A model can be consistent and still decline people for reasons nobody chose. Three industry voices explain how lenders turn an AI credit decision into a reason a customer and a regulator will accept.

Machine learning now sits behind a large share of credit decisions. Most customers never see the model. What they see is a short letter saying they were declined, and sometimes why. As a result, explaining AI loan declines has become one of the hardest jobs in consumer lending.

So we asked industry leaders one question: if you use AI or machine learning in lending decisions, how do you explain a declined application to a customer or a regulator? We wanted first-hand experience over commentary. Three respondents answered. Below, we set their replies on AI loan declines against the rules in the US and the UK.

AI loan declines and the threshold problem

Our first respondent runs credit risk analytics at a US consumer lender and recently published research testing AI models against human credit professionals.

“I recently released research comparing AI models with human underwriters on the same loan files, and it changed how I think about this question. I gave 17 loan files to 61 experienced credit professionals and to eight current large language models from three major AI labs, ten runs each. On borderline files the models were far more consistent than the humans. Their agreement was about 0.60, against 0.15 for the human panel, which is barely above chance. But the models approved 79% of the borderline files, where humans approved 46%. Asking a model to act like a 25-year veteran barely changed that. The models were consistent, but consistent at an approval threshold nobody had chosen.

That’s the explainability problem in one sentence. Regulators and customers aren’t asking why the model produced a score. They’re asking why you declined this person. If the model’s threshold has drifted from your credit policy, a fluent reason code doesn’t help, because it describes the model’s logic, not your policy.

In practice, check the threshold before you explain it. Compare the model’s approval rate on borderline cases with your own policy’s; in my study the gap was more than 30 points. Explain the decision, not the model: the notice should state the specific reasons the applicant fell below your cutoff. Don’t use AI-written rationales as decline reasons, because a language model will write a convincing explanation for any decision. And when an underwriter overrides the model, record the underwriter’s reason at the time.”

Ashutosh Bhargave, Senior Director, Credit Risk Analytics and Portfolio Management

His figures on AI loan declines match his paper, “Consistent But Lenient“, posted on SSRN in October 2026. The paper adds that the models reversed the human majority on three of nine borderline files. Their implied approval threshold also varied by 26 points between AI labs. However, it is a preprint, not a peer-reviewed study. It also tests general-purpose language models, while most lenders still decide with purpose-built scoring models.

Reason codes in the US

The US rule is clear and still in force. Under Regulation B, a lender that declines credit must give the specific principal reasons. Those reasons must reflect the factors the lender considered, and the rule applies to AI loan declines too. Our second respondent built consumer lending systems at LendingClub, now Happen Bank, before joining Amazon.

“You never explain a decline with the model. You explain it with reason codes. The model produces a score, and a separate layer maps the features that pushed that score down into a few readable reasons, the adverse action codes that go on the notice. In the US that’s an ECOA and Reg B requirement, not a nice-to-have. The customer gets ‘your revolving balance is high relative to your limits’, not ‘gradient boosting node 47 fired’.

The hard part is keeping those reasons honest. They have to reflect what the model did, not a plausible story added afterwards. As we moved to more complex models, the gap between the score and an explanation a person could act on got wider, and closing it was real engineering work.

For a regulator the bar is higher. They want to see the decision was consistent, documented and not using a protected class as a proxy. That means versioning the model, logging the inputs for every decision and testing outcomes across groups. Explainability is mostly a data and systems discipline, designed in from the start, not rebuilt when a complaint arrives.”

Shubham Kumbhalkar, Software Development Engineer II, Amazon

The federal backdrop for AI loan declines has shifted, though. In May 2025 the CFPB withdrew guidance that had said complex models do not excuse vague reasons. In April 2026 it also removed disparate impact from Regulation B, effective 21 July 2026. Still, the duty to give specific reasons for AI loan declines remains. States are filling the gap too. Massachusetts reached a $2.5 million settlement with student lender Earnest in 2025 over its AI underwriting models, which Earnest denied breached the law.

The UK has no reason-code rule

Our third respondent framed the problem through UK rules.

“Explainability in lending depends on mapping every automated decline to specific, readable reasons at the point of decision. Relying on complex models without a dedicated attribution layer creates real regulatory risk. The system should show which variables, such as a high debt-to-income ratio or many recent credit searches, most influenced each outcome, so lenders can move away from generic rejection messages.

Under the UK Consumer Duty, transparency also means keeping a human involved to check model logic and test outcomes across groups, so the AI isn’t using variables that act as proxies for protected characteristics. A common mistake is treating the explanation as a justification written afterwards rather than a live output of the decision. The best explanation for a regulator shows the model’s logic is consistent, documented and based on factors a human underwriter could defend.”

Abhishek Pareek, Founder and Director, Coders.dev

His advice on attribution is sound, but UK rules on AI loan declines differ from the US. UK lenders do not have to give reasons for a decline. Instead, a lender that declines on credit reference agency data must name the agency. The Consumer Duty expects firms to show good outcomes and clear communication, but it does not prescribe reason codes or a human reviewer.

Rules on automated decisions, including AI loan declines, have changed too. Since 5 February 2026, new UK GDPR rules allow fully automated decisions, but people must be able to make representations, get human intervention and contest the outcome. In October, the FCA also warned that a nominal human reviewer is not a control in itself. Firms must show meaningful review and challenge.

Less detail can be better

More explanation of AI loan declines is not always better. In an FCA experiment with 8,860 UK adults, extra detail about how an algorithm worked raised people’s confidence in challenging decisions. However, the regulator found it could also lead to worse decisions. For lenders explaining AI loan declines, the goal is a reason a customer can act on, not a tour of the model. For a sense of how other lenders use AI in credit work, see our report on TD Credit Assist.

How to explain AI loan declines

Taken together, the answers point to a short order of work for AI loan declines. First, set and check the approval threshold against credit policy before launch. Next, generate reason codes from the scoring model itself, not from an AI-written summary. Then log every input and model version, and test outcomes across groups even where federal law no longer requires it. Finally, give UK customers a real route to human review.

What to watch next

Several changes will shape AI loan declines in 2027. The EU AI Act treats credit scoring as high-risk, with obligations now due from 2 December 2027. In the UK, the ICO’s final guidance on automated decisions is expected this winter. Pricing transparency is also under scrutiny, as our report on Revolut’s UK personal loans showed.

Above all, the lesson on AI loan declines from our respondents is consistent. Explain the decision, not the model. A reason that matches policy will satisfy a customer and a regulator far better than a fluent account of what an algorithm did.

FintechBits covers payments, banking and fintech developments for readers in the US and UK. This article is for information only and is not financial or legal advice. Views expressed are those of the FintechBits editorial team.