Outcome · Put people where judgment belongsFederal health payments

Ten million pages processed, with a person making the determination.

~10M
Pages processed to date: >30%
Less document-review time: 0.90 / 1.00
Accuracy classifying 25 document types

01 · Problem

What was in the way?

Clinical review volume outran reviewer capacity. Every claim needed documented determination, and the backlog was growing faster than staff could be added to it.

02 · AI at work

What did the AI teammate do?

The AI Medical Records Reviewer classifies 25 document types and bookmarks the page that carries the evidence, so cases that need judgment reach reviewers first.

03 · Human role

Where did people keep authority?

A clinician makes every improper payments determination. No change to who decides, increasing reviews with no reduction in review staff.

04 · Result

What changed?

Document-review time fell by more than 30% across roughly ten million pages processed.

One record, before and after the teammate.

Before the teammate

A thousand pages, and no way in.

Reviewers searched 1,000+ page records for discharge summaries, labs, physician notes and EKGs. Tens of thousands of pages per clinical case, under compliance deadlines, with real risk of human error.

After the teammate

The page that carries the evidence, bookmarked.

The teammate bookmarks the page that carries a section, not every page a class appears on, and maps the class to the section as it is actually titled. The reviewer confirms. Every bookmark is still a reviewer decision.

Drag to compare

0.90 weighted-average F1 across 25 classes, and more than 30% less document-review time. We ruled out deterministic rules before building. We piloted a retrieval system and killed it.

Have a problem like this? Let's talk.

Talk to us →Download the case study ↓