INSURANCE
CONCEPT STUDY
CASE 02
An AI agent that files car insurance claims after an accident. It filled forms fast, but users couldn’t tell what it had guessed. We audited it and redesigned how it shows its work.
AGENT
Files auto insurance claims after an accident
USERS
Drivers after a collision, often stressed
ENGAGEMENT
Trust Audit + UX Sprint
FOCUS
Transparency · Supervision · User control
Concept study: a composite of patterns common in consumer insurance agents, not a client project. Scores come from our TRUST audit of the “before” state.
01 — THE PROBLEM
The agent filled claims in minutes. But it mixed verified facts with its own guesses, and submitted before anyone checked. The errors surfaced weeks later, as disputes.
✕
Guesses looked like facts
Estimated repair costs appeared exactly like verified documents.
TYPICAL USER REACTION
“I didn’t know that number was a guess.”
✕
Submitted without review
The agent sent the claim as soon as the uploads finished.
TYPICAL USER REACTION
“It submitted before I even read it.”
✕
No human when it mattered
Disputed claims led to a generic support form.
TYPICAL USER REACTION
“I just needed to talk to someone.”
02 — THE AUDIT
Each action was mapped to the Risk Ladder, then scored against 40 checks across the five TRUST pillars.
TRUST SCORE · BEFORE
40
Untrustworthy
Transparency
30
Reversibility
48
User control
42
Supervision
36
Traceability
44
Read the policy
L1
Reads silently and logs it
OK
Collect photos and documents
L2
Collected, but sources not shown
Gap
Estimate the repair cost
L2
Shown as fact, with no confidence or source
Gap
Submit the claim
L4
Submits automatically after upload
Critical gap
Answer an insurer’s question
L4
Replies for the user without asking
Critical gap
Accept a settlement offer
L5
One tap, terms not explained
Critical gap
CRITICAL FINDINGS
F1
T4 · Transparency
Estimates look like facts
Users couldn’t spot the fields most likely to be wrong.
F2
S1 · Supervision
Submission without review
A binding claim went out before the user saw it.
F3
S7 · Supervision
No human handoff
Stressed users in a dispute hit a dead end.
03 — THE REDESIGN
We redesigned the three moments where users lost track of what the agent was doing on their behalf.
F1
Every field shows where it came from
→ Verified, estimated and missing fields look different
→ A source link for every document and number
→ A plain reason behind each estimate
BEFORE
Claim details
Repair estimate
$2,340
Police report
Attached
Damage
Rear bumper
AFTER
Claim details
Police report
Verified ✓
Photos
4 · Verified ✓
Repair estimate
$2,340 · Estimated
Based on 3 similar repairs nearby · See sources
F2
A review step before anything is submitted
→ A checklist of what’s ready and what needs you
→ Estimated fields highlighted for a final check
→ Nothing is sent until the user says so
BEFORE
✓ Claim submitted
Reference #88213
AFTER
REVIEW BEFORE SUBMITTING
✓
Photos and police report
✓
Policy and vehicle details
!
Repair estimate: please confirm
Submit claim
Edit
F3
A person, the moment it matters
→ Disputes route straight to an adjuster
→ The claim context goes along, so nothing is repeated
→ Clear status and next steps while waiting
BEFORE
Claim status: Disputed
Please contact support.
AFTER
DISPUTED · WE’RE ON IT
Talk to an adjuster
We’ll share your claim, photos and notes, so you don’t have to repeat yourself.
Call adjuster
Chat
04 — THE RESULT
We re-audited the redesigned concept with the same 40-point checklist.
BEFORE
40
Untrustworthy
AFTER
88
Trusted
Transparency
30 → 90
Reversibility
48 → 84
User control
42 → 86
Supervision
36 → 91
Traceability
44 → 88
Before
After
WHAT WE’D MEASURE AFTER LAUNCH
01
Claim accuracy
Claims needing correction after submission
02
Dispute rate
Share of claims that end in a dispute
03
Fields corrected
How often users change the agent’s estimates
04
Time to submit
From accident report to submitted claim
05
Handoff rate
How often users ask for a person, and why