Where AI Helps, and
Where I Still Have To
I want to showcase how I use AI in my daily work and how it helped reduce the timeline of a real project from 3–5 weeks to 4–8 days.
Role
Product Experience Designer
Role
Product Experience Designer
Service
Product design, Design research, Agentic engineering
Service
Product design, Design research, Agentic engineering
Year
2026
Year
2026
The Stakes
The Stakes
Adyen had no centralized way to end a merchant relationship. When Risk flagged a customer for termination, the entire process ran through people, not systems. An Ops team was manually tracing account hierarchies, drafting notice letters, and making sure every step happened correctly, every time.
Nothing had gone wrong. That's a credit to Ops, not to the process. The system worked because they were careful, not because it was built to hold up on its own, and that's a lot of weight to keep resting on manual diligence.
We wanted to get Ops out of that position; asking skilled people to be the only safeguard on something this consequential isn't a system. It's a bet. We needed something that let them stop being the fallback and focus more on making judgment calls to ensure we stayed in line with compliance audits.
Adyen had no centralized way to end a merchant relationship. When Risk flagged a customer for termination, the entire process ran through people, not systems. An Ops team was manually tracing account hierarchies, drafting notice letters, and making sure every step happened correctly, every time.
Nothing had gone wrong. That's a credit to Ops, not to the process. The system worked because they were careful, not because it was built to hold up on its own, and that's a lot of weight to keep resting on manual diligence.
We wanted to get Ops out of that position; asking skilled people to be the only safeguard on something this consequential isn't a system. It's a bet. We needed something that let them stop being the fallback and focus more on making judgment calls to ensure we stayed in line with compliance audits.
The Approach
The Approach
Pull the scope together
To understand the whole picture of the offboarding process, I interviewed the fraud, risk, and compliance teams to understand their current workflows, expectations, and pain points. I fed the interview data into NotebookLM to summarize the workflows and journey maps, then aligned on the problem statement and scope with my PM and tech lead in our trio meetings.
Spin up something real
I converted the PRD into a prompt and shared it with Figma Make, picked mainly because it's collaborative and easy for people to comment on directly. After a few iterations, I had an initial prototype, a few hours' work. I shared it directly with my PM, tech lead, and potential users so they could comment and I could collect initial feedback fast.
Validation loop
I ran interviews with the same group, then used NotebookLM to help synthesize the testing findings into patterns instead of doing it by hand. From there, I could fine-tune the design together with my PM and tech lead in the same session, prompting and adjusting live instead of scheduling a separate alignment round.
Pull the scope together
To understand the whole picture of the offboarding process, I interviewed the fraud, risk, and compliance teams to understand their current workflows, expectations, and pain points. I fed the interview data into NotebookLM to summarize the workflows and journey maps, then aligned on the problem statement and scope with my PM and tech lead in our trio meetings.
Spin up something real
I converted the PRD into a prompt and shared it with Figma Make, picked mainly because it's collaborative and easy for people to comment on directly. After a few iterations, I had an initial prototype, a few hours' work. I shared it directly with my PM, tech lead, and potential users so they could comment and I could collect initial feedback fast.
Validation loop
I ran interviews with the same group, then used NotebookLM to help synthesize the testing findings into patterns instead of doing it by hand. From there, I could fine-tune the design together with my PM and tech lead in the same session, prompting and adjusting live instead of scheduling a separate alignment round.
Make it dev-ready
I switched from Figma Make to Windsurf, copying over the design details I'd already fine-tuned in Figma. Windsurf reads from our design system MCP, so what it generated matched the visual and interaction patterns we'd already defined closely, not approximately.
Dev build from a real starting point
That Windsurf output became something my frontend dev could actually adapt instead of building from scratch. We had a few check-ins so he could quickly review the code quality before I pushed it to our internal GitHub repo for him to pick up.
Confirm the match
I started from AI-generated code anchored to the Figma design, so testing was mostly about confirming quality, not hunting for drift.
Make it dev-ready
I switched from Figma Make to Windsurf, copying over the design details I'd already fine-tuned in Figma. Windsurf reads from our design system MCP, so what it generated matched the visual and interaction patterns we'd already defined closely, not approximately.
Dev build from a real starting point
That Windsurf output became something my frontend dev could actually adapt instead of building from scratch. We had a few check-ins so he could quickly review the code quality before I pushed it to our internal GitHub repo for him to pick up.
Confirm the match
I started from AI-generated code anchored to the Figma design, so testing was mostly about confirming quality, not hunting for drift.
The Outcome
The Outcome
What used to take 3 to 5 weeks now takes 4 to 8 days, largely because of the AI tooling in the loop. That's not just a speed win; it's fewer weeks where the process depended entirely on people getting it right every time.
On the frontend side, based on the retro with my frontend dev, honestly, it made things somehow faster, not noticeably faster. What did change: for custom components, which our design system allows certain customization on, he could easily catch the design details and reference the code instead of building from scratch.
What used to take 3 to 5 weeks now takes 4 to 8 days, largely because of the AI tooling in the loop. That's not just a speed win; it's fewer weeks where the process depended entirely on people getting it right every time.
On the frontend side, based on the retro with my frontend dev, honestly, it made things somehow faster, not noticeably faster. What did change: for custom components, which our design system allows certain customization on, he could easily catch the design details and reference the code instead of building from scratch.
The Judgment Call
The Judgment Call
One thing I benefited a lot from is the speed of validating. I didn't need to spend a lot of time tweaking the prototype to make it look real. With AI, I could get a high-fidelity prototype ready to test within a few hours. I uploaded the testing recordings to Condens and NotebookLM to expand the context for my AI agent.
I tried asking NotebookLM to convert the user testing findings into a prompt and shared that with Windsurf. It covered about 50 to 60 percent of the feedback. That made me realize trio alignment, with product and tech, is still essential. Prioritization, trade-off decisions, even the critical judgment calls, those all come down to context AI doesn't have. Without that context, it can't actually tell what matters or why.
One thing I benefited a lot from is the speed of validating. I didn't need to spend a lot of time tweaking the prototype to make it look real. With AI, I could get a high-fidelity prototype ready to test within a few hours. I uploaded the testing recordings to Condens and NotebookLM to expand the context for my AI agent.
I tried asking NotebookLM to convert the user testing findings into a prompt and shared that with Windsurf. It covered about 50 to 60 percent of the feedback. That made me realize trio alignment, with product and tech, is still essential. Prioritization, trade-off decisions, even the critical judgment calls, those all come down to context AI doesn't have. Without that context, it can't actually tell what matters or why.
There's a lot of discussion right now around AI design copilots, specifically, whether designers should be the ones building and merging directly on the production site themselves. I'm naturally excited to try new things, but I keep questioning where the real value actually is.
My current take:
for a brand new feature, dev should still own the production build. They know the API and handle the long-term maintenance, but AI did change his starting point (like this project).
For UI refinements and content changes, the stuff that often gets deprioritized, designers iterating directly in the repo and opening a PR is a real option now.
And that line keeps moving anyway. As I mentioned previously, AI changes fast enough that I'm still experimenting case by case, not settling on one fixed rule.
There's a lot of discussion right now around AI design copilots, specifically, whether designers should be the ones building and merging directly on the production site themselves. I'm naturally excited to try new things, but I keep questioning where the real value actually is.
My current take:
for a brand new feature, dev should still own the production build. They know the API and handle the long-term maintenance, but AI did change his starting point (like this project).
For UI refinements and content changes, the stuff that often gets deprioritized, designers iterating directly in the repo and opening a PR is a real option now.
And that line keeps moving anyway. As I mentioned previously, AI changes fast enough that I'm still experimenting case by case, not settling on one fixed rule.
