What if underwriters could review more submissions
without compromising the rigour and judgement behind every decision?
Designed the Agentic Experience to move underwriters
from analysts to decision-makers
Context
Nirvana is a commercial trucking insurer writing $100M+ in premium. A submission is a trucking company's application for insurance.
The overview shipped behind a feature flag against a real holdout, so the results are measured against underwriters still working the old way.
What the ai does
Gathered the insured’s safety, loss, insurance and driving records and any other pending details.
Review the submission against 16 guidelines written by the underwriting team.
Read the submission for red flags.
My Role
I led design with a product manager, a senior underwriter and the engineering squad.
The underwriting team owned the rules.
I owned how they appeared to underwriters, and designed the flows and prototype myself.
The stakes
Underwriters carry the risk on every policy they bind, so the AI could recommend but the decision always stayed with them.
If the interface showed risk unclearly, an underwriter could make a decision that looked right on screen and still bind a policy that hurt the book.
Every submission is different.
Some are clean enough to quote in two minutes.
Some are clearly outside what Nirvana will insure.
Others are ambiguous, need real judgement.
And some aren’t ready to review at all.
10 - 20 mins
Data-gathering time per submission
Equal review cost per submission
Good risks were buried


Underwriters don't trust a verdict they can't verify
An underwriter rarely asks whether a submission passes. They ask how close it sits to the limit, and what that number is based on.
My earliest version went straight to the answer, and underwriters ignored the page altogether. More on that later.
The recommendation is broken into 3 parts:
The Recommendation: What's the verdict?
The Signals: What is it based on?
The Guideline: What threshold was it checked against?

Underwriters think in degrees, not in pass or fail
One level down, every value is traced to its source and dated.



The underwriter opens a submission as
a decision-maker, not a data-gatherer.




Trust Signal
Every check shows its working
Each check shows the guideline’s value, the rule, its source and the date. A denser table, but underwriters can tell a bad risk from bad data.
I hid each value’s details in a tooltip to keep the table clean. But underwriters couldn’t see how close a value was to the limit, which told them more than pass or fail.


2x submissions reviewed in a year,
with the same underwriting team
52%
Reduction
In submissions needing manual review
6,000
hours
Of review work removed in five quarters, roughly three underwriter-years.
1 in 1,000
Of review work removed in five quarters, roughly three underwriter-years.

