Our AI claims model wasn't accurate enough. Manual reviews were high, and costs were climbing. Most members are first-timers, navigating alone on mobile with no fallback when something goes wrong.
I led design, getting business, claims, engineering, and AI to agree on one problem before we started:
- Business need
Cut costs by having AI clear more claims on its own. - User need
Clear guidance on what to submit for a claim. - Tech need
Complete, well-structured data for claims and payments.
The claim submission flow that guides you.
Guided document upload
Required documents are marked clearly.
Optional ones come with helper text so members know what applies to their situation.
Fewer fields to fill in
Fields pre-fill from the scanned document, cutting manual entry.
Members only step in to review and adjust when something needs it.
Guided toward the right selection
Members are recommended the most likely service type upfront.
Clinic and diagnosis search return close matches, not just exact database names.
Consent, only when it's needed
Manual input steps in only where it must.
Members flag if the claim is from an accident or a third party, and give consent, before submitting.
When the process is complex, data and user validation guided me.
The model itself pointed to why: a lot of the data it was receiving was messy or incomplete.
That sent me into the submission flow itself, where I ran a UX audit alongside the product's own analytics to identify four points of friction.
What users told us about the four friction points.
I decided to test with four users to validate why: three with no experience submitting a claim, and one who had submitted before, on someone else's behalf.
- Benefit type
18% of users picked the wrong benefit type.
1/4 users said
“I don't understand the benefit type's name.” - OCR loading
33% of users dropped off during scanning.
4/4 users said
“It's stuck on a page. Did the app crash?” - Data entry
20% of users gave mismatched answers.
3/4 users said
“Did my document say I paid last Friday? Not sure.” - Diagnosis search
26% of users couldn't find their diagnosis.
2/4 users said
“I have a flu and cough, I don't see the selection.”
Two product strategy proposals. One would lay the foundation.
From the previous data, I learned that claim submission is a scenario-based experience. I explored two directions that could support and scale with that in mind.
Direction 2: Upload once, let AI classify. Where we were heading, but the model wasn't ready.
We focused on improving four main experiences.
1. Upload document→Scenario guided document upload
Before
After
What changed
Users can now see which documents are required at a glance, with helper text guiding the optional ones.
2. Static loader→Scanning document animated loader
Before
After
What changed
Users can now see visible scanning progress, so the wait reads as working, not broken.
3. Free for all→Controlled and smarter search
.png)
Before
After
What changed
Users can now find the closest match instantly, with an option to add a new clinic only when it's genuinely missing.
4. Medical terminology→Everyday language every user knows
.png)
Before
After
What changed
Users can now search in plain language instead of medical terminology.
I saw the real user experience at a client's office.
Phase 1.0 shipped in Q4 2025.
I spoke with members and realized claims can be highly stressful too.
Two problems stood out, and how I solved them post-launch.
- Upload screen
Still felt overwhelming. Options sat at equal weight, so members scrolled past what they needed.
Fix
Gave required documents clear priority over optional ones, with helper text guiding the rest. - Diagnosis field
One wasn't enough. Prescriptions often cover more than one condition.
Fix
Added a second field.
The AI gets better data. The team chases fewer corrections.
What I'd do differently.
- Push for the bigger fix earlier. The friction wasn't a surprise, it was the corner I'd cut.
- The form feeds the model. Structured data mattered more than a polished screen.
