Designing a smarter GenAI workflow for AI developers

A vague brief

A global tech company wanted to add GenAI capabilities to its existing AI development platform. But the brief was vague. It started as a paid add-on, a separate thing bolted onto the side. Nobody was sure who the primary users were, how technical they'd be, or how a new partner integration should fit into the experience. We had 12 weeks to figure it out.

Figuring out the users

We started with three initial personas provided by the client. Over time, through 6 user interviews, 12 usability tests, and focus groups, we narrowed in on the Gen AI Developer as the primary user. We spoke with 32 international data scientists and AI developers across the project. Access to users was a problem initially, but once we got in front of them, the insights came fast.

What we learned

Twelve insights shaped the direction. The most significant ones:

• Project creation was already a strength - users loved how easy it was and didn't want it disrupted
• Full customisation, not just quickstarts - templates were welcome, but only if users could go under the hood
• AI assistance was a baseline expectation - not a feature to be discovered, something users assumed would be there
• Technical language matters - the UI needed to respect the proficiency of data scientists and developers, not dumb things down
• Bundling created confusion - grouping features together made the experience harder to navigate, not simpler
• The partner tool's brand name meant nothing - leading with it confused users who had no context for what it was

The product definition shifts

This was the most important part of the project. What started as a paid add-on ended up somewhere very different. Through research and collaboration sessions with the client, we reframed it as a core feature of the platform, embedded in its project creation workflow. Not a separate thing. Not bolted on. Built in.

Designing through iteration

Three major design proposals, each shaped by what testing broke:

• Iteration 1 - split project creation into ML vs GenAI paths. Users didn't want to make that call upfront. Back to the drawing board.
• Iteration 2 - templates with customisation. This landed much better. Users felt guided without feeling constrained.
• Iteration 3 - integrated the partner evaluation tool as an entry point. Users found the branding confusing and the placement jarring.

The final design took the best of each: template selection with full customisation, a streamlined setup panel, and the evaluation tool integrated without exposing its brand.

Where it landed

The final designs were handed off with a development roadmap covering feasibility, phased delivery, and a backlog of future features. But the more significant outcome was upstream - the research fundamentally changed what the product was. It went from a paid add-on bolted to the side of the platform to a core feature embedded in the primary workflow. That's not a UX improvement. That's a product strategy shift, and it came directly from getting in front of real users.

What I took from it

Sometimes the most valuable design work isn't about the interface. It's about helping a team figure out what the product actually is before building it.