“If AI can already generate interfaces in seconds, why do we still need a design team?”

It’s a question Natalia Daffunchio, Nybble’s UX Lead, keeps hearing. With 25+ years in graphic design, a genuine passion for technology, and 8 years on our team, she speaks from both sides of the shift: someone who knows the fundamentals deeply and has been putting the new tools to work from day one.

In this article, she walks us through a real UX team’s process: how AI fits into each stage of design, what it does well, where it fails in predictable ways, and why the replacement narrative doesn’t match what we’re seeing in practice.

From experiment to method

Like many teams, ours started out tentatively. At first, AI was a place to ask questions and see whether the answers made sense. Over time, that exploration got more serious, and AI found a place in the workflow.

Our UX team works with the Double Diamond method, which organizes the process into four stages: discover, define, develop, and deliver. Instead of looking for one tool to do everything, they worked out which one delivered the best results at each stage. That might sound like an operational detail, but it says a lot about how the team sees AI: not as a catch-all solution, but as a set of capabilities you have to know how to combine.

Discover: feed it before you ask

The first stage carries the highest volume of information: technical documentation, interviews, recordings, client materials. This is where AI became central. The team sets up a project and feeds it the relevant material, creating a single space that holds the full knowledge base and supports understanding at this stage and every one that follows.

Once the project has enough context, the tool produces a first summary. That’s where the real work begins: asking the right questions. The information is there, but it needs to be condensed and organized to answer what’s actually needed. Like a recipe, the result depends on the ingredients and on how precisely they’re combined.

The time saved at this stage is enormous. What doesn’t change is the review. The team never leaves the loop of reviewing, validating, and reviewing again, especially when the documentation is complex.

Define: translate findings into concrete problems

Discovery produces the definitions that feed the second stage: framing the problem. What needs to be solved, what the business needs, what users need, and how the two align. The result is a clear problem statement and a brief that brings together all of the project’s definitions.

This stage demands a lot of strategic thinking. AI helps organize and structure, but deciding which problem is worth solving is still a human job.

Ideating with context (and friction)

For ideation, the team uses Claude or Codex set up with their professional profile: skills, experience, ways of working. That removes the need to explain the context in every prompt, and the answers come back in a language and framework the team understands.

Ideating doesn’t mean accepting the first thing AI generates. A model can churn out quick prototypes and code on the fly, but it lacks product judgment: everything looks brilliant to it. So we learned to push back: question its information architecture, make it justify every flow, discard lukewarm solutions, and check results with fresh eyes.

That friction isn’t a flaw. It’s part of the process.

Develop and Deliver: where judgment matters the most

In the develop and deliver stages, Claude comes in alongside the AI built into Figma and FigJam. This is where flows take shape, the product’s tone gets defined, and the team arrives at a prototype ready for handoff. And this is exactly where the team’s judgment carries the most weight.

The first AI-generated results have fairly predictable blind spots. Layouts tend to be conventional, and the typefaces are always the same. Colors show up in isolation, with no token system, and shift with every iteration, which makes it impossible to build a solid foundation for development. Text gets cut off inside components. Flows cover the happy path, where the user moves from A to B without a hitch, but leave out detours, errors, and edge cases.

Accessibility is probably the most sensitive area: light grays, illegible text, states that rely on color alone. At the same time, the opposite problem appears: over-signposted interfaces, with arrows and instructions like “tap to open” everywhere. Users have learned how to use digital interfaces. Real accessibility is built into the code so that screen readers work and the product is inclusive. Filling the screen with instructions doesn’t get you there.

Fixing all of that takes knowledge of patterns, accessibility, hierarchy, and scannability. Some of that knowledge can be taught. With skills, Claude, Codex, and other tools can absorb the team’s rules, from the token system to contrast criteria, so that each iteration starts one step further ahead. What can’t be delegated is judgment: knowing which rule applies in each case, and understanding the real context in which the product will be used.

Two cases, one lesson

The first case was a tour booking portal for a tourism company. In the initial AI-generated version, the selected day didn’t look selected. Key information sat at the end of a sentence, right where people read least. Data that changed when an option was chosen appeared on the opposite side of the screen, forcing users to scan across it to find what they needed. Some elements overlapped, and almost everything shared the same visual weight.

The final version groups information where it’s needed, clearly marks selected states, sets a reading hierarchy, and follows a design system. It works on tablets and cuts redundant text. The most important change wasn’t visible at first glance: cognitive load dropped noticeably.

The second case was an app for a window supplier, designed to be used in the customer’s home. That often means working outdoors, in direct sunlight, with a tape measure in hand or even wearing gloves. The first version ignored that context. It was generic, with small text that was hard to read and a stepper that showed the current step through color alone, which excludes people with color blindness or low vision.

In the final version, two primary actions replace four cards that competed with each other. The menu was cut down to three entry points, information is organized with labels that help users prioritize, and each step of the process is shown with text, not just color. The flow lets users photograph the window, measure it automatically, adjust the measurements, choose a product, and see it in place in the same image before generating a quote. What used to require a trip back to the office to prepare a visualization now happens on the spot.

In both cases, AI got the first version done quickly. What made it usable was a different question: who’s going to use it, where, and under what conditions.

20-60-20: where did the work go?

In a traditional process, most of the time went into manual production, leaving little room for strategy. Today the split looks more like a model that’s been circulating widely in the industry: 20-60-20. The first 20% belongs to the team: defining, making decisions, writing the specs. The middle 60% belongs to AI, which executes and delivers drafts in seconds. And the final 20% goes back to the team: validating, correcting, questioning, and redefining until the proposal works.

That last 20% is no small thing. No tool replaces a human criteria. When everyone uses the same models, with results that tend to look alike, the difference comes from a team that can spot what’s missing and has the nerve to go one step further.

Production time has genuinely gone down. What changed is where the remaining time goes: less on doing, more on thinking.

What we’re seeing in practice: the fundamentals haven’t changed

The narrative that AI will replace design teams is everywhere. What we see in day-to-day work is something else. AI has democratized speed, and today anyone can generate an interface in seconds. But judgment, experience, and point of view are still what turn that interface into a product that works.

What hasn’t changed is the craft of design itself: from how information is organized to how each flow is built.

If you’d like to know more about our UX design solutions, don’t hesitate to reach out.