AI in lending: what the enforcement record requires before the speed claims count (2026)

If you run lending operations, you are being sold speed. Every vendor deck promises fewer steps, faster approvals and a model that decides in seconds.

The speed is probably real somewhere. The problem is that none of it has been independently measured, while the AI lending outcomes a regulator actually examined are about something else entirely.

Take this month’s two headline claims. Colony Bank told American Banker that generative AI cut its SBA lending process “from 92 steps to 32” [source]. On its Q3 call, PayPal said its AI risk platform “nearly halved credit decision times in Canada, Italy and Spain” ([source]).

Both are the company describing its own system, relayed by a single outlet. Nobody outside either firm has measured them. TIN has not checked them either, and this post treats them as claims.

What AI in lending means

AI in lending is the use of machine-learning and generative models to make or support credit decisions: who gets a loan, on what terms and at what price, and how the paperwork behind each loan is read and processed.

The first half of that definition is where the legal exposure sits. A model that reads documents faster is an operations tool. A model that decides who is approved is a credit decision, and credit decisions carry a duty to explain.

What is required when using AI in lending decisions?

In the US, a lender using AI in lending decisions must still tell a denied applicant why. Regulation B requires that the statement of reasons for adverse action “must be specific” and indicate “the principal reason(s) for the adverse action” [source].

The same rule closes the obvious escape hatch. Statements that the denial rested on “the creditor’s internal standards or policies,” or that the applicant failed “to achieve a qualifying score,” are “insufficient” [source]. A model output is not a reason.

In 2022 the CFPB spelled out that the notice requirements “apply equally to all credit decisions, regardless of the technology used to make them” [source]. It withdrew that circular, and its 2023 follow-up, on May 12, 2025 ([source]).

Read that withdrawal carefully. It removed the Bureau’s commentary, not the regulation. The specific-reasons text above is still in Regulation B, and state attorneys general enforce fair lending law on their own, as the Earnest case below shows.

In the EU the bar is set by category. The AI Act lists “AI systems intended to be used to evaluate the creditworthiness of natural persons or establish their credit score” as high-risk [source], with an exception for fraud detection.

AI use cases in lending, and what the record shows for each

The AI use cases in lending that appear on the public record fall into four groups. What separates them is not the technology but whether anyone outside the lender has looked at the result.

Notice where the outside scrutiny landed. The two rows a regulator examined are both credit decisions. The two efficiency rows rest on what the lender said.

How is AI credit decisioning examined when it goes wrong?

AI credit decisioning is examined the way any credit decision is: by asking who the model disadvantaged and whether the lender tested for it. Earnest is the case where the answer was no.

The Massachusetts AG alleged that Earnest’s AI models made “determinations on loan applicants’ eligibility and loan terms and pricing,” and that the firm failed to take reasonable measures, “including failing to test its models for disparate impact” [source].

Two features carried the case. A subscore tied to the default rate at the applicant’s college left “Black and Hispanic applicants more likely to be penalized than White applicants,” and a “Knockout Rule” denied applicants based on immigration status [source].

Earnest paid $2.5 million [source] and agreed to discontinue both features ([source]). These were allegations resolved by settlement, not a court finding of discrimination. The full record is in the Earnest case file.

Can an AI lending model pass a regulator’s test?

Yes, and the Apple Card is the proof that it can. In 2019 a public complaint alleged its credit limits discriminated by sex, and New York’s regulator went after the underwriting itself.

The Department’s finding rested on “analysis of underwriting data for approximately 400,000 New York State applicants” [source]. It found “women and men with equivalent credit characteristics had similar Apple Card application outcomes” ([source]).

The clearance was not a clean bill. The regulator found “deficiencies in customer service and transparency,” and the firms dropped a rule making applicants wait six months to appeal credit terms [source]. See the Apple Card case file.

How is AI used in commercial lending?

In commercial lending, the best-known use of AI is reading loan documents, not deciding on loans. JPMorgan’s COIN classifies clauses in commercial credit agreements.

Bloomberg reported in 2017 that the work COIN took over had consumed 360,000 hours a year of lawyer and loan-officer time [source]. That figure has one origin, the bank, and has never been independently measured or updated.

That is the honest reading in the COIN case file: the most-quoted AI efficiency number in lending is nine years old and single-sourced. It is still quoted as if it were a benchmark.

What TIN’s case files show

Three TIN case files bear on AI in lending, and they split cleanly. The two credit-decision cases were examined by a regulator with power to compel data. The one efficiency case was never examined by anyone.

The tempting reading is that AI in lending is a speed story and fair lending law is a tax on it. The record says the opposite. Apple Card and Earnest both ran algorithms on consumer credit. One came through an examination of about 400,000 applicants; the other paid $2.5 million [source].

The difference was not the model. It was whether the lender could show what the model did to whom. Testability decided both outcomes, and it is the one property no speed figure tells you about.

The bottom line

The requirement for AI in lending is not a compliance checklist bolted onto a faster model. It is the product. Regulation B requires specific reasons for every denial, so a credit model whose denials you cannot explain fails the one test that matters, whatever it saves in steps.

So judge a lending AI pitch by the question a regulator will ask, not the one the vendor answers. A decision you cannot explain is not faster.

It is a liability with a delay on it.