AI in credit decisioning

Credit decisioning uses AI to assess a borrower: a default probability, a credit decision, or a price, from application and bureau data, and for some lenders from alternative data as well. The model scores; a person grants, denies or prices, or the score is applied directly. Here it is through the four lenses, the data, the model, the task, and evaluation and testing, the same way you would break down any use case.

The data

Application data, credit-bureau history, account behaviour, and for some lenders alternative data such as transactions, device or telco signals. The questions to ask of it: is it representative of the people the model will actually meet; does it carry historical bias from past lending decisions; does it contain, or proxy for, a protected attribute such as postcode or education; and what are the provenance and rights for each source. In credit, the data is where most of the risk enters.

The model

Usually tabular machine learning, such as gradient-boosted trees or logistic regression, not a generative model, because the task is a structured prediction from structured data and a generative model does not improve a tabular score. The choice that matters most is transparency: a simpler, more explainable model can be worth a small accuracy cost when you have to explain a declined application to the customer and to a regulator.

The task

A score-or-classify task feeding a decision. The model ranks or scores applicants; a person grants, denies or prices. The distinction that changes everything is where it sits between support and action. If a person genuinely re-decides with the information to overturn the score, the AI supports the decision. If the score is applied with a sign-off that no one really exercises, the AI makes it. Same model, very different use.

Evaluation and testing

The weakest link is the score itself, and the costs are asymmetric: a false approval is a credit loss, a false decline is a lost customer and a possible fairness problem. So test beyond overall accuracy. Measure performance group by group, with a fairness definition you have chosen and can defend. Check calibration, so a stated probability means what it says. Check for proxy variables. Confirm you can explain a single decision, not just the model in general. And re-test as the borrower population changes, because a model built on one population decays on another.

Judging the materiality

This page does not assign a risk tier, on purpose. How material credit decisioning is depends on how you deploy it: who it affects and how many, whether a person truly re-decides or only approves, whether the decision can be reversed, and your fairness and regulatory exposure. The same model can be a high-materiality use in one firm and a lower one in another. Put the four lenses above against your own deployment, rate it on impact, complexity and reliance as the AIRG expects, and let the intensity of the controls follow your own rating. For how that rating is done, see use-case risk and materiality.

Related on this site: AI bias and fairness, explainability and opacity, over-reliance and automation bias, the full AI risks taxonomy, and the guide for risk and control functions.

In an ebook: AI Risk Management for Risk and Compliance.