the On design, ai-assisted workflows, and the conversations we still need to have. space between us

Over the past six months, ai has brought me closer to engineers, product managers, and qa than I’d been in years.

I’ve been experimenting with ai-assisted front-end development. I’m a designer, not an engineer, and producing code I don’t entirely understand is still slightly terrifying. So I ask questions, try to understand how my decisions affect other people’s work, and spend more time talking to those who will eventually have to deal with what I’m producing.

Strangely enough, the technology has made the process more human.

But on the very same project, I experienced almost the opposite.

A screenshot doesn’t tell the whole story

We were redesigning a complex b2b product under a tight deadline. ai-assisted development allowed us to redesign a core part of the experience and introduce a new feature in time for release. The plan was to refine the ui further once the transition from legacy pages was complete.

With the deadline approaching, we missed an important opportunity to walk everyone through the final design. Feedback arrived asynchronously, mostly focused on the visual interface. One comment suggested we needed to be more consistent with rounded corners.

What made it interesting was that we already had a system. Different components used different degrees of rounding, deliberately. In a dense dashboard, shape was one way to distinguish functions without introducing more colors. What appeared to be an inconsistency in a screenshot was, in fact, a design decision.

Was that the best possible decision? I don’t know. But it was a decision, not an accident. And changing it would mean reconsidering more than the appearance of a few components.

A screenshot can show what we designed, but not necessarily why.

What bothered me wasn’t the criticism. It was how little room there was to discuss the reasoning behind the design, or the implications of changing it. Looking back, I also recognize that we hadn’t made that reasoning sufficiently accessible to everyone reviewing the work.

I don’t know whether ai was involved in producing that feedback, and I don’t think it really matters. Early on, when I was experimenting with ai-assisted design reviews, I encountered similar problems. The model might confidently identify six typography styles when there were actually three, or point to inconsistencies that didn’t exist. But the problem isn’t exclusive to ai. Anyone evaluating a complex interface through a handful of screenshots is working with an incomplete picture.

Give the same model access to the codebase, and it can inspect actual tokens, components, and patterns rather than guessing from pixels. That doesn’t make its judgment correct, but it gives it a better starting point. The prompt matters, too. If someone already believes a screen uses too many shades of one color, asking ai to identify that problem can reinforce the initial assumption.

Research into ai-assisted ui critique points to a related problem: feedback improves when designers can correct the model’s interpretation of an interface before it starts suggesting changes.

Understanding before evaluating

I think there’s a better way to approach this, and part of it starts with how I share my own work.

For complex product work, I’d rather share the actual front end, documentation, and critical design decisions, alongside a short brief explaining the goals, constraints, and reasoning behind the system. Reviewers could use that context with their ai tools instead of reconstructing the design from screenshots. They could question particular decisions or explore unfamiliar parts of the product before we meet.

I’m not expecting that to eliminate misunderstandings, or replace a conversation. But it might make asynchronous feedback more useful and give us a better starting point when we do talk.

Reviewers have a responsibility, too. When ai contributes substantially to criticism, sharing the prompt and assumptions could help both sides understand how that feedback was formed. None of this means dismissing someone’s immediate reaction. “I find this confusing” is valuable feedback and doesn’t need ai-generated validation. But proposing changes to a system requires understanding what those changes might affect.

We’re getting remarkably good at producing things faster. I’m less certain we’re adapting our habits of reviewing them at the same pace. Designers need to make their reasoning easier to access, and reviewers need to be willing to engage with it.

ai can help us understand each other’s work, but it can’t take responsibility for that understanding.

I’d rather see ai create more opportunities for those conversations, not fewer.