Skip to content
All projects
Case study · engineering

Document Analysis Automation

50,000 documents read in six hours for an $8B dispute.

50,000 Docs in 6 Hours

Context & Constraints

A major legal dispute worth $8 billion was running in parallel between the US and Brazil. The case involved retroactive tax exemptions with significant financial impact for the end client. The analysis required searching through 50,000 scanned documents — dating back to 1958 — of various types (board minutes, commercial registry records) for terms indicating profit and dividend distributions to shareholders.

The initial estimate required a team of 5 consultants working for 10 days, and even then, only a sample-based review would be feasible. The volume, variety, and age of documents made comprehensive manual analysis impractical, leaving critical evidence potentially undiscovered.

Architecture

Built an end-to-end automation pipeline: a Python script split documents into individual pages (dramatically improving analysis quality), which were then routed to a SharePoint processing folder. A Power Automate flow orchestrated the analysis using AI Builder and Copilot Studio. The deliverable was a spreadsheet with one row per page, including a column indicating whether the page mentioned dividends.

Document Analysis — end-to-end automation pipeline

My Role

Called in after multiple failed attempts by the business team to solve this with existing AI platforms. Designed the full solution architecture: Python page-splitting script, SharePoint integration, Power Automate orchestration with AI Builder and Copilot Studio. Delivered the end-to-end pipeline that produced auditable, page-level results.

Results & Validation

50,000 documents analyzed in 6 hours

+400 hours of manual work saved

98% accuracy validated by business team

Processed 50,000 documents in 6 hours with 98% accuracy (validated by the business team's sample audit). What would have taken 5 consultants 10 days — and still only covered a sample — became a complete, page-level analysis. The key was decomposing documents into pages before AI processing, which dramatically improved granularity and accuracy.

Stack

Copilot StudioPower AutomateAI BuilderPython

Building something where reliability matters?

I design, ship and operate systems like this one.

Let's talk

Building GenAI where reliability matters? Let's talk.

leonardo@leonardosa.pro

Madrid, Spain · from August 2026 · English / French / Portuguese / Spanish

© 2026 Leonardo Costa de Sá