The situation
For two to three years at Empower I ran end-to-end design on complex retirement products: research synthesis, requirements analysis, user flows, annotated wireframes, UX specifications, engineering handoff, and migration into the updated design system. The design work was never the bottleneck. The work around it was. Organizing requirements, drafting specifications, and preparing artifacts took repeated manual passes, and every hour spent there was an hour not spent with research, with engineering, or on the experience itself.
What I did
- ChatGPT for organizing requirements, synthesizing research, drafting UX specifications, design rationale, and stakeholder communication.
- Claude, through the Claude Code extension in Visual Studio Code, for reviewing long documents, refining deliverables, and pressure-testing alternative approaches.
- Gemini for comparing outputs and checking complex design and documentation questions.
- GitHub Copilot for reading implementation detail and front-end code, which made conversations with engineering sharper.
- Figma Make and Figma AI for early concept exploration and evaluating component ideas. Production work stayed on the enterprise design system with manual review.
The constraint that shaped it
Our enterprise Figma sat inside a security-restricted environment that ruled out most of what people reach for to speed up design-system work: Markdown-based documentation, reusable AI skills, and automation generally.
So I built a method that worked inside those limits. Screenshots, structured prompting, and enterprise-approved tools, used to analyze existing designs, identify the right design system components, draft UX annotations, and prepare implementation guidance. The clearest example was Withdrawals by Money Source, and I adapted the same approach for Auto-RMD and notary services.
The goal was never to generate designs. It was to move a complex financial workflow into the updated design system without losing consistency, accessibility, or implementation quality.
How I checked it
AI output was a first draft. Never a deliverable.
Everything went back against business requirements, research findings, meeting notes, design files, engineering discussions, accessibility standards, and the design system before anyone else saw it. AI did sometimes invent assumptions or summarize a requirement wrong, and catching that became part of normal design review rather than a separate step.
Final design decisions stayed mine. That distinction matters more in a regulated product than anywhere else.
Working in a regulated environment
I worked within Empower's approved AI policies and used enterprise-approved tools throughout.
Prompts covered generalized design problems: research synthesis, component mapping, UX specifications, implementation planning, documentation. No participant information, no personally identifiable information, no confidential customer data, no sensitive production detail. Even inside approved tools, I kept prompts to the minimum information the task actually required.
The tooling changed how I prepared work. It did not change who was accountable for it.
Outcome
The method let me put working concepts and leadership demonstrations in front of people much earlier than the traditional workflow allowed. Earlier feedback meant the team aligned on direction sooner, refined more efficiently, and handed engineering something more complete.
The signal I trust most is that leadership and engineering came back with fewer clarification questions, because the deliverables were more structured and easier to review.
Leadership asked me to demonstrate how I was working, and I volunteered further sessions. Those covered research synthesis, UX specifications, design system migration, and engineering handoff, using Withdrawals by Money Source to show how the approach worked around our Figma constraints. The point was never to promote a tool. It was to show AI fitting inside an existing UX process with governance and human review intact.
Parts of the workflow were picked up informally by teammates. I want to be accurate about the boundary: Empower was still actively working on AI integration, and there was no defined organizational process in place when I left.
What I'd tell a team starting now
- Start with your process, not the technology. Point it at the repetitive work: research synthesis, organizing requirements, specifications, handoff. AI pays off when it supports a workflow that already exists.
- Keep human judgment in the loop. It can accelerate preparation and documentation. Designers still own research validity, accessibility, compliance, usability, and the product decision.
- Experiment inside your constraints. Every enterprise has different security, governance, and tooling limits. Build repeatable methods that fit the environment you actually have rather than waiting for a perfect capability. Consistency beats chasing the newest tool.