NN/g: Designing AI agents — 4 lessons from China's Qwen

Feifei Liu and Maria Rosala from Nielsen Norman Group studied Qwen, an AI agent from Alibaba widely used in China, to understand what happens when everyday users encounter agent-based interfaces for the first time. The result is a set of four design lessons grounded in observed user behavior rather than design theory.

The article is useful because Qwen operates at scale with a genuinely broad user base — not early adopters, but people with varying levels of AI familiarity. Watching how those users navigate agent interactions reveals failure modes that usability testing with motivated participants can miss.

Support discoverability through redundancy. Most users default to familiar input patterns rather than exploring new interaction modes. Qwen addresses this by offering multiple entry points: both a chat interface and a browsing layer with categorized task templates. Auto-filled prompts help when categories are narrow, but broader task types benefit from the agent asking clarifying questions before proceeding. The lesson for designers is that a single entry point for AI features will underserve users who do not know what to ask.

Use familiar patterns to reduce learning cost. When the interface borrows from known conventions — Qwen’s task browsing resembles a delivery app — users feel oriented faster. The caveat is that borrowed patterns must fit the context. Carousels that show only one item at a time mislead users about how many options are available. Familiarity helps onboarding; misapplied familiarity creates confusion at the wrong moment.

Handle personal data carefully. When Qwen displayed users’ full addresses without explanation, participants reported feeling their privacy was violated even though they had authorized the data access. Transparency about what data the agent is using, and why, needs to happen before the agent acts — not after. Clear authorization screens and minimal data disclosure at each step are more trustworthy than comprehensive data access with a buried explanation.

Prioritize transparency to protect user autonomy. Users abandoned agent-suggested options when hidden fees or missing decision-critical information appeared late in the flow. Efficiency gains from AI assistance disappear if users lose confidence in the agent’s outputs. Showing enough of the agent’s reasoning to let users evaluate recommendations is not a nice-to-have — it is what keeps users in control.

The throughline: capability alone does not determine whether an AI agent succeeds with new users. Usability, transparency, and user control determine adoption.