AI could fundamentally change the future of bank lending — and make it a lot less risky.
As AI infiltrates every corner of the business world, banks have been experimenting with the technology for loan underwriting. That experimentation has caught the attention of academics. Harvard researchers conducted a case study on a California-based startup, Zest AI, which uses a machine learning model to assess credit risk for banks as an alternative to traditional measures (i.e. credit scores). Zest is used by more than 180 banks and credit unions, from big institutions like Freddie Mac to small, local ones across the U.S. Its competitors offering similar services include Pagaya Technologies, Chetu, Lender Toolkit, and Informed.IQ, among others.
Zest AI gained traction because “it was able to show that its credit risk model provided more accurate assessments of credit risk than standard credit scores available from credit rating agencies such as Equifax, Experian, and TransUnion,” Harvard’s David Scharfstein and Ryan Gilland wrote in their report. While their case study was reviewed and approved before publication by a Zest AI representative, it wasn’t funded by the company.
Credit unions and banks using Zest AI have seen a 25% increase in loan approvals, holding risk constant.
“So you’re not just saying yes to more people, you’re taking the same amount of risk, but able to say yes to more of your customers and members,” Mike de Vere, Zest AI’s CEO, told Quartz earlier this year.
That’s because, where credit scores might give a grainy, pixelated image of a borrower, AI models give a high definition, 3D video. De Vere said Zest uses “hundreds of variables” to determine loan approvals, whereas a credit score is a “blunt instrument that usually has 15 to 20 variables in it.” For example, the model uses proxies for debt-to-income ratio, a factor in traditional scoring models that doesn’t take into account gender pay disparities and results in lower loan approvals for women. It accounts for patterns, say, if someone has a late credit card payment during the holidays every year, but not any other time.
“While the Zest AI model categorized a significant number of applicants as low risk when the standard model categorized them as high risk, there were also cases where the Zest AI model categorized borrowers as high risk when the standard model would categorize them as low risk. Thus, the economic gain from applying the Zest AI model came from both expanding the pool of eligible borrowers and turning away risky applicants that might otherwise have been approved but defaulted.” – David S. Scharfstein and Ryan Gilland for the Harvard Business School Case Collection
In deploying its model, Zest is also showing how AI can expand access to personal, auto, home, and small business loans for people of color. Lenders using Zest AI saw loan approvals increase by 49% for Latinos, 41% for Black applicants, 40% for women, 36% for elderly applicants, and 31% for Asian American Pacific Islanders applicants.
That impact is more pronounced at some institutions, in particular. For example, Verity Credit Union in Washington state saw major rises in loan approvals for Black Americans (177%), people over 62 years old (271%), and AAPIs (375%).
Correction: A previous version of this article said Harvard conducted a study on Zest AI showing that it increased loan approvals by 25%. Harvard did not conduct an independent analysis; it reported on numbers shared to them by Zest AI.