Designing a smarter GenAI workflow for AI developers
Three iterations and 32 users in, the product had stopped being a separately-named add-on
- Who
- Enterprise AI Platform (Anonymous)Jul 2024 – Oct 2024 (4mo)
- Scope
- UX researchProduct designProduct strategy
- Team
- DesignerLead strategist Product Designer
- Tools
Figma
Miro
Context
Briefed as a separately-named add-on, with the user as the open question
An enterprise AI platform wanted to add GenAI capabilities, and the initial spec framed this as a separately-named, separately-priced add-on, scoped to launch in a planned marketplace alongside the platform. Extensive on the technical side, but vague on users, with three predefined personas, no evidence on which was primary, and an undecided third-party integration.
Three open questions sat under the brief. How would this integrate into users' actual workflow, when most AI developers already had preferred tools they didn't want to leave? Who was the primary user, given three personas in the spec but no evidence on which one mattered most? And how would the third-party evaluation tool slot in, given the commercial relationship between the two companies was still being negotiated mid-project, including how prominent the tool's own brand should be inside the workflow? Each of those would shape the design in different ways, so the first job was figuring out which one we could answer first.
The product definition shift
The most valuable design work wasn't the interface, it was figuring out what the product should be
Through 6 user interviews, 12 usability tests, focus groups, and an in-person playtest with the client's AI scientists, we narrowed the persona to the AI developer as the primary user. 32 international data scientists and developers spoke with us across the project, and access was the hard part initially, but once we were in front of them, the insights came fast.
The brief had named three personas without telling us which was primary. We narrowed it through the research itself, by surfacing workflow patterns and seeing which persona those patterns actually belonged to. The AI developer emerged as primary because the patterns we kept seeing, custom code, model experimentation, evaluation as a separate step, all sat naturally with them. The other two personas hit those same steps at a higher abstraction, or through someone else's hands. So the persona call landed as a conclusion from the research, instead of an assumption we'd brought in with us.
Twelve key insights shaped the direction. Users wanted full customisation, where the brief had assumed quickstart options would do. AI assistance was expected as a baseline, and bundling features added confusion where it was meant to add simplicity. The third-party tool wrapped into the experience had its own brand, but most users didn't recognise it, so we removed the brand references and surfaced the feature by what it did for the user.
The shift: the product stopped being a separately-named add-on and became a renamed premium feature of the core platform, embedded in the project creation workflow.
Iteration discipline
Three iterations, with each round teaching the next what to fix
I structured each round to answer a specific question the previous one had opened up. The first split project creation into ML vs GenAI paths upfront, which users liked for simplicity but didn't like for the early commitment. The second introduced templates with customisation, which landed much better, although users weren't always sure if a template was local or cloud-based. The third integrated the third-party evaluation tool as an entry point, which users found confusing because the unfamiliar brand compounded it.
Because each round was designed to close a specific question, the final design resolved three sequenced design calls rather than averaging across them. The final took the best of each: template selection with full customisation, a streamlined setup panel, the evaluation tool integrated without exposing its branding, and a redesigned project landing page focused on experiment runs.
Outcome
A changed product direction and a complete product vision, handed off with a clear roadmap
Five core screens delivered: template selection, customised template setup, custom project creation, premium trial modal, and experiment runs. The roadmap covered feasibility, development, and future features. 32 users, 3 iterations, and 12 key insights later, the product had gone from separately-named add-on to embedded premium feature of the core platform.
