← All work
Owl
· 2024 – 2025
An insurance AI that flags problems and helps resolve them
Owl reads every document in an insurance file. Over three versions I took it from a dashboard of AI summaries to Escalations, where Owl flags the problems that need a person and helps resolve, defer or follow up on each one.
Role
Design engineer (freelance). Product strategy, interaction design, AI workflow and prompt design, prototyping
Team
Owl’s founders, head of product, and engineering
Timeline
Dec 2024 – 2025
Platform
Web app
Problem
Adjusters and underwriters could read Owl’s summaries, but still had to work out for themselves what was missing, contradictory or overdue in a file.
Approach
Lead with problems instead of summaries. Owl raises escalations with evidence attached, explains them on request, and does the follow-up work when someone decides what to do.
Outcome
The insurer saw it as a clear step up from OwlVision: more reliable extraction, faster decisions and smoother workflows. Owl’s sales team used that rollout as the case study that brought on six more customers.
Hundreds of pages, and a decision to make
An insurance file is years of applications, inspections, medical notes, letters and call logs, scanned into PDFs hundreds of pages long. The person responsible for it has dozens more like it.
Owl’s models read those files and pull out the facts. It took three versions to turn those facts into something an adjuster could act on and trust.
Version one: a dashboard of facts
My first project took a read-only proof of concept to production for a Fortune 500 disability insurer. It was a dashboard: a claims list, structured medical records, and a document viewer where every fact linked to the page it came from. It launched on schedule, and about 65% of adjusters ended up opening documents mainly through those citations.
OwlVision. Summaries with sources. Finding the problems was still up to the adjuster.
Version two: a conversation
The second version organized each claim into threads of work. Adjusters wrote notes, brought Owl in with an @mention, and Owl filled in tasks and templates field by field, marking anything it couldn’t find as unknown.
OwlGPT. Answers with sources, and tasks Owl fills in as it reads.
It was more useful, and it showed us where the value was. People got the most out of Owl when it pointed out something nobody had asked about: an employer no one had called in months, a job description missing from the file.
Version three: Escalations
So the third version starts there. Owl reviews every file as documents arrive and raises an escalation whenever something needs a person: two documents that disagree, a hard stop in underwriting rules, a step that’s overdue. The whole product is built around four moments.
1. See what needs attention
Home opens on a count and a list. Each escalation has a plain-language title, a status, and a due date.
2. Open the file
Inside a file, escalations sit under Needs attention, above the summary and the sources, so the problem is the first thing you read.
3. Ask about it
Open an escalation and Owl Assist offers the obvious next steps as chips. “Tell me about this escalation” gets the same three-part answer every time: what it means, why it was escalated, and how to proceed, including what Owl recommends.
4. Act on it
Every escalation ends in one of three actions: resolve, defer, or add a note. You can pick one from the escalation itself or just ask Owl to do it. Either way Owl does the follow-through. Deferring asks for a follow-up date and suggests one. Resolving asks for the reason and records it. Each step shows up in the task list as Owl works, and in the activity log afterwards.
Resolve, defer, note. Owl recommends a follow-up date when you defer, and a resolved escalation says why it was closed.
The paper trail. Tasks show what Owl is doing right now. Activities show who did what, Owl included.
Six months of tuning
Most of the work on Escalations happened after the screens existed. For about six months I worked with the team on the prompts and the workflow around them, running real files through it and reading what Owl got wrong. A few rules came out of that, and I designed each one into both the interface and the model’s instructions:
No escalation without evidence. If Owl can’t cite the page, it doesn’t raise it.
Owl recommends, people decide. It never resolves or defers on its own.
Same shape every time. Every explanation answers the same three questions, so people learn where to look.
“Unknown” beats a confident guess.
Outcome
Fortune 500 insurer moved from OwlVision to Escalations
6 new customers signed using that rollout as the case study
65% of adjusters reach documents mainly through a citation
The insurer’s team saw it as a clear step up from OwlVision. They found the extraction more reliable, since every escalation had to point at evidence, and they made decisions faster with the problems waiting on the home screen instead of buried in a summary. Having resolve, defer and note in one place, with Owl handling the follow-up, also cut down on jumping between tools.
Owl’s sales team turned that rollout into their main case study, and it brought on six more customers. Some feedback was useful in a different way: a few people wanted to pin notifications, and starting the day on a list of problems instead of a list of claims took some getting used to.