JPMorgan COIN: machine learning for commercial loan agreements

JPMorgan Chase uses a machine learning system called COIN to interpret commercial loan agreements. According to Bloomberg in 2017, the review work had taken lawyers and loan officers about 360,000 hours a year; COIN does it in seconds.

At a glance

Item	Detail
Company	JPMorgan Chase
Industry	Banking and finance
Use cases	Document understanding, contract review
Reported results	About 360,000 hours a year of manual work replaced; review in seconds; fewer loan-servicing errors. No investment cost disclosed

Case details

Background

Interpreting commercial loan agreements used to be done by hand by lawyers and loan officers, taking about 360,000 hours a year. According to the system’s designers, most loan-servicing mistakes came from human error in interpreting the roughly 12,000 new wholesale contracts the bank handles each year.

What they did

COIN stands for Contract Intelligence. It uses machine learning to find patterns and relationships in data in order to interpret contracts, and runs on Gaia, JPMorgan’s new private cloud platform. When Bloomberg reported on it in February 2017, the system had been live since the previous June. JPMorgan planned to extend it to other complex legal documents such as credit default swaps and custody agreements, and possibly later to interpreting regulations and analyzing corporate communications.

Results and data

Metric	As reported
Manual work replaced	About 360,000 hours a year (lawyers and loan officers)
Review time	Seconds
Contract volume	About 12,000 new wholesale contracts a year
Accuracy	Less error-prone than manual review; fewer loan-servicing mistakes

The same report noted JPMorgan’s technology budget that year was $9.6 billion, with about 20,000 developers.

What to take from it

• Start with high-volume, fixed-format documents. Commercial loan agreements are numerous and their clause types fairly stable, which suits a dedicated model doing bulk extraction.
• Count fewer errors as a return. Beyond hours saved, cutting the business mistakes caused by human misreading is a source of value in its own right.

Limits

• This predates large language models. COIN is a purpose-built machine learning model, so it does not show that a general LLM would achieve the same results.
• The 360,000 hours measure the manual effort replaced, not money saved or jobs cut, and the reports do not disclose what the system cost.
• All figures come from 2017 press coverage, with no public update on later performance. Companies with varied contract types and few samples should not benchmark against it directly.

Sources

Original source: Bloomberg · JPMorgan Software Does in Seconds What Took Lawyers 360,000 Hours Additional source: The Irish Times (via Bloomberg) · JPMorgan’s tech strategy turns to machine learning Additional source: ABA Journal · JPMorgan Chase uses tech to save 360,000 hours of annual work

This is an industry research entry: a publicly documented third-party AI case, not a Moments client case. For projects Moments delivered itself, see client cases.