On someone else's screen: what a brand still holds when an AI assistant draws the interface
With A2UI 1.0, Google wants AI assistants to answer with screens instead of text. For brands, something fundamental shifts: the design largely goes to whoever draws the screen, while the building blocks, the data and the moment of confirmation stay with the company. How much visual control remains depends on the route, and for most SMEs this is, today, something to prepare for.
An imaginary print shop has been supplying brochures to a handful of regular business clients for years. One day in the near future, a marketing employee at one of those clients asks her employer's AI assistant to repeat last spring's order. The answer is a screen showing the print run from back then, a choice of three paper types, a delivery date and a button to confirm. The prices are correct and the order only goes out after the click. Of the print shop as it once had itself designed, little is visible: colours, typeface and buttons come from the assistant. The scene is made up and lies in the future for now. The question it raises has been on the table since Google's announcement on 11 November: what happens to a brand when an AI assistant answers with a screen instead of text?
An order on someone else's screen
What Google launched
That day Google released A2UI 1.0, according to the company the first stable version of an open protocol for generative interfaces. That means a screen an AI assistant puts together on the spot, such as a form to confirm something or a chart, as an alternative to an answer in running text. The division of labour is simple: the agent describes what should be shown, and the application that displays the answer draws it with its own components. According to Google, agents only place components the developer registered in advance, and the announcement promises that agents use “your design system and nothing else”.
The question is whose design system that is. The version 1.0 specification gives a sober answer. Theme and primary colour have been taken out of the protocol, so an agent can no longer impose a colour that way, and the design follows the theme of the application that displays the screen. The design system in the promise is therefore that of whoever draws the screen. Whether that is the brand itself depends on the route by which the customer comes in.
Who describes and who draws
Three routes, three degrees of control
The first route is the brand's own app or website. According to the project site, there are renderers for several frameworks on the web and on mobile devices, and anyone who uses such a renderer in their own application chooses both the components and their design. For the house style, little changes there.
It is different when the assistant itself draws. For Gemini Enterprise, Google's assistant for business, the documentation describes that A2UI agents appear with Gemini's built-in styling and with components from a fixed component gallery. The brand supplies content and structure there, and Gemini decides what it looks like. Whether a company can bring in its own component library is not mentioned anywhere in that documentation, and it also still refers to earlier versions of the protocol.
Routes 1 and 2: who draws the screen
The third route runs through MCP Apps, since January 2026 an official extension of the Model Context Protocol, developed together with OpenAI. A company brings its own interface along, which runs inside the assistant in a sandboxed frame. Claude shows interactive apps this way from Asana, Figma and Canva among others, and according to the protocol's documentation, as it stood in November, it also works in ChatGPT and Microsoft 365 Copilot. According to Google, it has also been possible since October in Gemini Enterprise, which thus turns up in two routes, because the technical path by which a brand comes in determines the control, regardless of which assistant shows the screen. A2UI fits into that route: Google described together with the makers of MCP Apps how an app can bring its own renderer and component library inside that frame.
Here the strongest objection is right. A brand that chooses this route loses little and gains a channel, with ample visual freedom. That freedom does have edges. The host may pass on theme settings and ask permission for what the app does, and in its design guidelines OpenAI explicitly restricts the use of custom colours and fonts. So the brand paints for itself, in a frame the host chooses.
Route 3: painting for yourself in the host's frame
“A brand that exists only in colours disappears as soon as someone else chooses the paint. What remains is the way it lets a customer decide.”
What the brand still holds
Across the three routes, three things stay with the company. The first is the building blocks: which components a brand offers determines which actions an agent can propose. The second is the data that ends up in those components, from prices to delivery times. The third, and the weightiest for trust, is the moment a customer confirms something.
The announcement includes an example scenario in which nothing goes out before the user agrees. Google itself presents that scenario as an illustration, and anyone who reads the specification sees a design choice that the protocol makes possible. The protocol supplies the building blocks for it: some functions, such as opening a link, only work after an action by the user, actions start from what the user clicks, and MCP Apps lets a host ask for permission without requiring it. Whether an order only goes out after a deliberate click is decided by whoever builds the agent. Something similar applies to the promise that sensitive data never reaches the model. Google phrases it broadly, while the project's documentation shows that it depends on how the developer links the data on the server side.
Three things stay with the company
That makes the moment of confirmation brand work in the literal sense. On screens that look different with every assistant, the way a company lets a customer choose, check and agree is one of the few things that can remain recognisable everywhere. Nielsen Norman Group warned back in 2024 that interfaces that are assembled anew each time cost users predictability and bring privacy risks. A company that stays careful and predictable at that point wins back something of what it gives up in design. A brand book that stops at colours and fonts says remarkably little about that moment.
The moment of confirmation as brand work
An honest note of caution
How far this is from the daily practice of an SME deserves a sober look. That people use AI assistants is certain: according to Eurostat, in 2025 42 per cent of Belgians aged 16 to 74 had used generative AI tools in the three months before the survey. That figure still says little about contact with companies, and what is known about that points mainly to text and links. Adobe sees traffic from AI assistants to American retail sites growing strongly, measured as click-throughs to the retailer's own site. For interactive brand screens in assistants, no public usage figures are known as yet, let alone for SMEs, and the early examples come from software companies with a business audience.
The list of companies that, according to Google, are already building with A2UI also calls for caution. ByteDance confirms it itself via Lynx, while for the other names the list rests on the announcement for now. On top of that comes the question of which standard will prevail. Besides A2UI and MCP Apps there are other approaches, and according to InfoQ the developer community voices criticism of an overly restrictive component catalogue, of security and of that fragmentation. Whoever builds today runs the risk of investing in a standard that ends up in the margins a few years from now.
What is certain and what is not yet
What an SME can already do now
What follows is a line of reasoning based on what is certain. As long as customers mainly meet a company through an assistant in text and links, preparation pays off most in clear content and reliable data: prices, terms, delivery times and product information that an assistant can take over correctly. A screen that is drawn later relies on exactly the same data, so that work pays off twice.
In addition, a brand can put on paper which building blocks are typical of its way of working, at which moments a customer must always decide for themselves, and which data never belongs with a language model. Such a document costs little, survives a change of standard and gives developers a clear starting point later. Investing in interactive screens of its own can wait until it is clearer which standard remains and whether customers in its own sector actually use such screens.
Prepare now, build later
How much remains
In the imaginary print shop, the customer would hardly recognise the house style on her assistant's screen, and yet there is a lot of the company in it: the paper types it offers, the prices that are correct, the delivery date that is met and the confirmation that comes at the right moment. On someone else's screen, less of a brand remains than a design agency would wish, and more than the vanished colours suggest. What remains is what the company always promised, only now it has to keep that promise in an environment it did not design itself.