Role: Lead Product Designer
Domain: Fintech · Desktop platform
Context
Loan refinancing involves complex financial logic: multiple loans, interest calculations, deductions. All needing to be visible simultaneously.
The users are bankers: experienced, fast, with zero tolerance for confusion.
I was designing the refinancing form as part of a broader flow. All the inputs on the form are interdependent: change the spread, and the monthly payment changes. Change the loan period, and the interest changes. Every field affects every other field.
To make that visible, we decided, as a design choice, not a requirement, to add a live calculator alongside the form. Every input the banker makes instantly updates the full repayment breakdown on the right. But that meant we'd introduced a layer of complexity we owned: would bankers actually understand that the numbers are connected? Would they notice when something changed, and understand why?
That's what this case study is about: how I built a fully functional research prototype using AI to test exactly that, and what we changed as a result.
The focus
The Challenge
The form feeds live calculations - every input on the left instantly updates the repayment breakdown on the right. I wanted to test whether this was intuitive and get feedback on the form overall.
The catch: our users are bankers. They needed to see real, accurate calculations as they interacted with the form.
That meant I needed a fully functional prototype with working math, not a simulation — without going through development.
Process
1. Prompt (ChatGPT) — Described the form logic, uploaded the initial design, defined the research goal. The first prompt is the most important.
2. Prototype (Figma Make) — Pasted the prompt with the design as a base frame. Got a functional prototype with working validation and live calculations on the first try.
3. Interview guide (ChatGPT) — In the same conversation, generated a structured research script focused on calculator visibility and field relationships.
4. Testing — 4 participants (international bankers). Sessions auto-transcribed via Gemini. Feedback documented directly in Figma; transcripts summarized by ChatGPT into Facts vs. Insights.
Design Changes After Research
-
Loan cards were information-heavy — users scanned past key details. Stripped each card down to loan number and remaining balance only.
-
Users didn't realize they needed to actively select loans — the list read as display, not input. Redesigned to make selection feel like an explicit action.
-
Deduction fields came after interest — users expected the opposite. Reordered the form to match their mental model.
-
The calculator summary didn't reflect how users think about forgiveness — they expected deductions to appear before interest in the breakdown. Adjusted the order accordingly.
-
With multiple loans selected, users lost track of the base amount — they wanted to see the consolidated total before interest was added. Added a sub-summary row for that state.
Before
After
Outcome
The research session - from prototype to insights - took 5 days, with no development resources involved.
5 concrete design changes came directly from 4 user sessions. Every change was grounded in a specific observation, not a gut feeling.
The AI-assisted workflow (prototype → testing → synthesis) has since become my go-to method for validating complex form logic before handoff. It compresses a research cycle that would typically take 2–3 weeks into a single sprint.
If I could go back...
I'd test with more participants, and specifically with less experienced bankers. Some of the friction points we found might have been even more pronounced with junior users, and some of our fixes might have introduced new confusion for experts.
I'd also spend more time on the cancellation flow. A banker who starts the refinancing process and decides mid-way to abort - where does she land? What happens to any data she entered? That path was out of scope for this research round, but it's a gap I'd close next time.