AI Simulated Users Can Find Friction. They Can’t Stand In for Your Users

A simulated user can click through a confusing form in seconds. The harder question is whether the trace tells a designer something true about a person who needs to use it. UXCascade, a research tool presented at UIST 2026, tries to make simulated-agent feedback easier to inspect and turn into design hypotheses. Its best role is an early warning system between sketches and real user sessions.

From a pile of traces to a design question

UXCascade organizes agent-generated feedback across personas, goals, and outcomes. The tool links a suggested issue to the reasoning and interaction steps that produced it, then gives practitioners a way to explore a change and simulate its effect across different personas. Instead of treating a generated paragraph as a usability finding, the interface tries to connect an observation to an inspectable path through the product. ^[1]

That structure addresses a practical problem: simulated agents can produce lots of free-form notes, but volume is not the same as useful insight. If a report says “the form is confusing,” a designer needs to know which task failed, where the agent hesitated, and whether the complaint appears across more than one simulated path. A trace can make the hypothesis easier to inspect.

The paper reports a study with eight UX professionals and says UXCascade fit into existing workflows for iterative feedback during early interface development. That finding is about practitioner interaction with the tool. It does not establish that the simulated agents accurately represent a product’s actual audience or that their suggestions improve outcomes after a change. ^{[1] [2]}

The boundary between rehearsal and evidence

An agent follows a model, a prompt, a browser environment, and a persona description. If those assumptions are wrong, the trace can be internally consistent and still miss what matters. A simulated “first-time shopper” may not reflect a screen-reader user, a person using an older phone, someone with limited connectivity, or a customer who interprets a label differently from the model.

Simulation also risks repetition disguised as coverage. Ten runs using the same model and similar prompts are not ten independent people. A pattern repeated across agents can reveal that a flow is brittle to particular instructions; it cannot establish how common the issue is among real visitors.

This is why UXCascade’s traceability is useful but limited. A designer can inspect why an agent flagged a step, decide whether that scenario is relevant, and use the issue to improve a prototype before recruiting participants. The agent’s explanation is part of the simulated output, not independent evidence that a human would encounter the same problem. ^[1]

A practical way to use simulated agents

Use them to find obvious breaks and generate test questions. Give agents clearly bounded goals; vary device, experience, and constraints; inspect the recordings and paths; then verify promising issues with people who match the intended audience. Keep accessibility checks tied to established standards and assistive technology rather than relying on a persona prompt to emulate disability.

Measure the tool against the cost it is supposed to reduce. Does it help teams identify reproducible navigation failures earlier? Does it produce fewer false alarms after the interface changes? Can reviewers distinguish observed interface behavior from model assumptions? A useful simulation should help a team decide what to test next, not create a new score that management mistakes for user satisfaction.

The paper’s workflow also pairs well with other interface research, including the text-native video authoring prototype Doki and a separate study of human-AI collaboration in 3D editing.

The strongest promise is speed at the hypothesis stage. The strongest risk is treating a plausible simulated journey as a substitute for human evidence. UXCascade is most useful when it makes that boundary visible: an agent can help designers explore possibilities, while real users remain the authority on whether a design works for them.