How Lenders Use AI in Credit Decisions, and the Fairness Rules That Apply
Explainer
- Topic: How lenders use AI, and the fairness and explainability rules around it
- EU: The AI Act treats AI used to evaluate the creditworthiness of individuals as high-risk. The Digital Omnibus (Regulation (EU) 2026/1744, in force July 27, 2026) moved those obligations from August 2, 2026 to December 2, 2027
- US: Creditors must give applicants specific principal reasons for adverse action under ECOA and Regulation B
- Takeaway: Design for explanation, testing and human review now, whatever the deadline
Lenders increasingly use machine learning to decide who gets credit, on what terms and at what price. AI can find creditworthy borrowers that traditional scores miss, and it can also hide unfair patterns inside a model nobody can explain. This explainer sets out where AI is used in lending, the fairness risks and how the rules in the US and EU differ.
Where AI is used in lending
- Application triage and document reading.
- Identity and income verification.
- Credit risk scoring, often combining bureau data with cash-flow or other signals.
- Pricing and limit setting.
- Fraud detection, which has its own treatment under the EU rules (below).
- Collections, such as prioritising accounts and timing contact.
- Customer service through chatbots.
The fairness problem
- Proxy variables. A model can use data that stands in for a protected characteristic, such as location.
- Biased history. A model trained on past decisions can repeat past unfairness.
- Opacity. If nobody can say why a model declined an applicant, the lender cannot give a proper reason or check for bias.
United States: reasons still required
Under the Equal Credit Opportunity Act and Regulation B, a creditor taking adverse action must give the applicant specific principal reasons. The CFPB withdrew its 2022 and 2023 circulars on complex-algorithm notices in May 2025, but the underlying requirement remains, as Global Banking & Finance noted. The same report says the CFPB amended Regulation B in April 2026 to state that ECOA does not recognise disparate-impact liability. A separate review describes the same change and notes that state-level enforcement has moved in the other direction. The practical point for lenders: federal rules on reason-giving still apply, and states and courts may look at fairness independently.
European Union: high-risk, with a later date
- Classification. The AI Act lists AI systems intended to evaluate the creditworthiness of natural persons or establish their credit score as high-risk under Annex III, with an exception for AI used to detect financial fraud, according to Credisense.
- New timetable. Regulation (EU) 2026/1744, published on July 24, 2026 and in force from July 27, moved the high-risk obligations for stand-alone Annex III systems from August 2, 2026 to December 2, 2027. Transparency obligations still applied from August 2, 2026, as the same source notes.
- What already applies. EBA guidelines on loan origination (EBA/GL/2020/06) govern automated creditworthiness assessment today, according to BankingNewsAI, and EU consumer credit rules give borrowers a right to human intervention and an explanation.
- Obligations to expect. For high-risk systems: data governance, logging, human oversight, documentation and registration.
The delay is not a reprieve. A lender starting from zero will still need time to build documentation, testing and oversight.
What lenders should build now
- An inventory of every model and AI tool that influences a credit decision.
- Data governance: where training data comes from, how it is checked and which variables are used.
- Fairness testing across groups, including checks for proxy variables, where the law allows it.
- Reason codes that turn a model’s output into specific reasons a person can understand.
- A human review path for declined or borderline applicants, and for complaints.
- Logging and retention of inputs, outputs and overrides.
- Vendor oversight, because buying a scoring model does not transfer your responsibility.
- Monitoring for drift, so a model that worked last year is still fair this year.
A worked example: explaining a decline (illustrative)
Suppose a model declines a $5,000 loan application. The example is invented, but it shows what a useful explanation looks like.
- Weak: “Your application did not meet our criteria.” It tells the applicant nothing and may not meet the requirement to give specific reasons.
- Better: “Your card balances are high compared with your limits; your credit history is under two years; and a payment in the last six months was late.” These are specific, true and actionable.
To produce the second kind of message, a lender needs a way to translate a model’s output into the factors that actually drove the result. That is the job of reason codes, and it is much easier to design for them at the start than to add them later.
How to test a model for fairness
- Define the groups and outcomes you will compare, within what the law allows you to collect and use.
- Compare approval rates, pricing and limits across groups, and investigate large gaps.
- Look for proxies, which are variables that stand in for protected characteristics.
- Consider less discriminatory alternatives: could a different model or variable achieve similar accuracy with fairer results?
- Document the findings and decisions, including why a variable stays or goes.
- Repeat after changes and on a schedule, because data and behaviour drift over time.
Who is responsible for what
| Role | Typical responsibility |
|---|---|
| Board and senior management | Set risk appetite and ensure oversight and resources |
| Model owner | Accountable for performance, documentation and monitoring |
| Data science team | Builds, tests and explains the model |
| Compliance and legal | Interpret rules, review notices and fairness testing |
| Internal audit | Independent review of controls and evidence |
| Vendors | Provide documentation and access, but do not take over the lender’s duties |
EU AI Act key dates
| Date | What applies |
|---|---|
| February 2025 | Bans on prohibited practices and AI literacy duties |
| August 2025 | Rules for general-purpose AI models |
| July 27, 2026 | The Digital Omnibus on AI (Regulation (EU) 2026/1744) enters into force |
| August 2, 2026 | Transparency obligations begin to apply |
| December 2, 2027 | High-risk obligations for stand-alone Annex III systems, including credit scoring |
| August 2, 2028 | High-risk obligations for AI embedded in products covered by Annex I |
Frequently asked questions
Does using a vendor’s model shift responsibility to the vendor?
No. The lender still has to be able to explain decisions, test for fairness and oversee the model. Vendors should provide documentation and access, but the duty to the customer stays with the lender.
Is a system that detects fraud and scores credit exempt?
Only for the fraud function. One reference guide says a system that both detects fraud and evaluates creditworthiness is still high-risk for its credit-scoring function.
What are the penalties in the EU?
One analysis reports fines of up to €15 million or 3% of global turnover for non-compliance with high-risk requirements, with higher tiers for prohibited practices. Check the final text and national enforcement.
Can lenders use alternative data?
Yes, in many markets, but they must be able to justify it, test it for fairness and explain decisions that rely on it.
The Funding Journal view
The regulators’ directions differ, but the practical test is the same: can you tell a declined customer why, and can you show a supervisor you checked the model for unfair outcomes? Lenders that can answer both quickly have an advantage whichever rulebook applies. Those that cannot will find the explanation problem harder to solve after launch than before.
Related reading
- How to Choose a Digital Lending Platform: the modules, pricing and compliance questions to compare.
Sources and further reading
- Credisense: Europe moved its AI deadline for credit scoring
- AI Act reference: AI credit scoring under the EU AI Act
- Avanto: The AI Act timeline moved (Regulation (EU) 2026/1744)
- Global Banking & Finance: How AI is rewriting lending decisions
- BankingNewsAI: AI credit underwriting and lending
This article is general information, not financial, legal or tax advice. See our disclaimer. Facts last checked October 26, 2026; rules, prices and companies change, so confirm current terms before you act.
Published by The Funding Journal Editorial. Practical, sourced writing on fintech, crowdfunding, investment platforms, SaaS and AI, and the stories of founders who raised capital.
