Domain Atlas / Lending & credit collections AI
Wells Fargo refinance underwriting & the bridge nobody could build
Explore this deployment in the PAN Lab ↗
In the PAN Lab, the readouts of this case's model organization carry a shaded evidence band whose width follows the least-established class among the modeling inputs the readings rest on.
The least-established input behind this case's model organization's readings is an assumption, not a measurement. Evidence base: 1 assumed · 9 published baseline.
The public class-certification order of 5 August 2025 in In re Wells Fargo Mortgage Discrimination Litigation recites, on facts the plaintiffs did not dispute, a three-part home-lending stack that is not an artificial-intelligence system. CORE is a front-end workflow tool that 'guides the flow of the loan origination process' and 'retrieves, stores, and displays' application data, Risk Engine outputs and final lending decisions, and it 'is not an automated underwriting system and it does not make lending decisions or calculate or assign Credit Risk Classes.' A separate application, the Risk Engine, assigns Credit Risk Classes using business rules, business services and data attributes; it 'currently executes tens of thousands of separately identifiable business rules' grouped into 16 business services, the largest of which, Get Risk Decision, 'contains 14,000 rules' and '1,300+ data attributes.' ECS comprises exactly two scorecard models, 11419 for government loans and 11960 for conventional loans, each of which per the bank's declaration 'is a simple scorecard whose logic can be fully stated on 2-3 sheets of paper,' 'neither utilizes artificial intelligence,' and each is 'simple enough that an applicant's score could be calculated by hand'; they map credit-bureau attributes to a score and thence to a Credit Risk Class and also generate risk insight messages displayed to underwriters. Wells Fargo's position, accepted as undisputed for certification purposes, is that these are distinct applications and that 'there is no such thing as CORE/ECS.' Underwriters 'are instructed that the Risk Engine never supersedes the judgment of the underwriter' and '[u]nderwriting is done by human underwriters.' Separately, Bloomberg News reported on 11 March 2022, from an analysis of Home Mortgage Disclosure Act data covering roughly 8 million 2020 refinance applications, that Wells Fargo approved 47 per cent of Black homeowners' completed 2020 refinance applications, 53 per cent of Hispanic or Latino homeowners' and 72 per cent of white homeowners' — the largest racial gap among major lenders and the only major lender to reject more Black refinance applicants than it approved. Those figures are carried here because a Senate letter of 16 March 2022 and a Senate Banking Committee release of 17 March 2022 independently restate them; the article itself was not retrievable for verification and no claim is made about what its analysis controlled for. Wells Fargo did not dispute the arithmetic of the counts, attributed the gap to 'additional, legitimate, credit-related factors' including credit scores, home appraisals and broader economic inequity, and a spokesman said the analysis was 'designed to present a skewed picture of our lending efforts.' That attribution has never been tested by any adjudicator, and no court and no regulator has ever found that Wells Fargo discriminated in refinance underwriting.[5]
What happened
On 11 March 2022 Bloomberg News published an analysis of Home Mortgage Disclosure Act data covering roughly eight million 2020 refinance applications. It reported that Wells Fargo had approved 47 per cent of Black homeowners' completed refinance applications that year, 53 per cent of Hispanic or Latino homeowners' and 72 per cent of white homeowners' — the largest racial gap among major lenders, and the only major lender that rejected more Black refinance applicants than it approved. The bank did not dispute the arithmetic. It attributed the gap to "additional, legitimate, credit-related factors" — credit scores, home appraisals and broader economic inequity — and a spokesman told Fortune the analysis was "designed to present a skewed picture of our lending efforts." That attribution has never been tested by any adjudicator.
Read the figure with two cautions attached, because both matter downstream. The 72 per cent is Wells Fargo's own white-applicant rate; it is not the roughly 70 to 71 per cent Black-applicant approval rate reported across all OTHER lenders in the same coverage, and the two are frequently confused. And the public loan-level file those counts come from excludes the applicant credit score by rule, so no analysis built on it can control for the very factor the bank named first.
Six senators wrote within a week. On 16 March, Senators Elizabeth Warren and Ron Wyden demanded from the chief executive, by 28 March, "All data and algorithms used by Wells Fargo to evaluate applications to refinance residential home mortgages from January 1, 2010 to December 31, 2021," together with any internal analyses of the causes of the low approval rates, any remediation plans, and all communications with regulators about them. The next day eleven senators led by the Banking Committee chair asked the Secretary of Housing and Urban Development and the Director of the Consumer Financial Protection Bureau to review the bank's 2020 and 2021 refinance processes for compliance and to "take appropriate action if disparities are found." On 13 April the bank announced a $150 million Special Purpose Credit Program to lower rates and refinancing costs for Black homeowners it already serviced, plus $60 million in WORTH grants aimed at roughly 40,000 homeowners of colour in eight markets through 2025 — a programme announced a month after publication, conceding no disparity in underwriting, with operator-reported and unaudited figures, and with no change to any underwriting rule announced alongside it.
Six private actions were consolidated in January 2023 before Judge James Donato. The operative complaint pleaded the system as the wrong: a decision "to employ centralized, universal, race-infected lending algorithms to differentially assess, delay and ultimately reject residential lending applications," implemented through a "centralized pioneering automated underwriting system -- sometimes referred to as CORE -- without sufficient, or sometimes any, human supervision or involvement." That is the pleading's language and none of it was adjudicated.
Then the class-certification record reframed the object, and this is the part worth slowing down for.
On the undisputed facts the court recited from the bank's own sworn declarations, CORE — which the bank's own filings expand as Common Opportunities Results Experiences — is a front-end workflow tool that "guides the flow of the loan origination process" and "retrieves, stores, and displays" application data, Risk Engine outputs and final lending decisions — and it "is not an automated underwriting system and it does not make lending decisions or calculate or assign Credit Risk Classes." A separate application, the Risk Engine, assigns those classes using business rules, business services and data attributes. It "currently executes tens of thousands of separately identifiable business rules" across 16 business services; the largest, Get Risk Decision, "contains 14,000 rules" and "1,300+ data attributes." One example rule appears on the record: a minimum credit score of 700 for retail, non-conforming, fixed-rate purchase loans up to $3,000,000. And "ECS" is two scorecard models, 11419 for government loans and 11960 for conventional. Each, per the declaration, "is a simple scorecard whose logic can be fully stated on 2-3 sheets of paper"; "neither utilizes artificial intelligence"; each is "simple enough that an applicant's score could be calculated by hand." They read credit-bureau attributes, output a number that maps to a risk class, and also generate risk insight messages displayed to underwriters. The bank's position, which the plaintiffs did not take issue with, is that these are distinct applications and that "there is no such thing as 'CORE/ECS'."
So the object at the centre of a four-year discrimination case is not a machine-learning model. It is a very large deterministic rule base, two hand-computable scorecards, a workflow screen, and a person. Fannie Mae's Desktop Underwriter and Freddie Mac's Loan Product Advisor run alongside, returning their own determination on the same application; the proposed classes were eventually narrowed to minority applicants approved by one of those external systems, or by the bank's own scorecard, and ultimately denied.
The human is the hinge, and the hinge turns both ways. The bank's declarant states that "underwriters are instructed that the Risk Engine never supersedes the judgment of the underwriter" and that "[u]nderwriting is done by human underwriters." The plaintiffs did not dispute it; they argued disparate impact survives "some discretion in the process." They also alleged — and every word of this is allegation from a case that ended without a merits ruling — that the discretion had been hollowed out: loan processors previously expected to handle about 30 applications a month later handling "more than 50 and sometimes nearly 100," sometimes "as many as three times the normal monthly volume"; processors and underwriters terminated or departed and not replaced; staff "systematically disincentivized to check the work"; and system changes removing the ability to make in-system adjustments that would raise approval likelihood.
The expert exchange is the centre of the file, and the two sides agreed. Plaintiffs' expert Dr. Amanda Kurzendoerfer ran a regression controlling for key underwriting factors and found statistically significant approval-rate disparities favouring white applicants that "cannot be explained by legitimate underwriting factors." At a concurrent expert hearing in February 2025 — both statisticians in the same session, on the same questions — Wells Fargo's expert Dr. Marsha J. Courchane "did not substantially disagree with the logistic analysis ... or with her findings of a statistical disparity along racial lines." Her objection was causal: without reviewing individual files you cannot tell whether any particular denial was wrongful, and "you will almost always find an underwriting disparity, on average, for minorities." Neither report, neither specification and none of the underlying data is public. The expert record is sealed.
Where the plaintiffs got closest to the machine, they were pointing at the bank's own documents, and it counted against them. The order records their characterisation that Wells Fargo "could not itself identify what was driving the disparity" and "found that there were three potential proxies for race in its ECS Model — major derogatories, average months in file, and recent inquiries — which could be causing the disparity." The court held that this cut against the plaintiffs, because the burden of showing class-wide causation was theirs. The bank's inability to explain its own outcome was, in law, the claimants' problem.
The bank had published on exactly this hazard. In May 2021, six members of its Corporate Model Risk group posted "Bias, Fairness, and Accountability with AI and ML Algorithms," warning that historical data skew and automated feature engineering can "miss the potential for correlated surrogate variables causing proxy discrimination," that black-box algorithms have "potential for serious harm" in consumer lending, and that models "must be continually monitored for disparate impact testing." The plaintiffs quoted it back at the bank in the operative complaint. What that monitoring produced is the fact the record does not settle.
On 5 August 2025 Judge Donato denied class certification for failure of Rule 23(a)(2) commonality. Numerosity had been conceded on the plaintiffs' own estimate of "at least 119,100" members. The court wrote that plaintiffs "did not present any classwide evidence whatsoever of robust causality" and had "focused like a laser on the statistical disparity in application denial rates, without anything in the way of explanatory factors." Then the sentence this whole file turns on: "the fact that Wells Fargo potentially analyzed 1,300+ data attributes pursuant to 14,000 rules indicates that commonality with respect to denial of a mortgage is not at all obvious here." The articulation of the system became the defence. Resting on Wal-Mart Stores v. Dukes, the court also treated the undisputed presence of human discretion as bringing the case within the post-Dukes line refusing to certify discretionary-policy lending classes: "Without some glue holding the alleged reasons for all those decisions together, it will be impossible to say that examination of all the class members' claims for relief will produce a common answer to the crucial question why was I denied." It reached no other Rule 23 element and made no merits finding.
Both Rule 23(f) petitions were denied on 8 January 2026. The case then ended without any ruling at all. On 7 May 2026 interim lead counsel filed a Notice of Settlement covering four named plaintiffs "and any and all parties represented by Counsel that were potential class members"; the court granted dismissal with prejudice and relieved interim lead counsel on 14 May. Three further plaintiffs were dismissed with prejudice between 8 May and 1 June. Only Christopher Williams' individual claim remains, stayed by order of 28 July 2026 through 25 September, with the court noting that "further requests to continue the stay are not likely to be granted." Terms are confidential and no amount is public; the notice extinguishes claims without any concession. The summary-judgment motion was never decided.
No regulator has made a finding here, and the absence is documented rather than assumed. The senators' March 2022 request to two agencies produced no publicly announced investigation, enforcement action or finding. Wells Fargo's Form 10-K for fiscal 2025 records the litigation sequence — "In August 2025, the district court denied class certification and plaintiffs' interlocutory appeal of the decision was denied in January 2026" — and discloses no government proceeding or investigation concerning mortgage-lending discrimination, while disclosing several unrelated government matters. Two near-misses must not be folded into that. The December 2022 consumer-bureau consent order at the same bank, $3.7 billion, terminated in early 2025, covered auto lending, deposit accounts and mortgage rate-lock extension fees — a separate matter. And a 2023 press account of a supervisory Matter Requiring Attention concerns discretionary mortgage PRICING exceptions rather than underwriting approval; it is supervisory, non-public in its particulars, and could not be verified at source.
One more separation. This is the 2022 to 2026 underwriting-approvals litigation. A differently captioned 2011 case and a 2012 federal settlement at the same bank concerned discretionary pricing and subprime steering; the 2011 case appears in the 2025 order as precedent against certification, not as a related claim.
The independent literature does not close the question in either direction, and leaving it out would imply that it does. Bhutta, Hizmo and Ringo, working from the confidential regulator-held file that DOES contain credit score, find that minority applicants have significantly lower credit scores and higher leverage, are less likely to receive approval from race-blind government automated underwriting systems, that observable applicant-risk factors explain most racial disparities in lender denials, and that differential treatment has played a limited role in recent denial disparities — a finding about the industry that speaks to no single lender's internal stack, and one the plaintiffs themselves cited in their complaint. Bartlett, Morse, Stanton and Wallace find that algorithmic lenders discriminate less than face-to-face lenders in pricing while Latinx and Black borrowers still pay measurably more.
What is left is a system nobody outside a protective order has read, a measurement everybody agrees on, a cause nobody decided, and a 10-K that discloses the litigation and no proceeding.
The sociotechnical reading
Most files in this atlas turn on what a system did. This one turns on who was allowed to find out, and it is worth being exact about the difference, because it changes what the file is evidence of.
Nothing here is a finding. No court and no regulator has found that this bank discriminated in refinance underwriting. No rule, attribute, weight or scorecard column has been identified as defective by anybody. What exists is a MEASUREMENT that both sides' statisticians accepted, a CAUSE that no adjudicator decided, and a proceeding that ended on a question about proof architecture. That is not a hedge; it is the content.
Start with the object, because the popular description of it is wrong in a way that matters. On the bank's own sworn account, accepted as undisputed, the thing at the centre of this case is not an artificial-intelligence model. It is a workflow screen that decides nothing, a rule base of tens of thousands of separately written rules, and two scorecards a person could compute with a pen. Almost every intuition people bring to algorithmic discrimination — opaque weights, learned proxies, a model nobody can explain — has to be set down at the door. This system is, in principle, the most explainable kind of automated decision there is. Every rule was written by somebody. Every threshold is a number a person chose.
And it made no difference at all, because explainability in principle and legibility in practice are different properties. Fourteen thousand rules over thirteen hundred attributes is a rule base no human reads end to end. When the court reached for a reason why commonality was not obvious, it reached for exactly that: "the fact that Wells Fargo potentially analyzed 1,300+ data attributes pursuant to 14,000 rules indicates that commonality with respect to denial of a mortgage is not at all obvious here." A system with a rule for everything has fourteen thousand places for a defect to sit, and the size that makes each rule changeable is the size that made the claim unprovable. This is the atlas's cleanest demonstration that transparency is not the same as accountability. Nothing here was a black box. It was a very large white one, and nobody could see through it either.
Now the two walls, because they are what makes this dispute unresolvable from outside rather than merely contested.
The first wall is a rule about publication. Every reportable lending decision becomes a public row: originations, denials, amounts, purposes, applicant demographics. The applicant credit score is not on the row, because the Bureau excludes it from the public loan-level data by rule, along with the fields naming the scoring model. So the one factor the bank named first as its explanation is the one factor no outside analysis can hold constant. That limit is measured, not asserted, and it is independent of this dispute: models built on the public record have been found to explain at most around 56 per cent of ten large lenders' 2024 decisions, and even the confidential regulator-held version that carries the score leaves most of the variation unexplained across nine million applications. A disclosure regime designed to make lending legible produced, here, a record with a hole in exactly the shape of the argument.
The second wall is a protective order. The rule base and the two scorecards exist, complete and readable, and they reached a court file once, under seal, in a proceeding whose certification and summary-judgment briefing is extensively sealed. Two experts read them. They agreed about the measurement. Then the case stopped before the merits, both appeal petitions were denied, seven plaintiffs settled confidentially, and the complete copy went back behind the wall carrying whatever answer it holds.
Between the two walls sits the burden, and it points the wrong way for anyone trying to look. The claimant must show robust causality. So when the bank's own documents — on the other side's characterisation of material still under seal — suggested it could not itself identify what was driving the disparity, that counted AGAINST the claimants rather than against the bank. An operator's inability to explain its own outcome became the claimants' evidentiary problem. Read that structure carefully: the party holding the complete decision record has no obligation to explain it, and the party seeking the explanation cannot reach it, and the law allocates the consequence of the resulting ignorance to the second party.
Then there is the discretion paradox, which is the same fact doing two jobs. The human underwriter is the bank's safety claim: the engine never supersedes their judgment, underwriting is done by people. The human underwriter is also the claimants' obstacle: individualised discretion is precisely what defeats the common answer Rule 23 requires. Both propositions are true at once and neither side could give up its half. Meanwhile the pleadings describe that same discretion being squeezed — caseloads two to three times historical volume, people not replaced, an alleged disincentive to check the work — allegations that were never tested, in a case that ended before anyone tested anything. A safety claim resting on a human check whose actual capacity nobody measured is a safety claim nobody can evaluate, and that is a governance fact independent of whether the allegations were true.
One structural detail deserves its own paragraph because almost no other deployment in the atlas has it. Two external enterprise underwriting systems, run by parties this bank does not control, return their own answer on the same application. The proposed classes were eventually narrowed to exactly the applicants for whom those two answers diverged: approved by an outside system, or by the bank's own scorecard, and ultimately denied. So a comparison that would tell you something real was not merely possible here — it was already sitting in the file, on every application, for years. Nothing in this record sets one answer against the other or asks why they differ. The check the whole matter needed was in the building the entire time.
And the escalation channel that did work says the rest. A named claimant pleaded nine months of lost paperwork and repeated re-requests, and then approval the very next day after he notified a federal housing agency. Allegation, settled, never adjudicated — and still the record's clearest picture of a route that opens only when an outside authority is invoked. It resolves one person. It asks nothing about the nine months. Seven claimants settled confidentially and with prejudice; the rule base was never adjudicated; the summary-judgment motion was never decided; and nothing in the public record changed about how the engine decides.
That is the shape this board draws. Not a proven defect, and not an exoneration either. A governance topology in which everybody with the power to ask lacked the access, everybody with the access lacked the obligation, and the one forum that had both stopped short — and in which the system's own size, the thing that should have made it auditable, is what the court cited when it said the question could not be answered in common.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.