What I learned designing for systems that think

UX design has always demanded that we understand complex systems and find the clearest path for people through them. That work, rooted in research, behavioral observation, and validation, hasn't changed with AI.

What has changed is the system itself. Before, we designed for predictable outputs: the same input always produced the same result. Now, the system learns, interprets, and responds in variable ways. It can fail in ways you didn't anticipate. It can get things right in ways nobody expected.

This adds a new layer to the designer's responsibility. We still need to understand users. We now also need to understand how the model behaves so that we can design experiences that remain trustworthy even when the output is uncertain.

A Real Problem, on a Production Line

In partnership with an automotive company and Google Cloud, we built a manufacturing defect inspection PWA integrating Google Cloud Vision AI with SAP. The computer vision model classified defects in industrial parts in real time.

High precision, but not absolute precision. On a production line, a wrong classification has a real cost.

The most important design decision wasn't about layout or visual hierarchy. It was about when the human operator needed to step in.

The flow was designed so the system would present its classification alongside its confidence level, and the operator always has the final word before any irreversible action. The interface didn't hide the model's uncertainty. It made that uncertainty readable.

That project was recognized at Google Cloud Next '26, and it confirmed something I had been building toward: transparency in AI products isn't an optional feature. It's a foundational UX requirement.

Four Principles That Guide My Work

Make uncertainty visible. If the model is 67% confident in a response, the user needs to know. Not to confuse them, but to help them calibrate their own decision. Hiding uncertainty creates false confidence and increases the risk of error.

Design for failure, not just success. The happy path is the easy part. The real work is in the hard cases: what happens when the AI returns something unexpected? When the user disagrees with the suggestion? These states deserve as much attention as the main flow.

Return control when it matters. Some decisions have real consequences for real people. No model, no matter how accurate, should make those decisions alone. Design needs to make human intervention accessible and frictionless.

System feedback is part of the experience. When AI acts, the user needs to understand what happened, why, and what they can do next. In AI products, the response doesn't follow a fixed script. Feedback needs to be clear even when behavior is variable.

Why This Matters Now

Many products are incorporating AI quickly, but most interfaces still treat the model as a black box that delivers answers. The user clicks, the AI responds, end of interaction.

That approach hides a cost that shows up later: people who stop using the product because they don't understand what happened, because they made an error they couldn't fix, or because they simply don't trust what the system does.

Designers don't build AI models. But we are the professionals best equipped to translate the behavior of those models into something people can understand, use, and correct when needed.

This work is still being defined. And the answer, as it turns out, isn't entirely new. As Kate Moran, VP at Nielsen Norman Group, puts it:

"When designing with new tech, lean on old best practices."

The foundations of UX, clarity, user control, error recovery, trust, have never been more relevant. The systems got smarter. The work of understanding people didn't get easier. It got more necessary.