From Screens to Goals
Nielsen Norman Group named this shift directly in its research on generative UI: instead of one interface serving many users adequately, an AI-generated interface can be assembled per person, per session, tuned to that individual's context. The designer's job moves from placing components on a screen to defining the goal the system should optimize for, and the guardrails it should never cross while doing it.
A Concrete Case
NN/g's own example: Alex is booking a flight. Because the system knows she's dyslexic, it renders the page in a higher-contrast, dyslexia-friendly font. Because it knows she usually flies out of Miami, it pre-fills her origin. Because her history shows she avoids red-eyes and prioritizes window seats, it surfaces those flights first, without her filtering for them. None of that is a screen a designer drew. It's a set of rules a designer defined, executed differently for every user who hits the flow.
What the Designer's Job Becomes
- From screens to systems: the deliverable shifts from a specific layout to a design system, a set of content rules, and behavioral constraints the AI operates inside.
- From flows to intents: instead of mapping thirty-seven onboarding screens, the designer models the intents those screens existed to serve - what the user is actually trying to accomplish at each moment of uncertainty.
- From one path to many: a fixed checkout is one path. An outcome-oriented checkout completes the purchase through chat, voice, one-click, or a traditional form, chosen by the system based on the user's behavior in that session.
- Evaluation changes shape: you can't run a heuristic review against a single static design anymore. You have to evaluate the range of experiences the system generates across genuinely different users.
What This Costs, Right Now
NN/g flags three real constraints, not hypothetical ones. Privacy: generative UI needs deep personal data to personalize anything, which is a harder sell every year. Consistency: an interface that looks different every visit works against the learnability that keeps users oriented. And capability: current models still hallucinate and inherit training-data bias, so a generated interface can confidently get something wrong in a way a static, tested one won't.
The Same Shift, Seen from Service Design
NN/g's research on service design describes the same transition from the operations side: as AI assistants start acting on a user's behalf - comparing options, executing a plan, surfacing exceptions for approval rather than every step - services start competing on AI-to-AI compatibility and data accuracy as much as on the visible interface. Its retirement-planning example: the AI assistant analyzes the numbers, builds the plan, and executes portfolio changes with the user's approval at the decision points that actually matter.
Key Takeaways
- Outcome-oriented design replaces a fixed interface with a goal, a set of guardrails, and a system that generates the path per user.
- The designer's core artifact shifts from screens and flows to intent models, content rules, and behavioral constraints.
- Privacy requirements, interface inconsistency, and current model limitations are real near-term costs, not distant hypotheticals.
- The same shift is visible from the service-design side: AI assistants executing plans on a user's behalf change what organizations compete on.
