AI-Assisted Credit Evaluation Models for Improving Risk Assessment Accuracy in U.S. Banking Systems

Abstract

This study investigated a persistent problem in U.S. banking credit decisioning: traditional scorecards and manual underwriting can produce inconsistent judgments and avoidable misclassification, especially when borrower profiles are complex and decision speed is high. The purpose was to quantify how AI-assisted credit evaluation, deployed within enterprise banking environments, improves perceived risk assessment accuracy and which enabling conditions most strongly drive those gains. Using a quantitative, cross-sectional, case-based design, data were collected via a structured 5-point Likert survey from n = 214 eligible banking professionals (usable response rate 71.3%) across underwriting (38.8%), credit analysis (27.1%), risk management (22.0%), and model risk/compliance (12.1%), representing enterprise-grade, cloud-supported decision workflows in the case banks. Key variables included Risk Assessment Accuracy Improvement (dependent) and five predictors: AI Model Capability, Data Quality and Availability, Explainability/Transparency, Governance and Compliance Alignment, and Monitoring and Drift Management. The analysis plan applied reliability testing (Cronbach’s α), descriptive statistics, Pearson correlations, and multiple regression. Measurement reliability was strong (α = .81–.90; DV α = .90). Descriptively, respondents agreed that AI improved accuracy (DV M = 3.97, SD = 0.63), with high ratings for data quality (M = 4.05) and governance (M = 3.94), while monitoring was lower (M = 3.72). Accuracy improvement correlated significantly with all predictors (r = .39–.56, p < .001), strongest for data quality (r = .56) and governance (r = .51). In regression, the model explained substantial variance (R² = .46; Adj. R² = .44; F(5,208) = 35.4, p < .001), with Data Quality (β = .29, p < .001), Governance (β = .22, p = .002), and AI Capability (β = .18, p = .006) as significant drivers; explainability was marginal (β = .11, p = .071) and monitoring was not significant after controls (β = .09, p = .104). Practically, the strongest perceived operational gain was improved underwriter consistency (M = 4.06), alongside reduced false approvals (M = 3.84), implying that banks realize the largest accuracy benefits when enterprise AI is paired with disciplined data pipelines and governance controls rather than model sophistication alone.