AI Underwriting
AI underwriting is the use of machine-learning models to decide whether to lend to, or insure, an applicant and on what terms, scoring risk from credit files, bank transactions and other data rather than from fixed rules alone. In consumer lending it remains a regulated credit decision: the lender must explain refusals and show the model does not discriminate.
- Also known as
- Machine-learning underwriting, ML underwriting, AI credit decisioning
- Used at
- consumer and small-business lending, buy-now-pay-later, credit cards, merchant underwriting by payment companies, and insurance pricing
- Typical data
- credit-bureau files, open-banking transaction data, application details and repayment history; for businesses, accounting and card-processing records as well
- EU
- the AI Act lists AI systems that assess individuals' creditworthiness or set their credit score as high-risk, with an exception for fraud detection, and does the same for risk assessment and pricing in life and health insurance
- US
- under the Equal Credit Opportunity Act and Regulation B, a lender that refuses credit must state the specific principal reasons, and those reasons must reflect the factors its model actually scored
- Data protection
- the EU Court of Justice ruled in December 2023, in a case about the German bureau SCHUFA, that calculating a credit score is itself an automated decision under the GDPR when lenders rely mainly on it
- Parties
- the lender or insurer, model developers and data providers, model-risk and compliance teams, and conduct and data-protection regulators
How AI underwriting works in practice
A traditional scorecard uses a dozen variables with fixed weights. An AI underwriting model — commonly gradient-boosted decision trees, sometimes neural networks — uses hundreds, including patterns in transaction data: regularity of income, gambling spend, days in overdraft, how balances move before payday. It is trained on past applications and their outcomes, tested on data it has not seen, and deployed behind a policy layer that still applies hard rules — affordability, exposure limits, sanctions — before any offer is made.
The gain is usually measured one of two ways: more applicants approved at the same loss rate, or the same number approved at a lower loss rate. The cost is governance. Each refusal needs reasons a customer can understand, so lenders run explanation methods over the model to extract the factors behind each decision. Each variable needs checking for whether it stands in for a protected characteristic; postcode and spending patterns can track ethnicity or age. And the model needs monitoring, because one trained in a single economic period can drift when rates or employment change.
For a buyer of a lender or of an underwriting-software company, the model is worth what its evidence shows: performance on later, out-of-time data, actual losses by vintage against what the model predicted, fairness testing, the logic that produces refusal reasons, and independent validation reports. A model that cannot be explained to a regulator is a liability, however well it predicts.
Example
A UK consumer lender replaces a 14-variable scorecard with a gradient-boosted model using 260 variables, 180 of them from open-banking data. On 50,000 historical applications held back for testing, the new model approves 41% at a predicted 12-month default rate of 3.2%, against 34% for the scorecard at the same rate. After a year live, approvals are up from 34% to 39% and defaults run at 3.4%. The fairness review removes two variables that tracked applicants' age too closely, costing half a point of approval rate. When the lender is sold, the buyer's team re-runs the model on the last six months of applications, checks in a 300-file sample that refusal reasons match the model's top factors, and prices the loan book on observed defaults rather than predicted ones.
Common mistakes
- Judging a model on test-set accuracy without tracking actual defaults by vintage. - Giving refusal reasons that the model did not actually use. - Removing protected characteristics and assuming the model is therefore fair; proxies remain. - Retraining only on approved applicants, so the model never learns about the people it turned away. - Leaving the model unmonitored through a change in interest rates or employment. - Letting the model override affordability rules instead of operating inside them.