Regulation B covers business credit, so the effects test left your merchant files too. New Jersey's statute reaches any loan for whatever purpose, and it did not move.

Yes, and this is the part most coverage skipped. Regulation B implements ECOA, and its scope was never limited to consumers. The regulation defines business credit as "extensions of credit primarily for business or commercial (including agricultural) purposes," defines an applicant as "any person who requests or who has received an extension of credit from a creditor," and defines person to include a "corporation, government or governmental subdivision or agency, trust, estate, partnership, cooperative, or association."²⁰ Business applicants get modified adverse action notice mechanics keyed to a $1 million gross revenue threshold, but they are inside the regulation.

So the effects test that left Regulation B on July 21 left it for commercial files too, not only for mortgages and cards. If you read the July 21 coverage as a consumer story and concluded your own exposure was unchanged, the opposite happened: it moved, and it moved in your favor at the federal level while staying exactly where it was at the state level.

The same rule did two other things worth knowing. It narrowed what counts as discouraging an applicant, shifting the test toward statements showing intent, and it restricted for-profit special purpose credit programs from using race, color, national origin, or sex as eligibility criteria.^{4 5} Consumer advocates read the package as a substantial reduction in ECOA protection.⁶

One caveat specific to this audience, and it matters more than the rest of this section. Whether a given merchant cash advance is credit at all is contested, and providers have long argued that a true purchase of future receivables is not an extension of credit. Regulation B's coverage follows that characterization. If your product is credit, Regulation B reaches it. If your position is that it is not credit, that position is now carrying more weight in your compliance posture than it was carrying in June, and it is worth confirming with counsel that your documents actually support it.

This is the case worth reading closely, because it shows what a state enforcement theory looks like when the defendant is a nonbank lender using models rather than a bank. In an assurance of discontinuance dated July 10, 2025, Earnest Operations LLC agreed to pay $2.5 million to Massachusetts.^{8 22}

The Attorney General alleged three practices. Earnest allegedly assigned weighted subscores from a Cohort Default Rate variable reflecting the average default rate at an applicant's college, which the AG alleged produced disparate impact, with Black and Hispanic applicants more likely to be denied or offered worse terms. It allegedly applied a knockout rule automatically declining applicants without at least a green card. And underwriters allegedly overrode the models without clear standards or documentation.⁸ The AG's framing was that Earnest failed to mitigate the risk of disparate harms to Black, Hispanic, and non-citizen applicants from its use of AI underwriting models, under state consumer protection and fair lending law.^{22 23}

Two things matter for how you read this. Earnest denied the allegations and denied violating Massachusetts or federal law; the settlement resolved the matter without any admission.⁸ And the forward-looking remedy was governance, not a model ban. Earnest agreed to a governance structure for covered AI models setting out requirements for "written policies, risk assessments, testing, inventories, documentation, and an oversight team," plus ongoing compliance reporting to the AG's office.^{23 22}

An honesty note before you extrapolate. Earnest is a student lender, so this is a consumer credit matter, and we are not going to pretend a merchant file is the same file. What reads across is the theory rather than the asset class: a state attorney general, a nonbank defendant, model-driven underwriting, and a governance remedy rather than a prohibition. Two of the three alleged practices are model design choices a licensee would never see. The third, undocumented overrides, is a pure process failure sitting inside the lender, and it does not care what asset class you write.

New Jersey is the one we can answer precisely, because the statute says so on its face. N.J.S.A. 10:5-12(i) reaches any bank, banking organization, mortgage company, insurance company, other financial institution, lender, or credit institution "involved in the making or purchasing of any loan or extension of credit, for whatever purpose," and prohibits discrimination in the granting, withholding, extending, modifying, renewing, or purchasing of loans or credit, or in the fixing of rates.²¹ That phrase, for whatever purpose, is what carries it past consumer credit.

On top of that statute, the Division on Civil Rights adopted disparate impact rules effective December 15, 2025, building on January 2025 guidance addressing algorithmic discrimination.^{10 9} The guidance treats discrimination produced by automated decision-making tools as reachable in the same way as any other discriminatory conduct, and holds that a tool serving a legitimate interest is still prohibited where a less discriminatory alternative exists, even if that alternative requires somewhat more labor, time, and resources.^{9 11 25} Counsel reads the state as preserving the disparate impact theory for borrowers regardless of the federal retreat.¹²

Beyond New Jersey we will be careful. Husch Blackwell's assessment is that "state antidiscrimination and fair lending statutes in California, New York, Massachusetts, and other states still authorize effects-based claims. State attorneys general have already signaled they'll keep bringing them."⁴ That is a statement about effects-based claims generally, not a finding that each of those statutes reaches business purpose credit the way New Jersey's does. Whether your borrower's state gets there is a state-by-state question for your counsel, and the phrase to have them look for is the one New Jersey uses. The Fair Housing Act also retains disparate impact liability on its own footing.^{4 7}

It does not reach you unless you sell to the GSEs, and we will be exact about that before describing it. Lender Letter LL-2026-04 binds Fannie Mae sellers and servicers, which means mortgage originators. If you fund merchant advances, buy receivables, or write equipment paper, this letter creates no obligation you owe today.

Fannie Mae issued it on April 8, 2026, with an effective date 120 days later, which landed on August 6.^{19 1} Worth noting for anyone reading a causal story into the calendar: the letter predates the July 21 Regulation B rule and was published two weeks before that rule even appeared in the Federal Register. These are two independent developments that happen to land in the same quarter, not a sequence.

The obligation is documentary, not technical. A seller using AI or ML must produce written policies and procedures spanning development, implementation, use, and maintenance of those systems, refreshed at least annually, transparent and communicated to appropriate personnel, reflecting an understanding of legal and regulatory requirements, and reflecting risk management calibrated to the firm's risk tolerance.^{1 19} Vendor and subcontractor use of AI or ML must be subject to governance "no less protective" than the framework's own requirements.^{19 2} Fannie Mae also reserves the ability to ask what AI is being used, why, and what safeguards sit around it, and does not specify penalties for noncompliance.² It created an information right and left the remedy unstated.

Fannie Mae does not answer this. The letter uses "AI/ML" without a technical definition, and reporting on it describes the scope as broad: it does not distinguish internally developed tools from vendor-provided systems and does not limit itself to underwriting, capturing any AI or ML system used in origination or servicing.^{19 2} The undefined boundary is the most common complaint from mortgage lenders working through it, since most firms are using something that might qualify without having inventoried it.

For your own scoping we would not start from the technology label at all, because the label is the wrong question. ECOA and the state statutes attach to the credit decision and its outcomes, not to the sophistication of the tool that produced them. The Cohort Default Rate variable in the Massachusetts matter was not deep learning; it was a variable that correlated with race and carried weight in a decision.⁸ A hard-coded scorecard, a bank statement rules engine, and a licensed cash flow score all produce the same exposure profile as a neural network if they materially influence who gets approved and at what price. The practical filter is whether a system makes or materially influences a credit decision, and the inventory follows from that, not from whether the vendor calls it AI.

In the most exposed position, and it is worth being blunt about why. Buying a scoring product does not move the obligation to the vendor. Fannie Mae's framework pushes the seller's governance standard onto vendors and subcontractors rather than excusing the seller.^{19 1} New Jersey goes further and says it directly on the liability side: a covered entity may be liable if unlawful discrimination results from reliance on a vendor's product or system, and covered entities are expected to take reasonable steps to evaluate the design and testing of automated decision-making tools.^{25 9}

Put those together and the vendor contract is not a shield in either channel. A lender who cannot list the variables in a purchased model, or produce override logs, has a governance gap that no indemnity clause closes.

It says the models are being pointed at thinner files while this is happening. TransUnion's Q2 2026 Credit Industry Insights Report, released August 6, put unsecured personal loan originations at 6.4 million, up 19.5% year over year, with subprime originations up 29% against 9% for super prime. Balances reached $281 billion, up 9.6%. The 60-plus-day delinquency rate was 3.81%, up 44 basis points year over year.¹³ Note the reporting convention: the Q2 report carries first-quarter origination data, so the origination figures and the delinquency figure are not the same period.

TransUnion's Josh Turnbull framed it as lenders "reaching more consumers than ever, particularly at the subprime end, but they are doing it with smaller loan sizes and tighter underwriting," with subprime borrowers up 18.4%, subprime accounts up 20.5%, and average new subprime loan size down 6.8%.¹³ This is consumer data and we are not going to dress it up as a commercial print, since TransUnion's CIIR does not cover merchant advances or equipment paper. It is directional for one reason: growth concentrated at the thin end and executed through automated decisioning is the pattern that produces outcome disparities, and that pattern is not asset-class specific.

Three things are worth having on the shelf, and all three are cheap relative to assembling them under deadline. An inventory of every system that makes or materially influences a credit decision, including purchased and embedded tools, with a named owner for each. The variable list for those systems, obtained from vendors in writing, with specific attention to proxies that correlate with protected classes, which is the structural problem the Cohort Default Rate allegation describes.⁸ And an override log with a documented standard, since undocumented human overrides were a distinct count in the Massachusetts matter rather than a footnote to the model claims.⁸

On jurisdiction, the practical filter is where your borrowers sit, not where you sit. New Jersey's statute reaches credit extended for whatever purpose, which is the clearest commercial hook currently on the books.²¹

The DIY version needs a scheduler, per-state connections, a diffing layer, and permanent maintenance. That project never beats the roadmap.

Cobalt's Business Monitoring replaces it: enroll a business, set a 1 to 30 day cadence, and severity-ranked changes arrive at your webhook at one credit per check on your existing plan.

Schedule a FREE demo call

United Wholesale Mortgage reported a second-quarter net loss of $451.9 million on $888.0 million of revenue, against net income of $170.4 million in the first quarter. The swing was driven by a $603.2 million loss on interest rate derivatives, which the company attributes to portfolio hedging around its mortgage pipeline and secondary market positioning rather than to credit performance. Origination volume was $39.7 billion, flat against Q2 2025. The company said that "subsequent to June 30, 2026, the Company's Board of Directors determined to suspend its quarterly dividend," and announced "a $2.05 billion equity capital investment by Oaktree Capital Management and SFS Group Capital, LLC, a newly formed investment vehicle wholly owned by the Ishbia family."¹⁸ Shares closed at $1.19 after the announcement.^{14 15} A hedging loss rather than a credit loss, but the capital structure response is the part worth watching.

Edward Jones is on track to open its in-house bank in January 2027, built for lending rather than the retail checking it already offers through a U.S. Bank partnership. The firm holds FDIC and Utah Department of Financial Institutions approval for an industrial bank, and Andrea Moss will serve as president. David Chubak, head of wealth management and field management, said that "being able to get into some of those lending areas allows our advisors to have a full conversation on all of our clients' needs."¹⁶ The charter arc that ran through OppFi, Upstart, Klarna, and LendingClub now includes a distributor rather than a lender.

Serious auto delinquency, measured 60 or more days past due, stood at 1.33% in the second quarter, up just 2 basis points year over year, with TransUnion noting that the pace of deterioration has slowed.¹³ Subprime and near prime origination volume kept climbing over the same period.¹⁷ Set that against the unsecured personal loan print, where 60-plus delinquency rose 44 basis points to 3.81%,¹³ and the two asset classes are moving apart rather than together.

_{Sources
1 Fannie Mae | Lender Letter LL-2026-04, Governance Framework for the Use of Artificial Intelligence and Machine Learning, issued April 8, 2026
2 Cooley Finsights | Fannie Mae Issues AI/ML Governance Framework for Sellers and Servicers, commentary published April 24, 2026
3 Federal Register | Equal Credit Opportunity Act (Regulation B), CFPB final rule 2026-07804, published April 22, 2026, effective July 21, 2026
4 Husch Blackwell | CFPB Finalizes Major Regulation B Overhaul, April 24, 2026
5 Venable LLP | CFPB Makes Significant Changes to Regulation B
6 National Consumer Law Center | CFPB Guts Core Equal Credit Opportunity Act Protections in Regulation B
7 Norton Rose Fulbright | CFPB amends Regulation B, changing approach to fair lending
8 ABA Banking Journal | Massachusetts AG Reaches $2.5M Settlement With Earnest Operations Over AI Lending Bias, July 10, 2025 assurance of discontinuance
9 New Jersey Office of the Attorney General | Guidance on Algorithmic Discrimination and the New Jersey Law Against Discrimination, January 9, 2025
10 Consumer Financial Services Law Monitor | New Jersey Adopts Disparate Impact Rules Under LAD, effective December 15, 2025
11 Littler | AI in the Garden State: New Guidance on Algorithmic Discrimination and the NJLAD
12 Holland & Knight | N.J. Attorney General Preserves Disparate Impact Theory of Lending Discrimination for Borrowers
13 TransUnion | Q2 2026 Credit Industry Insights Report, released August 6, 2026
14 The Detroit News | UWM reports Q2 loss, suspends dividend, sets $2.05B capital plan
15 HousingWire | UWM raises $2.05 billion after $451.9 million Q2 loss
16 WealthManagement.com | Edward Jones Set for January Bank Launch to Widen Client Lending Options
17 Auto Finance News | Auto delinquency growth slows even as nonprime lending increases
18 UWM Holdings Corporation | Announces Second Quarter 2026 Results, company release
19 Orrick InfoBytes | Fannie Mae issues AI and machine learning governance framework, issuance date April 8, 2026, effective 120 days from publication
20 CFPB | Regulation B, 12 CFR 1002.2, definitions of business credit, applicant, and person
21 N.J.S.A. 10:5-12(i) | New Jersey Law Against Discrimination, lending and extensions of credit for whatever purpose
22 Massachusetts Attorney General | AG Campbell Announces $2.5 Million Settlement With Student Loan Lender for Unlawful Practices Through AI Use
23 DLA Piper | State action targets use of biased AI underwriting models, key points
24 Goulston & Storrs via National Law Review | Lenders Beware: Artificial Intelligence (AI) in Credit Agreements, March 2, 2026
25 Mayer Brown | Attention Lenders: New Jersey Issues Disparate Impact Regulations, December 2025}