JPMorgan Chase & Co.

JPMorgan Chase's commercial banking division generated thousands of credit agreements, NDAs, and custody contracts each year, all requiring manual review by attorneys and loan officers to extract key terms and flag non-standard clauses. By 2016 this rote document classification and data-extraction work consumed an estimated 360,000 attorney and loan officer hours annually, at premium professional billing rates. Individual credit agreements took 8–12 hours of attorney time to review, and non-standard provisions could delay loan closings by days.

Starting in 2016, JPMorgan's technology team built COIN (Contract Intelligence), an internal machine-learning platform trained on thousands of historical loan documents to extract roughly 150 key data attributes — loan amounts, maturity dates, collateral provisions, interest rate clauses — with accuracy matching experienced attorneys. COIN was deployed against the bank's full volume of about 12,000 commercial credit agreements reviewed annually, integrated into existing document management systems so standard documents were processed automatically while non-standard provisions were routed to human attorneys. The bank later expanded COIN's scope to NDAs, custody agreements, and credit default swap documentation.

COIN processed in seconds the commercial loan agreement review work that had previously consumed 360,000 attorney and loan officer hours per year. The system also reduced data-entry errors in loan servicing, a prior source of compliance incidents and remediation costs, and attorneys freed from document classification were redeployed to higher-value advisory and litigation work.

• Automating rote classification and extraction — not legal judgment — was the key design choice: non-standard provisions were still routed to human attorneys, keeping COIN inside its competence envelope.
• The labor-hour savings and the error-reduction/compliance framing were both true simultaneously, which is why the business case held internally.
• Replication is bounded by resources most firms lack: a proprietary decades-deep training dataset, substantial annual tech investment, and joint General Counsel/CFO sponsorship to reallocate (not defend) headcount.