Case study · Build
Legal Services
High-volume litigation · Canada
Intake accuracy ~20% → near-perfect
A Canadian litigation practice processing around 700 court filings a year.
The operation.
A Canadian firm running a high-volume litigation practice. Around 700 matters a year, three to ten new filings a day, each arriving as a scanned court document.
Every filing triggers the same admin sequence: an intake form, a tracker update, and an email routed to the right contact at an institutional counterparty.
The problem.
Each filing was processed by hand. Read the scan, find the client details on page one, retype them into a form, update the tracker, draft the email. Two or more hours a day, pure transcription.
Routing depended on knowledge held in people's heads. With only a spreadsheet for a record, matters that were never paid slipped through unnoticed.
What we did.
Conventional OCR failed on these scans, so we skipped it. We send the scanned filing straight to a vision model as a native document, the way a person reads it, stamps and handwriting included. Accuracy went from about 20% to near-perfect.
The pipeline pre-fills the form, updates the tracker, drafts the routed email, and flags low-confidence reads for review. Routing now follows a documented registry instead of memory.
What changed.
They didn't force a generic solution onto our practice. They tested the documents we actually receive, discovered that conventional OCR wasn't reliable enough and developed an approach that achieved near-perfect extraction on our filings.
Managing partner, litigation practice
Status
Phase 1 in production. A custom case-management app is in build, and a predictive layer is scoped.
Methodology
Recovered hours measure the elimination of manual transcription across three to ten filings a day.
Have a workflow like this?
Schedule a Conversation→