This is the story of A2UI in Gemini Enterprise: agents that answer with an interface you can use, when a request calls for one.
AI answers in paragraphs. But when you need to choose, compare, look at numbers or do something, a paragraph is the slow way. You need something to use, not more to read.
And every team asks differently. The research here was a collaboration: we interviewed most of Google Cloud's biggest clients, and they brought what their own in-house research teams had learned. They told us what they needed and wanted; we came back with the best ways to show that information. A legal team, a finance team, an events team, a hotel, a store: each needed the answer in a different shape. That was the part I was best at: reading a need and turning it into something you can see.
We're visual creatures, and some things are better shown than told. Nielsen Norman Group has measured it for decades: most people scan a page rather than read it, and on an average visit they get through a quarter of the words at most. The same text, just laid out to be scanned, tested almost half again as usable. So the case was engagement. With the computing power we have now, an agent can draw the answer, not just write it: a form, a table, a chart, a button. That's what A2UI makes possible, and safely. It turns the chatbot you read into a workspace you use. And there's a second case: an answer with interface can bring in other Google products. Suddenly the answer has a map, a calendar invite, an email or a call in it. People stay in the ecosystem, and every answer makes requests to more services. It's cross-selling, and for the person asking, a more seamless experience.
This came naturally, after four years at Search for commerce making custom solutions for clients of every size and industry. The ideas came from the clients themselves, and from Google product managers who brought requests of their own. We usually worked from scripts and customer journeys, so nobody told me how the interface should look. I was the one who drew it, request after request, until we finally trained the AI to do it by itself.
So the idea: ask to book something, and the agent answers with the options, a time picker and a Book button. Interface, streamed in, instead of a paragraph about it.
Under the hood, each request is read first: what the person wants, what they mention, and what decisions they'll need to make.
Then the answer is built from a catalog of components the app has approved. The agent sends data, not code, and can only ask for what's in the catalog, so whatever it shows is safe and looks like it belongs there.
The vision isn't widgets everywhere. With no choice, comparison, data or action involved, the answer stays plain text. Interface when it helps, a sentence when that's enough.
It started with the canvas: a side panel that opens with the deliverable on the right, and you can still use it that way today. Then, little by little, we shipped pieces of interface that appear right in the middle of the conversation. First quick calls to action, data tables and simple forms. Later, maps and data visualization.
We showed it with six requests and six kinds of answer, each traced back to the request it answers. In December 2025 Google opened A2UI to everyone, and it's being built into Gemini Enterprise so agents there can answer with interface.
A question I give my students from this one: what should an answer look like? Sometimes a table, sometimes a button, and sometimes just a sentence. Thanks for watching.