From 4 hours to 8 minutes in legal contract review

The context

A 200-lawyer firm in Bogotá was processing between 40 and 60 contracts daily. The initial review of each document — identifying risk clauses, verifying expiration dates, cross-referencing data on involved parties — took between 3 and 5 hours per lawyer.

The problem wasn't the team's capacity. It was that the most qualified professionals were doing work that didn't require their legal judgment; it required time and attention to detail. Two things a well-designed system can provide more consistently.

The technical challenge

Contracts arrived in multiple formats: scanned PDFs, Word documents, natively digitized documents with different layouts. None had a standardized structure. The system needed to:

• Extract text with high fidelity from variable-quality documents
• Identify clause types without a predefined schema
• Cross-reference information with internal client and counterparty databases
• Generate a structured report that a lawyer could review in minutes, not hours
• Scale to 200 documents per day with acceptable latency

The obvious solution — applying a large language model directly — wasn't enough. LLMs hallucinate on specific data like dates, numbers, and proper names. For legal use, that's not acceptable.

The solution

We built a three-layer pipeline:

Layer 1 — Structured extraction. An OCR process enhanced with post-processing to normalize text quality, followed by a segmentation system that identifies contract sections (parties, subject matter, obligations, penalties, term) with high precision.

Layer 2 — AI analysis. A fine-tuned model on Colombian and international contracts, combined with a RAG system that queries in real time the firm's jurisprudence and historical contract database. This allows comparing clauses against known risk patterns.

Layer 3 — Actionable report. A structured summary generator that produces a one-page document with a traffic-light risk indicator per clause, automatic alerts, and suggested questions for human review.

Lawyers weren't replaced. They were freed: their work shifted from reading and summarizing to reviewing, questioning, and deciding.

The results

Average initial review time dropped from 4 hours to 8 minutes. The firm went from processing 40-60 contracts daily to over 200, with the same team. The error rate in identifying critical clauses dropped from 12% (human under pressure) to 1.3% (system + human review).

ROI turned positive in the third month of operation.

What we learned

AI projects in legal environments fail almost always for the same reason: the model's accuracy is optimized before understanding what decisions the human makes with its output. We spent the first three weeks of the project job-shadowing — accompanying lawyers in their real workflow — before writing a single line of AI code. That made the difference.