The Hypothesis

What if the interface rearranged itself based on what the user actually needs?

The Concept

Most UIs are static layouts designed by humans for average users. But no user is average. This experiment tests whether an AI agent can observe user behaviour in real time and dynamically restructure the interface — reordering menus, surfacing relevant tools, hiding noise — so the UI adapts to the person, not the other way around.

AI-controlled UI

The hypothesis

What if the interface rearranged itself based on what the user actually needs?

The concept

Most UIs are static layouts designed by humans for average users. But no user is average. This experiment tests whether an AI agent can observe user behaviour in real time and dynamically restructure the interface — reordering menus, surfacing relevant tools, hiding noise — so the UI adapts to the person, not the other way around.

How it works

1. User opens app
2. Observe behaviour — click patterns, dwell time, scroll depth
3. Infer intent — intent model, task prediction
4. Restructure UI layout — reorder nav, surface tools
5. Measure outcome — task completion speed, error rate

Continuous loop — the UI reshapes itself every session based on accumulated behaviour signals.

What it explores

• Can an AI reliably predict what a user needs before they click?
• How much UI rearrangement is helpful vs. disorienting?
• Does adaptive UI reduce task completion time or increase cognitive load?
• What’s the minimum observation period before the model is useful?
• How do you handle the cold-start problem for new users?

What we found

• Early prototype reduced average task completion time by 22% for power users
• New users were confused by adaptation before enough behaviour history had accumulated
• Subtle rearrangement (reordering sidebar items, surfacing recent tools) outperformed full layout restructuring

Learnings

• Gradual adaptation beats sudden rearrangement — users need continuity to build muscle memory.

• Intent prediction works best for repetitive workflows, poorly for exploratory tasks.

• A “reset to default” escape hatch is essential for user trust — adaptive UIs must always let the user take back control.

• Task sequence patterns are a stronger signal than individual clicks — what someone does after something matters more than the click itself.

Where it goes next

We’re exploring whether this adaptive UI approach could become a core feature of Flywheel’s campaign builder — letting the interface learn each marketer’s workflow and surface the right tools at the right time.