Leading a design team through the shift to AI.

The situation

I run the design org at Conviva: UX, research, and content, globally distributed. The team worked the way most enterprise design teams work. Designers made mockups in Figma. Engineers interpreted them. The distance between the two was measured in spec documents, review meetings, and rework.

That pipeline was built for a slower company than the one we were becoming. Product timelines were compressing while design headcount wasn't, and I partner with product and operations on design's budget, so I knew more designers weren't coming. The output per designer had to change, not the number of designers.

The decision

The obvious use of AI was to make the existing pipeline faster: plugins, quicker wireframes, the same artifacts sooner. I decided the artifact was the problem. A mockup is a promise someone else has to keep, and every promise costs interpretation, meetings, and drift. If designers worked in the codebase, the handoff would carry working software instead of a picture of it, and most of the interpretation would go with the picture. I'd made that argument in How to Lead Design in the AI Era. This was the chance to find out what it costs in practice.

The work

In late 2025 I moved the team off Figma and onto an AI-augmented workflow, first on Bolt.new and later on Claude Code, across the whole team and in the middle of the highest-stakes project of the year. Designers now work in the real production system and hand engineering working software instead of pictures of it.

For early discovery I built a prompt-to-prototype model on top of this: working, high-fidelity prototypes at conversation speed, before anything earns a roadmap commitment. Discovery that used to yield three wireframes in a month yielded three working prototypes in its first week.

Both parts were hard: learning the tools, and changing how designers thought about producing work at all. Moving from Figma to coding agents and a terminal changed the metaphor of design work. Designers lost the visual canvas they'd always used to explore and to compare ideas side by side, and that loss was legitimate. The new workflow is efficient and it works. It is also thinner on exploration than the old one, and I haven't finished finding that back.

The first concrete problem was comparison. You can't lay three directions next to each other in a terminal. I worked with designers on ways to get it back: a preview dashboard, a review page, a lightbox, a standalone HTML page. Some situations only needed a different prompt or a different shape of output. The designers then built their own dashboards to track their work and to show progress. I think building those tools sharpens their design thinking, since the products they design are ones customers use to build dashboards of their own.

The move to new tools was a team decision I made. How each person explores and communicates inside those tools is theirs. I didn't turn any of this into a standard workflow. Different designers approach different situations differently, and I ask each of them to choose the artifact that will communicate the idea to the audience in front of them: the pod, the CEO, me, or an engineer. The dashboards are an example of that freedom, not the prescribed answer. The standard that holds across all of it is the design harness, which is where the checkable rules live so the choice of method can stay with the person.

I ran a from-scratch design-system remediation across the codebase in parallel: standardized components, a consistent neutral token palette, and the brand green as the primary interactive color. If a feature was incomplete, I removed it rather than deliver confusion. Language got the same treatment as pixels. I directed the UX writing function and built tone and microcopy frameworks that govern content across product, docs, and marketing.

The outcome

The handoff got shorter and carries less interpretation, because engineers work from prototypes that already behave rather than pictures that suggest behavior. The prototyping model returned three high-fidelity prototypes in its first week. The tone and microcopy frameworks govern content across product, docs, and marketing. The people on the team decide how to show their work, and they've built tools of their own to do it. One of them, a principal designer I hired, worked entirely in Figma when she arrived; within a few months she spent most of her time in an IDE, delivering code and writing behavioral contracts with engineers. The exploration we lost in the move is not all the way back. That part is still being worked out.