Designing a compensation workflow in code, with AI

UBS is reinventing how it runs compensation. In a one-month pilot I led the UX for a workspace that helps managers prepare year-end incentive decisions for their teams, working with a product owner, developers and compensation experts.

Project details are described in general terms, and wireframes are redrawn with fictional data, to respect confidentiality.

The problem

Year-end preparation ran on spreadsheets, emails and chat messages, worked through line by line. Evidence and decisions sat in different places, issues surfaced late, and it was hard to keep original figures, suggestions and final decisions apart.

Few topics are more sensitive than pay. Whatever we built had to protect confidential data, show where every suggestion and warning comes from, and leave every final decision explicitly with the manager.

The pattern I didn’t reuse

In my hiring work, a hybrid of chat and an agentic workspace worked well. But I had also seen its limit: dense, data-heavy decisions need more room than a side panel offers. This was exactly that kind of work. So we chose a guided workflow in a conventional dashboard, designed so AI can be added on top later to aggregate and interpret the data.

How I worked

• Design methods as AI skills. I captured how we design, from jobs to be done to information architecture and workflow creation, as reusable AI skills, and used them to turn the messy process into a clear user workflow.
• Wireframes in Markdown. Quick to write, quick to change, and easy for both people and AI to read.
• Front end in code. I built the interface directly with an AI coding assistant, using dummy data, while developers built the APIs and the calculation logic.
• AI-ready inputs. Much of the work was writing the context files and requirements the AI needed to produce something useful.

Key decisions

1. Evidence beside the decision. Data, budget impact and rule warnings appear where managers decide.
2. Suggestions stay suggestions. Nothing becomes final without the manager’s explicit confirmation.
3. Original, suggested and final stay separate, with a full history.
4. Guided workflow first, AI later, so the foundation is trusted before AI interprets on top of it.

Outcome

• Built within a one-month pilot.
• Now in user acceptance testing.
• Very positive reception, including attention from senior leadership.

How we’ll know it worked

• First bar, user acceptance testing: the pilot scope works for every scenario managers need to handle.

What I learned

• Design and development blur. In AI-assisted work, front end and back end don’t always separate cleanly. Teams need a collaboration model for this.
• Spend AI effort where it pays off. Some activities are worth it; others are still better done the conventional way.
• Connect the design system. Without a link to the Figma design system, I could only work with the coded components.
• Make the process AI-ready first. Next time I would start by helping product and business partners understand what good requirements look like, and reshape the process into one AI can work with.