← All work

Owl

· 2025

One wrong line used to mean a whole new report

When a claim got new information, Owl’s reviewers had to regenerate an entire template to fix a single line. Working closely with Owl’s largest customer, I designed a way to select any part of a template, ask Owl to improve or explain it, and keep only the changes they want.

Role

Design engineer (freelance). Interaction design, AI workflow design, prototyping

Team

Owl’s head of product and engineering, with reviewers at Owl’s largest customer

Timeline

2025 (Owl v1.5), revisited after Escalations

Platform

Web app

Problem

Templates were all or nothing. When a claim received new information, reviewers regenerated the whole template, waited for it, and checked every field again.

Approach

Make templates editable in place. Select any text, ask Owl to improve or explain it in plain language, and commit only the changes you keep. Every edit is saved as a version, so there is nothing to publish.

Outcome

Fixing a template went from a full regeneration to a single request. We shipped what the existing stack could support, then returned with the client’s feedback, a conversational Owl Assist and inline diffs.

Owl template editing: a claim timeline with the claimant’s name corrected inline, the Improve and Explain menu, and Owl Assist confirming three updates

A whole template for one wrong line

Owl turns an insurance file into templates: claim summaries, timelines of key events, and the forms a reviewer would otherwise fill in by hand. Templates are what Owl’s customers pay for, and each one is shaped with the client field by field.

But a claim doesn’t stand still. New medical records arrive, a name turns out to be incomplete, a diagnosis gets corrected. The only way to update a template was to regenerate all of it, wait, and reread every field to make sure nothing else had moved. For reviewers and adjusters working through dozens of claims, that was slow and frustrating, and it made them hesitant to fix small things at all.

Designed with the people using it

We built this alongside Owl’s largest customer, a Fortune 500 disability insurer. Their reviewers walked us through real claims, showed us where templates went stale, and reacted to each round of designs against their own workflow. That set the brief: edits had to be ad hoc, small, quick to check, and never surprise anyone else on the claim.

Select, ask, keep

The interaction borrows from document editors people already know. Highlight any text in a template and two actions appear beside it: Improve and Explain.

A claim timeline with one event highlighted and a menu offering Improve and Explain for the selected text
Select anything. Highlight a line and Owl offers to improve it or explain where it came from.

Improve opens Owl Assist with the selection attached and the prompt already started, so the reviewer only has to finish the sentence: “include the claimant’s full name.” Explain answers the question reviewers ask most often, why a line says what it says, without touching the template. The conversation is scoped to whatever template is open, and anything can be typed free-form.

Owl Assist with the selected summary text attached and the prompt “Improve this region: include the claimant’s full name”
Ask in plain language. The selection becomes the context. The reviewer just says what’s wrong.

Owl makes the change across the template, not just in the selected line, and every edit waits for a decision. Reviewers can keep or undo each one, or the whole set at once. Nothing is committed until a person says so.

Scoping what we could ship

The ideal version showed every edit inline, struck through and highlighted like tracked changes. Owl’s tech stack at the time couldn’t support diff highlighting, so I worked with engineering to find the version that it could.

In the MVP, changes happen at the field level. Keeping a change keeps the whole field, and Owl only ever edits field values, never the names or titles the client had agreed to. Owl’s replies stayed simple too: a short confirmation of what changed, rather than a full walkthrough. It was a smaller feature, but it removed the full regeneration from a reviewer’s day.

A few product rules came out of that scoping, and they all held up later:

  • Every edit is a version. There is no publish step to forget, and everyone sees the same template.

  • Nothing is ever lost. Old versions are a click away, and restoring one creates a new version instead of erasing history.

  • No locks. Anyone on the claim can edit. A template only one person can touch is a less useful template.

A claim summary with a Version History button and a panel listing previous Owl Assist conversations by time
History. Conversations and template versions are kept, so any change can be traced and rolled back.

Coming back with diffs

After template editing shipped, I spent the next stretch of Owl designing Escalations: a workspace where Owl flags the problems in a file and works through them with you in conversation. That work moved the whole product forward, including the stack that had held the first version back.

So we went back to the client. We asked reviewers how editing had been going, what they still worked around, and where they hesitated before keeping a change. Then we brought Escalations’ conversational Owl Assist and inline diff highlighting to templates. Owl now shows its reasoning step by step as it works. Removed text is struck through, new text is highlighted, and every change has its own Keep and Undo.

A template with field-level diffs: the claimant’s name corrected in two fields and an unspecified diagnosis removed, with Owl Assist explaining that it updated two instances
Diffs on every field. One request updates every field it affects, and each change can be reviewed on its own.

Outcome

Fixing a template went from regenerating the whole thing to typing a single sentence. Reviewers could correct a name, add a missing detail, or ask why a line said what it did, without waiting on a full rerun or rereading fields that hadn’t changed. And because Owl only reworks what it is asked to, edits stay small and predictable.

Shipping the smaller version first paid off. By the time we returned, reviewers had months of real edits behind them, so their feedback was specific. The second pass went straight at what they asked for: seeing exactly what changed before keeping it.

Some details are simplified or left out to respect client confidentiality.