrong number in it is more convincing than a wrong sentence, and the new visuals make answers look more authoritative than ever. That puts more weight on where the data comes from. Live numbers should come from your systems through your own components.

Discovery is unsettled too. ChatGPT has a plug-in directory and better recommendations, Claude has its connector directory, and each assistant decides in its own way when to suggest a company's app. For businesses, being chosen inside an assistant is starting to look like a new kind of search ranking, with rules nobody outside the platforms fully knows.

What it means if you build a product

Your components are now your interface to AI. A clean, documented set of components for the key moments in your product, such as a booking card, a quote form, an order tracker or a comparison table, can now appear in your app, on your website and inside the major assistants. A messy component library cannot.

Build the server before the screen. The MCP server that exposes your actions and data is reusable across every assistant that supports the standard. The interactive view on top is the smaller part of the work.

Keep humans in the loop where it counts. Every action that spends money, changes a booking or shares personal data should require a clear confirmation in your own interface, whichever assistant it starts in.

A year ago, our advice was to start with the two or three moments where a text answer clearly falls short. That still holds. The difference is that those moments can now be served wherever your customers happen to be asking.