AFINIT Wins Patent for Generative AI Loan Screening Method

The problem it addresses

Generative AI can interpret structured data such as income and credit scores, as well as unstructured material like financial statements and consultation records, which has made it a promising screening aid. But when a model outputs only the text "approve" or "reject", it cannot distinguish differences in risk among customers who are all approved. Setting loan limits or differentiating interest rates requires a continuous numerical measure of risk.

How it works

According to the company, a screening prompt is built from customer information, and the AI is prompted to answer approve, reject or review. The system then extracts the token log probability, which expresses how likely the model was to generate that word, and converts it into a rejection probability. That value is combined with a lender's historical delinquency data to train a statistical calibration function, separating raw text-generation probability from a customer's actual likelihood of default and producing a final default probability.

Lenders can use the result to sort approvals, rejections and manual reviews, or to set rates and limits by customer. When markets or regulations change, only the lightweight calibration function needs retraining, not the entire generative AI model, which the company says cuts operating cost and time significantly.

Global expansion

The same logic applies abroad. For countries with different languages, legal environments and loan products, the calibration function can be tuned using local delinquency data. Lee said the patent provides a technical foundation for linking qualitative AI answers to quantitative risk metrics, and that AFINIT will use its ability to reflect each country's and lender's standards to speed up its global expansion.