AI Product Design · 5 min read
The Right-Sized Interface
The best predictive experience offers the next useful step, keeps the person in charge, and gets out of the way when the task is done.
Most AI features today make you choose between two awkward options. You get a chat box that waits for you to ask the right question, or an assistant that appears uninvited and takes over half the screen. Predictive design lives in the space between: notice what might help, offer a small next step, and let the person decide.
Prediction is an offer, not an action
Google described Pixel 10 Magic Cue in an August 29, 2025 post as a feature that “intelligently anticipates your needs, linking information across your apps to offer help right when you need it.” For example, while someone is on a call with an airline to change a seat, Magic Cue brings up flight details from Gmail. When someone asks about dinner details, it pulls the time and place from Calendar or a confirmation email and offers them as a suggestion to share.
The important word is “offer.” The feature doesn’t decide to share the dinner details. It makes a useful next step available, and the person can accept it or ignore it. That’s the pattern worth borrowing: a small suggestion at the moment it may help, not an action taken on someone’s behalf without permission.
Google says Magic Cue runs on-device with Gemini Nano. On-device means the work runs on the person’s device rather than elsewhere. People can turn the feature off at any time and manage which apps it can access. Those controls matter because a prediction depends on information. People should be able to see and control what the system can use to make one.
A good offer does two things. It reduces effort, and it leaves the decision where it belongs. If the person has to ask for every useful detail, the system isn’t anticipating much. If it acts before they’ve agreed, it’s anticipating too much.
Right-size the surface
The size of an interface is part of its behavior. A small suggestion makes a different promise from a panel that takes over the screen. OpenAI’s UI guidelines for apps inside ChatGPT define display modes by the job: inline cards for single-purpose actions such as confirming a booking, with at most two actions; fullscreen for rich, multi-step work or detailed browsing; and picture-in-picture for ongoing sessions such as games or video.
That’s a useful size ladder. Start with the smallest surface that can complete the job, like a chip or card. Grow to a panel or full screen only when the task needs more room. Once the task is done, collapse back.
Every step up should be the person’s choice or clearly justified by the work. A card that asks for one decision doesn’t need to grow into a full-screen experience. A multi-step task may need more space, but the interface should make that transition understandable. Bigger isn’t automatically better. Sometimes bigger is just louder.
The surface should fit the task, not the ambition of the feature. If a tiny offer is enough, let it be enough.
Progressive UI: reveal as the task unfolds
Progressive disclosure is “a design pattern that hides and shows information based on user interaction,” according to GitHub’s Primer design system. Primer advises using it sparingly, and only when content truly needs to be truncated. Hiding content by default isn’t automatically clearer. It helps when the person can reveal what they need as they go.
For AI features, the interface can grow step by step with the task. Start with the offer. If the person wants to continue, reveal the next piece of work. If the task becomes more involved, show more of the interface. Use known components from the design system rather than inventing a new layout from scratch each time. That keeps the experience fast, consistent, and accessible.
Generative UI means an AI model generates not just content, but an entire user experience. On November 18, 2025, Google Research described this approach in experiments in the Gemini app (“dynamic view” and “visual layout”) and in Search’s AI Mode. The authors note that it “can sometimes take a minute or more,” with “occasional inaccuracies.” In their evaluation, human expert-made websites were still preferred most, with generative UI close behind.
That’s worth keeping in view. An interface that changes with every request might sound flexible, but people still need to understand where they are and what they can do. The goal isn’t to generate a new product for every prompt. It’s to reveal the right amount of a familiar, well-designed product as the task unfolds.
The agent concierge
An agent is a system that can act on a person’s behalf. A concierge pattern gives it a clear scope: it can help, but the person stays informed and in control.
Five rules make that relationship easier to understand:
1. Say what it will do before it does it. 2. Ask before anything that spends money, sends a message, or can’t be undone. 3. Show who did what: the person or the agent. 4. Leave a receipt of every action. 5. Make undo or stop easy.
These are not details to add after the agent works. They’re part of how the agent works. If a person can’t tell what’s about to happen, who made it happen, or how to stop it, the interface hasn’t made the agent’s scope clear.
The boundary I use is simple: make the next useful step available, but leave consequential actions with the person unless they have clearly delegated them.
The case against prediction
The strongest case against predictive experiences is a real one. Predictions are often wrong. They can feel invasive. Generated interfaces can be slow and inconsistent. And many people prefer a predictable tool they control.
That’s an argument for offers over actions, for showing why something appeared, for easy dismissal, and for a stable home screen that never moves. It’s not an argument against prediction. People shouldn’t have to surrender control to get a useful suggestion.
The cost of a mistake matters. A wrong offer costs a glance. A wrong action costs trust. So make offers easy to ignore, explain why they appeared, and keep the person’s familiar way into the product where they expect it. Prediction should make the tool more useful without making it feel less theirs.
Icon lens
Predicted and confirmed are different states, so they need different visual treatments: a suggestion should never look like a decision. Agent actions also need a visible marker that shows who acted. In Digital Sanskrit, Agent (karta) is a root, so “done by the agent” can be composed onto any object rather than drawn new; the receipt is a status on that compound. The visual language can stay consistent while making the action’s status clear.
Try this
**Goal:** Design a predictive offer that helps with one repeated task without taking over the experience.
**Steps:** 1. Pick one repeated task in your product. 2. Write the signal that predicts it, then sketch the smallest offer - a chip or card - and its one primary action. 3. Mark the point where the experience would grow to a panel or full screen. 4. Write what the agent may do without asking, what needs permission, and the receipt it leaves. 5. Test with five people. Ask whether they can say why the offer appeared and how to dismiss it.
**Check:** Every person can dismiss it in one step and explain why it showed up.
**Reference:** Digital Sanskrit, Vol. IV (The Agent Is a Reader) and Appendix B.8 (One concept across every surface).
Sources
- [Google, 4 ways Pixel's Magic Cue can help you save time (Aug 29, 2025)](https://blog.google/products/pixel/google-pixel-magic-cue-ai-feature/)
- [OpenAI, UI guidelines for apps in ChatGPT](https://developers.openai.com/apps-sdk/concepts/ui-guidelines)
- [GitHub Primer, Progressive disclosure](https://primer.style/product/ui-patterns/progressive-disclosure)
- [Google Research, Generative UI: A rich, custom, visual interactive user experience for any prompt (Nov 18, 2025)](https://research.google/blog/generative-ui-a-rich-custom-visual-interactive-user-experience-for-any-prompt/)