AI Document Extraction
ux case study
Simplifying a complex AI workflow to improve clarity and reduce drop-off across the setup journey.
The Problem-
A powerful tool that no one finished setting up.
The original product, "Setup Policy", offered remarkable AI capability — but the interface spoke a language only engineers understood.
Users were asked to define "Referenced Objects", "Validation Conditions", and "Post-Completed Functions" before they could process a single document. Dense forms, technical labels, and no guidance led to a 47% drop-off before the first document was extracted.
MY
ROLE
Timeline:
14 weeks
Team
2 PMs · 4 Engineers
Platform
Web · B2B SaaS
Role:
Lead UX/Product Designer & Product manager
Responsibilities:
- UX research
- Wireframing
- User flows
- Prototyping
- UI design
- Collaboration with engineers
- Collaboration with the client
Old Version
02 · The Goal
Make the complex feel obvious.
Improve clarity across the five-step setup so first-time users could extract their first document without help.
“What users struggled with”
“I don’t understand what this step does”
AMIR/ JBUILDING ENGENEER
“I’m afraid I’ll break something, I don't know which option to choose to create new rule”
ADAM / BUILDING ENGENEER
03 · Process
From research to ship in four phases.
A focused, sequential process. Each phase fed evidence into the next.
Step 01
User Interviews
14 sessions with operations leads, analysts, and admins.
Step 02
UX Flows
Mapped the five-step journey and isolated friction points.
Step 03
Wireframes
Lo-fi explorations to test structure before visual design.
Step 04
Design
High-fidelity system, prototyped and validated with users.
Key Insight
The product wasn't broken. The language was.
------------ Synthesis from 14 user interviews
04 · Design Approach\
Three principles guided every screen.
-
01
Plain language first
Replaced jargon with verbs users actually said in interviews.
-
02
A guide on every step
Added DocBot — an embedded AI assistant that explains each screen contextually.
-
03
One job per screen
Each step has a single primary action and a clear exit.
:
User Flow
Upload → Define Fields → Add Knowledge → Validate → Export
THE FLOW