Credit scoring is the use of a statistical or machine-learning model to turn information about a borrower, such as repayment history, existing debts and length of credit history, into a number that predicts how likely they are to repay. Lenders use the score to accept or decline applications, set prices and decide credit limits.
Why does credit scoring matter once AI agents act?
Because the rules around it are among the strictest that touch AI. The EU AI Act lists credit scoring of natural persons in Annex III as a high-risk use, with obligations for providers and deployers. GDPR Article 22 gives people rights around decisions based solely on automated processing with significant effects, which a credit decision usually has. In the US, ECOA and Regulation B require specific reasons when credit is denied, including when a model is used.
An agent near a credit decision can fall inside these rules without anyone intending it. If its output ranks applicants or feeds the decision, it may be part of the scoring system. Is a credit scoring agent high-risk under the EU AI Act? works through when that happens, and EU AI Act Annex III credit scoring and AI agents in lending covers the obligations.
How does credit scoring work in practice?
The model takes inputs from a credit bureau, the application and, for existing customers, account behaviour. It produces a score. A strategy, set by the credit risk team, turns the score into action through cut-offs: above one line approve, below another decline, in between refer to an underwriter. The model and the strategy are validated, monitored and changed under governance.
An agent's useful roles sit around this machinery: running a credit check, gathering application data, explaining which factors moved a score, and preparing referrals. It should not move cut-offs, override a decline, or produce its own score that bypasses the validated model.
What does that look like in an example?
A credit card issuer runs annual reviews on its book. Its agent pulls each customer's current bureau score and account data, flags accounts where the score has dropped sharply or utilisation has climbed, and drafts a recommendation: keep, reduce or review the limit. A credit analyst approves each change. The scoring model did the prediction; the agent organised the evidence; a person made the decision. The steps are in how to prepare annual credit reviews with an AI agent.
How is credit scoring different from a creditworthiness assessment?
A credit score is one number from one model. A creditworthiness assessment is the lender's whole judgement on whether a borrower is likely to repay, drawing on the score, income, commitments and circumstances. A good score does not settle the assessment, and a borrower with a thin credit file may be creditworthy without a strong score. Both appear in Annex III, which is one reason lenders should be clear about which one an agent is contributing to.
Which related terms should you know?
- Cut-off: the score at which the strategy changes the outcome.
- Automated decision: a decision made without meaningful human involvement, the focus of GDPR Article 22 and AI agents in credit decisions.
The limits on agents in lending more widely are in AI agents in banking.
Questions people ask
Is an agent that summarises credit files doing credit scoring?
Not by summarising alone. If its summaries are used to rank applicants or decide outcomes, treat it as part of the scoring process and govern it that way.
Can customers ask why their score led to a decline?
In many markets, yes. US lenders must give specific reasons, and GDPR gives people rights around solely automated decisions. Keep a record that lets you answer.