AI-Assisted Underwriting Models for Improving Risk Assessment Accuracy in U.S. Insurance Markets

Abstract

AI-assisted underwriting has gained prominence as insurers seek to improve risk assessment accuracy within increasingly complex U.S. insurance markets. This quantitative study evaluated the performance of AI-assisted underwriting models relative to conventional underwriting models using policy-level data from U.S. personal automobile and residential property insurance portfolios. The analytical sample comprised 48,620 underwriting observations across 18 U.S. states, with personal automobile insurance representing 64.7% of policies and residential property insurance accounting for 35.3%. Risk assessment accuracy was operationalized as a multi-dimensional construct encompassing discrimination, calibration alignment, loss sensitivity, and stability. Descriptive results showed that AI-assisted models achieved higher average discrimination (mean = 0.748, SD = 0.058) compared with conventional models (mean = 0.692, SD = 0.041), alongside improved loss sensitivity (0.721 versus 0.667). Calibration alignment increased from a mean of 0.914 under conventional models to 0.941 under AI-assisted models, while stability declined slightly from 0.884 to 0.861, indicating greater segment-level variability. Regression analysis confirmed statistically significant effects of AI-assisted models on discrimination (β = 0.056, p < 0.001), calibration alignment (β = 0.031, p < 0.001), and loss sensitivity (β = 0.049, p < 0.001), with adjusted R² values ranging from 0.32 to 0.41 across accuracy dimensions. Enriched data inputs produced additional gains in discrimination (β = 0.043, p < 0.001) and loss sensitivity (β = 0.038, p < 0.001), independent of model family. Reliability analysis demonstrated strong internal consistency for composite accuracy constructs, with Cronbach’s alpha values between 0.816 and 0.889. Overall, the findings provided quantitative evidence that AI-assisted underwriting models improved multiple dimensions of risk assessment accuracy in U.S. insurance markets, while introducing measurable trade-offs in performance stability across states and underwriting tiers.