JPMorgan Chase

JPMorgan's investment banking division handled roughly 12,000 commercial loan agreements per year, each typically 10–20 pages with about 150 extractable attributes (rates, covenants, collateral, grace periods, and related terms). Before automation, lawyers and loan officers manually reviewed each agreement, entered data into internal systems, and ran a second quality check. At an average of about 30 hours per agreement, the process consumed approximately 360,000 person-hours annually, drove tens of millions of dollars in labor cost, and remained error-prone on repetitive, detail-heavy work.

In June 2017, JPMorgan launched COIN (Contract Intelligence), a Python-based NLP system to extract key terms from commercial loan agreements. The stack combined OCR for scanned PDFs, named entity recognition for parties/dates/amounts, pattern matching and classification into the 150 required fields, and confidence scoring that routed uncertain extractions to human reviewers. A mixed team of technologists, data scientists, and legal experts spent considerable time building, training, and validating the system before production deployment.

JPMorgan reported that time per agreement fell from roughly 30 hours of manual work to seconds of machine processing plus minutes of human review on flagged items. Annual hours consumed dropped from 360,000 to an estimated under 40,000 for human oversight of flagged extractions. Beyond direct savings, loan officers could access extracted terms within hours rather than days or weeks, extraction consistency improved across the portfolio, and professionals were redeployed to higher-value advisory, deal structuring, and risk analysis.

• Automate high-volume, structured extraction tasks and keep lawyers focused on negotiation, ambiguity, and judgment—not field entry.
• Pair domain experts with data scientists: defining the 150 fields, legal language training, and validation rules was as critical as the models.
• Use an 80/20 pattern: auto-handle routine cases, escalate low-confidence items, and invest first in the document data pipeline around the model.