What Is Real-Time Generative UI? Why 2027 Is the Year Interfaces Started Building Themselves

For thirty years, the deal was fixed: designers and engineers decide what the interface is, and every user gets that interface. Generative UI breaks the deal. The interface stops being a thing you ship and becomes a thing that happens — assembled in the moment, for the person and the task in front of it.

That sounds like science fiction until you look at what’s already in production.

Where does generative UI actually exist today?

This isn’t a 2030 prediction. Shipped examples, right now:

• Google Gemini 3, Dynamic View (November 2025). Ask about the microbiome and you get a different interface as a five-year-old than as an adult — different content, different features, built on the fly. Google’s research paper (“Generative UI: LLMs are Effective UI Generators”) reports raters strongly preferred generated interfaces to standard output.

• Google A2UI (late 2025). An open spec that lets AI agents generate widgets — forms, sliders, charts — inline in a conversation instead of describing them in text. An agent that needs a date, time, and party size generates the form, exactly when needed.

• Salesforce Einstein dashboards. Layouts adapt to role and usage: a sales manager sees pipeline metrics first, a rep sees activity tracking. Same product, different interface per person.

• Adaptive workspaces. Tools like Motion, Glean, and Asana’s Dash reorganize the workspace around your current focus rather than waiting for you to assemble it.

• Khanmigo, Khan Academy’s tutor. The teaching interface adapts to where the student actually is — first encounter or review — rather than following a fixed sequence.

Different companies, different stacks, same direction: the interface is becoming an output, not an input.

How does it actually work under the hood?

The architectures converging in 2026–2027 share three layers:

1. A reasoning layer — the LLM interprets the user’s intent and context and decides what the interface should contain.

2. A UI specification layer — the model doesn’t emit raw HTML (a security and consistency nightmare). It emits structured descriptions: A2UI “component blueprints,” Open-JSON-UI specs, or tool calls mapped to a component registry.

3. A rendering layer — your application receives the spec and renders it with your real, branded, tested components. Protocols like AG-UI handle the streaming and state sync between agent and frontend.

That third layer is the part people miss when they panic about “AI designing whatever it wants.” In serious implementations, the AI selects and arranges from a governed set of components. It’s composing, not painting.

Why is this happening now, and not five years ago?

Three things converged. Models got good enough at code and layout to produce interfaces people prefer — the Google rater study is the milestone. Frameworks matured: CopilotKit, the Vercel AI SDK, assistant-ui, and A2UI turned what was a research demo into a stack a product team can actually build on. And the chat bottleneck became obvious — once agents could book trips and analyze data, returning a wall of text was visibly the weakest part of the experience. The reasoning outgrew the interface. Generative UI is the interface catching up.

What does this mean if you build products?

Not “replace your UI with AI.” Nobody serious is doing that — Google’s own implementation wraps generated interfaces in a fixed, predictable shell. What changes now is more specific:

• Your component library becomes infrastructure. Components that are well-documented, strongly typed, and state-described are what a generative layer composes from. Messy libraries can’t be generated from.

• Some screens will stop existing. The settings page, the filter panel, the report builder — anything whose job is “let the user describe what they want” is a candidate for generation.

• Consistency becomes a design decision, not a default. When the interface can vary per user, deciding what must never vary is the new core of product design.

We’ll spend the rest of this series on each of those. The short version: the teams that will win at generative UI aren’t the ones with the best models. They’re the ones with the best-defined components.

If you’re wondering what this shift means for a product you’re building right now, that’s a conversation worth having early — talk to our product design team →