✓
Sudarvel
Product Designer
Case file 04 · NDA
V1.0 · 2025
Confidential · 2025 · Fintech · AI · Web
Project under NDA
The product, client and real screens are confidential, so everything you see here is an illustrative mock with dummy data. If you’re hiring, I’m happy to walk you through the full case study on a call.
00 / THE STORY
One question, two days.
▶ CLICK THE BOOK TO TURN THE PAGE
SUMMARY
Investors and leaders asked finance questions by message. Analysts dug through budget sheets, quarterly reports, dashboards and the web, then wrote the reply by hand.
The slow part wasn’t the maths. It was finding the right file, checking it was current, and writing the answer in words an investor could use. Every follow-up sent the question round the loop again.
question_trace.log
09:10 → day 2
01
Investor asks leadership by message
✕ which year, which numbers?
02
Leadership forwards it to finance
✕ context lost in the forward
03
Analyst hunts for the right files
✕ several versions, which is current?
04
Analyst adds market context from the web
✕ sources not recorded
05
Reply written and sent back
✕ a day or two later
06
A follow-up question arrives
✕ the loop starts again
WHY IT’S HARD — EVIDENCE
E1 · Where finance time goes
35%
Only about a third of planning time goes to insight. The rest goes to collecting and checking data.
So: Let the assistant fetch and stitch; analysts judge.
E2 · AI gets finance wrong
81%
A leading model with retrieval got most finance benchmark questions wrong, or refused them.
So: Never show an answer without its source.
E3 · Trust has to be calibrated
Not blind
People should trust it where it’s good and check it where it isn’t.
So: Show the working state and the sources, not a lecture.
E4 · Finance is high stakes
Ask back
A confident answer to the wrong question is a context error.
So: When a question could mean three things, ask first.
HIGHLIGHTS
I designed one prompt box that behaves like a careful analyst: it asks back when a question is vague, says what it’s doing while it works, and names the source inside every point. Google’s People + AI Guidebook was the checklist.
One prompt box with attach and voice — no dashboard to learn first
A clarifying question with numbered options before a vague request is answered
A plain working status — “Searching the web…” — instead of a spinner
Short answer points, each ending with its source, and +1 when others agree
4
states — empty, typing, working, answer
6
PAIR chapters used as the checklist
5
stakeholder groups, investors to admins
6
next-round changes from one prototype test
KEY DECISIONS
01
One box, not a dashboard
Questions change every week, so the first screen is a single prompt box with examples.
02
Ask back before guessing
A vague request gets a clarifying question with numbered options.
03
Say what it’s doing
A plain status like “Searching the web…” says where the answer is coming from.
04
The source sits inside the answer
Every point ends with its source; “+1” means another source agrees.
05
Speed or depth, your call
A model picker labelled “Fast” lets people switch to a deeper answer.
4 choices that shaped the assistant
I wrote down every choice, sorted them into big-picture and small-detail, and picked four. Each is a short story: the problem, before, what I chose, and what changed. Screens are mock-ups with made-up data.
Tap a row to see the choices in it. Things I chose not to build count too.
One box, not a dashboard
Press Next (or the → key) to read it step by step.
RETROSPECTIVE
Designing for AI changed what “done” meant. A screen can look finished while the conversation behind it is still wrong. Reading my own mock-up as an investor would caught a context error: the user asked for India’s FY 2025–26, and the answer drifted to global data for calendar 2025.
“
A good finance assistant isn’t the one with the fastest answer. It’s the one that asks what you meant.
— Reflection
WHAT I’D DO NEXT
Make clarifying options tappable chips, so one tap fixes the scope exactly.
Echo the scope back on every answer: India · FY 2025–26 · as of a date.
Add “not what I asked” and thumbs under each answer, feeding review.
Answer from the company’s own reports first, with figures in a table people can copy.
ASK AWAY. CHECK THE SOURCE.
Sudarvel · UX/Product Designer · Chennai
Designed by Sudarvel · Last updated October 2026
LinkedIn ↗