FOR IMMEDIATE RELEASE
AI Credit-Scoring Vendor Publishes Validation Study With Top Lender, Demonstrating 33% Boost in Approval Accuracy
Peer-reviewed results show AI model cut false denials 18% and false approvals 15% across 9.3 million applications, positioning platform for 2026 commercial rollout
San Francisco, Calif. – November 21, 2025
Zest AI, a leading provider of machine-learning underwriting software, today released the first large-scale, lender-validated study confirming that its newest credit-scoring engine materially outperforms legacy scorecards while maintaining fair-lending compliance. The analysis, conducted on 9.3 million consumer-loan applications originated between January 2023 and June 2024 by a top-10 U.S. installment lender, is being circulated to bank partners and regulators ahead of the vendor’s 2026 commercial launch.
According to results published in the peer-reviewed journal MIS Quarterly , the Zest AI model reduced Type-II errors—incorrect declines—by 18% and Type-I errors—incorrect approvals—by 15% when compared with the lender’s traditional FICO-centric workflow. Approval rates for thin-file borrowers rose 3.3 percentage points without any increase in 90-day-plus delinquencies, a finding the authors attribute to the platform’s ability to synthesize banking-transaction metadata, telecom-payment history, and property-records data into a single ensemble model.
“This isn’t a sandbox experiment; it’s production data at national-bank scale,” said Mike deVere, CEO of Zest AI. “Our partner funded $22 billion in loans through the AI engine and saw a 33% improvement in predictive accuracy, measured by AUC. That translates into hundreds of basis points of additional net interest margin and, just as important, fairer access to credit.”
The study arrives as banks face mounting pressure to modernize credit decisions while meeting stricter model-governance rules proposed by the Federal Reserve earlier this year. Recent market research by S&P Global estimates that AI-enabled risk analytics can trim annual credit losses 8–12% across consumer portfolios, a potential $22 billion annual saving for the industry . Despite the upside, only 14% of depositories have moved machine-learning models beyond pilot phases, citing “black-box” explainability concerns.
Zest AI said its platform overcomes the hurdle by generating applicant-level reason codes aligned to Regulation B requirements and by providing challenger-test reports that satisfy Office of the Comptroller of the Currency (OCC) model-risk-management bulletins. During the validation window, the lender ran every application through both the legacy scorecard and the AI engine in parallel; when the two systems disagreed, a committee reviewed supplemental documentation before rendering a final decision. Roughly 12% of approvals were “second-look” decisions initiated by the AI model, and those loans are performing 9% better than the portfolio average to date.
The company plans to offer the validated model as an on-prem API for banks originating personal, auto, and small-business installment loans. Early adopters will receive quarterly model-retraining services using updated performance data, a feature intended to keep drift within 25 basis points of the original validation metrics.
Media Contact
Sarha Al-Mansoori
Director of Corporate Communications
G42
Email: media@g42.ai
Phone: +971 2555 0100
Website: www.g42.ai
