JPMorgan Chase & Co.

JPMorgan Chase is the largest U.S. bank by assets, holding $2.5 trillion as of year-end 2017. Its commercial banking division generates thousands of credit agreements, non-disclosure agreements, and custody contracts annually — each requiring manual review by attorneys and loan officers to extract key terms, verify compliance provisions, and flag non-standard clauses. By 2016, the commercial lending team spent an estimated 360,000 hours per year reviewing commercial loan agreements alone. Lawyers and loan officers — billing at $300–$500 per hour — performed rote document classification and data extraction that required no legal judgment: work that consumed capacity, created processing backlogs, and added significant cost without adding analytical value. Individual credit agreements required 8–12 hours of attorney time; non-standard provisions could delay loan closings by days.

JPMorgan Chase began developing COIN (Contract Intelligence) in 2016 as an internal machine-learning platform for legal document review. Key actions:

• ML model training: The technology team trained COIN on thousands of historical loan documents, enabling the system to identify and extract approximately 150 key data attributes (loan amounts, maturity dates, collateral provisions, interest rate clauses) with accuracy matching experienced attorneys.
• Full-volume deployment: COIN was deployed to process the bank's full volume of approximately 12,000 commercial credit agreements reviewed annually — the primary document category driving the 360,000-hour workload.
• Workflow integration: The platform connected directly to JPMorgan's existing document management systems. Standard documents were processed automatically; non-standard provisions were flagged and routed to human attorneys for judgment calls, keeping legal review focused on genuinely complex decisions.
• Scope expansion: After validating performance on commercial loans, the bank extended COIN to non-disclosure agreements, custody agreements, and eventually credit default swap documentation — multiplying the addressable automation volume.