If models can assemble interfaces on demand, the screen becomes cheap. The scarce asset is permission to make the screen consequential. A generated button may look identical to a designed button while carrying very different authority. Once software can improvise the view, a business needs to decide who controls what the view is allowed to do.
A design system can answer that question, but only if its job expands beyond visual consistency. It becomes commercial policy expressed in software: which information a component may reveal, which business process it may start and when a human must confirm the result.
The AI Daily Brief reports that OpenAI's “intelligent UI” renders interactive visualisations, charts and images instead of walls of text. That is an attributed product report, not independent evidence that dependable interfaces can now be generated for every workflow. Even if the capability becomes broadly reliable, arranging a useful chart will be the easy part. The harder question is whether a business can vary the presentation without weakening control over commitments and access.
OpenAI's plugin documentation describes custom interface elements as optional and recommends them when people need to inspect, compare, edit, confirm or navigate structured information. Those verbs conceal very different commercial consequences. Comparing two products is not equivalent to confirming a payment, even when both appear as cards inside the same conversation. A governed design system would encode that difference before the model chooses a presentation.
It should tell the model that a comparison card can display permitted supplier data, while an approval control carries a separate rule about who may act. This turns a component from a reusable piece of design into a compact agreement between product, operations and risk.
Consider a hypothetical procurement assistant that assembles a comparison table and offers an “Approve supplier” button. The model may choose how to organise the permitted information. The button itself should invoke a server-side rule that checks the request's current status before applying the user's delegated spending authority. If the rule rejects the action, visual polish cannot make the decision valid.
The point is larger than familiar advice about keeping security on the server. It is an allocation of commercial responsibility. OpenAI's documentation says business data is the source of truth and should not be stored only in the interface. Its ChatKit guidance also requires the application server to authenticate users and pass a unique identifier when creating a session. These boundaries provide a useful base, but a business still has to decide what each authenticated person may commit the company to do.
This architecture changes the economics of product change. A new layout can move quickly when its permitted behaviour is already defined. Risk review can focus on a changed interaction rule rather than reopening an entire journey because a card moved or a chart appeared. When an interaction causes loss, the business can examine the rule that authorised it instead of arguing vaguely about whether the model or the interface was responsible.
Vendor accountability also becomes clearer. A supplier demonstrating attractive generated screens should have to explain how a consequential control is connected to the customer's permissions and official records. The buyer should be able to change the model or presentation layer without rewriting the rules that determine whether an action is valid. If those rules exist only inside a vendor's prompt or generated code, the apparent flexibility creates a new concentration of operational risk.
Governed generation is not free. Maintaining approved components and testing their behaviour adds work that a fixed screen may not justify. A stable interface is the better choice when the task barely varies or when predictability matters more than avoiding navigation. Generated presentation earns its additional cost when users arrive with materially different questions and the business can confine the variation to low-consequence choices about how information is shown.
That condition should shape measurement. A pretty interface proves little if a user misunderstands whether an action is a draft, a request or a binding commitment. A useful evaluation asks whether the person understood the consequence and whether the official system recorded the intended result. Presentation quality matters because it affects those outcomes, not because generation deserves a novelty metric.
For an Australian company exploring this capability, the practical starting point is one workflow with a real commercial consequence. Map each visible control to the rule that gives it meaning. Record who owns that rule and what happens when it fails. Then ask whether a model can vary the surrounding presentation without changing the decision itself.
If no owner can answer those questions, generated interfaces are premature. On-demand presentation could make visual composition abundant while making disciplined constraints more valuable. If the capability improves, the durable advantage will belong to businesses that can release new experiences without improvising authority. Let the model design the screen only after the company has designed the rules.
sources
Prepared with assistance from a Useful Engine agent workflow. Factual and editorial checks are part of the publishing process.