The Adverse-Action Explainability Ladder: What a Bank Owes an Applicant It Rejects

A bank can test a credit model for bias, pass, and still be wrong about the thing that matters. That is not a hypothetical. It is a documented sequence: a regulator clears the model on the statistics, criticises the bank for being unable to explain a decision to the customer who asked, and three and a half years later a second regulator bills the same institution over 89 million dollars — not for the model, and not for the explanation, but for what happened after the customer picked up the phone.

This article lays out the discipline that closes that gap: The Adverse-Action Explainability Ladder — five rungs describing what a bank owes an applicant it refuses. The ladder has one unusual property, and it is the whole point: it does not reward climbing. Two rungs are indefensible for AI models rather than merely immature, one is the floor, one is the rung that looks most generous and should not be automated, and one is the mandate. Running underneath all five is the rung nobody draws on the diagram — the reason has to survive the trip to the person answering the phone.

Why this matters now

Most banks are watching Brussels. For a bank in Vietnam, Brussels is not the first clock — and as of July 2026 it is not even the near one. Three deadlines were running. Vietnam's two have struck; Europe's has just been pushed further away.

Vietnam's Personal Data Protection Law, Law 91/2025/QH15, has been in force since 1 January 2026. Not approaching — live. What it gives a data subject is the rights to object to, to restrict and to delete the processing of their data, and the right to withdraw consent. Separately, at Article 21(1), it carries a data-protection impact assessment duty: one original filed with the specialised authority within 60 days of first processing, refreshed every six months under Article 22(1). That duty is unconditional — there is no high-risk trigger, no profiling trigger and no automated-decision trigger anywhere in Article 21, the only exemptions are for competent state agencies and for small enterprises, start-ups, business households and micro-enterprises, and no bank qualifies for either. It also carries, at Article 27(1)(b), something more specific than the general architecture: credit information may not be used to score, rank, assess credit information or assess the creditworthiness of a data subject without that subject's consent. That is a banking-specific consent precondition sitting directly on credit scoring, and Decree 356/2025/NĐ-CP Article 8(2) details what the consent request must disclose — the purposes including credit scoring and creditworthiness assessment, the collection sources and sharing parties, the retention period, and the mechanism for withdrawing consent.

Vietnam's AI Law, Law 134/2025/QH15, took effect on 1 March 2026, with risk tiers and, at Article 35, an 18-month transition for pre-existing systems in the finance sector running to approximately 1 September 2027. That is a transition for systems already in place, not a sector-wide holiday for anything a bank builds next.

The EU AI Act's duties for credit scoring now apply from 2 December 2027 — and that date moved five weeks before this article was published. Regulation (EU) 2024/1689 classifies creditworthiness evaluation of natural persons as high-risk at Annex III point 5(b), requires human oversight of high-risk systems at Article 14, and explicitly carves out systems used to detect financial fraud. None of that changed. The date did. As enacted, Article 113 put Annex III high-risk obligations in the general-application tranche of 2 August 2026. Then Regulation (EU) 2026/1744 — the Digital Omnibus on AI, adopted 8 July 2026, published 24 July 2026, in force 27 July 2026 — replaced that provision at its Article 1(40)(b), and Annex III high-risk obligations now apply from 2 December 2027 (Annex I products move to 2 August 2028). The amendment took effect six days before the original deadline would have arrived, so that deadline never bound anyone.

Two things worth being exact about, because the second is where commentary is going wrong. The enacted amendment is an unconditional calendar date. Recital 40 gives the delayed availability of standards as the reason for the change and then says the date "is set to" 2 December 2027; the operative Article 113 contains no trigger, no Commission decision, no earlier-of mechanism and no backstop. If you find a source describing the new date as conditional on standards availability, it is describing the Commission's proposal rather than the regulation that passed. And the deferral moved the compliance machinery, not the individual right — Article 86, the right to obtain an explanation of an individual decision, sits in Chapter IX rather than the deferred Chapter III, was untouched by the amendment, and applies from 2 August 2026, as do the Article 50 transparency duties. Whether Article 86 bites on a credit decision before the Chapter III regime starts is this article's reading of the text rather than a settled question, so treat the application date as the fact and the practical effect as an argument.

So the ordering holds, and the gap just widened. Brussels is still last of the three — 1 January 2026, then 1 March 2026, then 2 December 2027 — and now last by roughly twenty-three months. A bank that sequenced its programme around the European calendar has been handed sixteen more months for the compliance machinery and no extra time at all on the two obligations that actually bind it. The binding clock for a Vietnamese bank struck first and struck quietly, in January.

The problem: two wrong reactions, and the duty that outlived both

When a model denies a loan and someone asks why, banks tend to do one of two things.

The first is to hide behind complexity. It has never been a defence: in the United States, the Equal Credit Opportunity Act (15 U.S.C. 1691(d)) and Regulation B (12 CFR 1002.9) require a creditor taking adverse action to give the specific principal reasons for it, and the duty does not scale down as the model scales up. The CFPB said so directly in Circular 2022-03 (May 2022) — guidance the Bureau withdrew in May 2025 without touching the statute or the regulation underneath it. "Too complex to explain" is not an engineering constraint. It is a compliance finding.

The second is to reach for a checklist. Lenders have long ticked a box from the sample forms in Regulation B — insufficient credit file, purchasing history — and treated the list as a safe harbour. It never was one on those terms: Regulation B's own commentary requires the reasons disclosed to be the applicant's actual principal reasons, which the CFPB restated in Circular 2023-03 (September 2023) before withdrawing it in May 2025. If a model denied someone because of the type of merchant they shop at or the hour at which they transact, "purchasing history" is not a reason. It is a euphemism.

That withdrawal is worth being exact about, because a compliance officer will raise it and because it is the strongest available illustration of this article's own closing caution. In May 2025 the Bureau withdrew a large batch of its guidance documents, interpretive rules, policy statements and advisory opinions applicable as of 12 May 2025; Circulars 2022-03 and 2023-03 both appear in the enumerated list. Neither may be cited as the current Bureau position.

It changes far less than it appears to, and that is the point of anchoring on statute rather than on guidance. The circulars were never the source of the duty; they were the Bureau's reading of it, and the withdrawal notice's own rationale is that non-binding guidance should not create obligations outside notice-and-comment rulemaking. The duty is untouched: ECOA at 15 U.S.C. 1691(d)(2) and (d)(3) requires a statement of specific reasons; Regulation B at 12 CFR 1002.9(b)(2) requires that the statement "be specific and indicate the principal reason(s) for the adverse action"; and the Official Interpretations to Regulation B state that the specific reasons disclosed "must relate to and accurately describe the factors actually considered or scored by a creditor." Regulation B's definition of adverse action at 12 CFR 1002.2(c) expressly includes refusing to increase a credit limit — which is precisely the decision at the centre of the case below.

A statute does not get withdrawn by notice. That is the whole reason to build against one.

Most banks fail here the same way regardless of jurisdiction: they fund the model, pass the fairness test, and never check whether the reason for a denial survives the trip to the person who answers the phone.

The framework: five rungs, and the direction is not up

Read the ladder as a legal-obligation gradient, not a capability curve.

Read the verdict column again. On this framework's reading, R1 and R2 are not early stages of a journey — they are indefensible for AI models, because neither produces the applicant's actual principal reasons. R3 is not an aspiration; it is where this framework argues deployment is permitted to begin. And R4, the rung that looks most generous to the customer, is the one to avoid automating.

The verdict column is this framework's own reading of the duty, not a taxonomy any regulator publishes. It is grounded in the statutory and regulatory text cited above — not in any enforcement action against an AI-generated reason code, because as the proof section sets out, no such published case exists. Read the rungs as an argument with a mechanism, and test the argument rather than inheriting the labels.

R3 is a model-selection constraint, not a reporting feature

The duty is specific and accurate: the reasons must describe the factors the model actually scored. That single word — accurate — quietly decides which models a bank may deploy, and it decides it upstream of anything the compliance team can fix.

The reason is a paper called Fooling LIME and SHAP: Adversarial Attacks on Post hoc Explanation Methods, by Slack, Hilgard, Jia, Singh and Lakkaraju, presented at the AIES conference in 2020. It demonstrates that perturbation-based explainers can be deliberately gamed: an adversary can build a discriminatory classifier and wrap it in scaffolding so that when the explainability tool probes it, the tool emits an innocuous, false explanation. Read the finding at its actual scope — it establishes a technical vulnerability under an adversarial-deployment threat model, not a claim that every SHAP value in production is wrong.

The legal step from there is this framework's inference, not a holding: no adjudicator has ruled that an unfaithful explainer voids an adverse-action notice. But the regulation's words point one way. If the explanation does not accurately describe the factors the model actually scored, it is hard to see how the notice satisfies 12 CFR 1002.9(b)(2) — and a creditor who cannot validate the accuracy of its own approximation is not well placed to argue otherwise.

So the bank has a choice, and it is an architecture choice rather than a reporting one: bolt an explainer onto a black box and hope it is faithful, or build a model that is interpretable by construction.

The usual objection is that interpretability costs accuracy. On tabular credit data — rows of income, history and obligations — that objection is not merely folklore; on the only published numbers from a real lending deployment, the sign is the other way round.

Intuit's QuickBooks Capital has scored incoming loan applications since July 2019 with monotonic XGBoost, weight-of-evidence binning and exact TreeSHAP, and the team published the result: Wang, Lesner, Ran, Rukonic, Xue and Shiu, Using Small Business Banking Data for Explainable Credit Risk Scoring, IAAI-20. XGBoost with monotonic constraints beat the scorecard it replaced by 7% on the K-S statistic — 0.437 against 0.410. And the paper's own ablation reports that monotonic constraints improve K-S by about 8% on average and ROC-AUC by about 3%, rather than costing anything. Constraining the model made it better, not worse.

Two honest limits on that. It is one deployment, in small-business lending rather than retail consumer credit, and the 7% is measured against the scorecard it replaced rather than against an unconstrained challenger — so read it as an existence proof, not a benchmark. And the paper attributes the US explanation duty to the Fair Credit Reporting Act, which is the wrong statute for adverse-action reasons in credit; cite it for the engineering, not the law.

The direction has support from the literature too. Grinsztajn, Oyallon and Varoquaux showed at NeurIPS 2022, in the Datasets and Benchmarks track, why tree-based models still outperform deep learning on typical tabular data — which is the data shape credit underwriting actually has. The Intuit deployment reproduced it industrially: its neural network scored 0.385 on K-S, losing outright to the 0.410 scorecard. The models that are hardest to explain are not the ones winning here, so the trade-off leaders think they are making is mostly not on the table. When that is the situation, stop calling it a trade-off.

What this article still cannot give you is a bank. The Intuit paper is a lender's published account, but it is a software company's small-business lending arm, not a regulated retail bank publishing its discrimination metrics and its adverse-action notices side by side — and no bank has done that. The engineering pattern is well understood; the regulated-bank practice base behind it is thin enough that any vendor claiming otherwise should be asked for the citation.

The Intuit paper is worth reading for a second reason, closer to this article's own argument than the accuracy numbers are. It builds its adverse-action reason codes by mapping the largest SHAP value per application onto roughly a hundred reason codes grouped by feature, adopts monotonic constraints expressly "to ensure credit decisions have acceptable explanations", and documents a "Monotonic SHAP Pitfall" in which missing-value encoding skews the attribution and produces incorrect explanations of credit decisions. That is R3 failing in production for a reason no fairness test would surface — and it is the same class of failure the adversarial-explainer literature describes, arrived at by accident rather than by attack.

R3.5 — the rung nobody puts on the diagram

Here is the failure the Apple Card case actually demonstrates, and it is not a modelling failure.

A reason can exist, be mathematically exact, be logged, and be auditable — and never appear on the screen of the agent taking the call.

R3 is satisfied on paper. The customer still gets "I can't see why, sir."

So treat R3 as met only when the reason renders on the frontline screen. There is a test that costs an afternoon and needs no budget line: pull ten recent denials and ask a contact-centre agent to explain each one using only what they can see. Whatever they cannot explain is the gap between your compliance documentation and your customer's experience — and it is the gap that generates complaints, because the applicant experiences the silence, not the audit log.

R4 is the trap

Counterfactual explanations answer the only question an applicant actually has: what do I change? Tell someone that clearing two thousand dollars of debt would have flipped the decision and you have given them something genuinely useful. This is the rung that feels like financial inclusion, and it is the one to be most careful about.

Then the model gets retrained — which sound model-risk practice pushes a bank towards — and the decision boundary moves. Upadhyay, Joshi and Lakkaraju showed at NeurIPS 2021 that recourse already issued stops holding once the model shifts, and Rawal, Kamar and Lakkaraju had documented the same fragility under data and model shift the year before. The applicant does exactly what the letter said, reapplies, and is denied again.

The bank has now manufactured a grievance, with its own letter as the evidence.

Note what the regulation does and does not ask for here, because it cuts against the instinct to over-deliver: the Official Interpretations to Regulation B provide that a creditor need not describe how or why a factor adversely affected an applicant. Counterfactual recourse is therefore a service decision rather than a disclosure duty — which is exactly why automating it is an unforced error. Methods designed to survive retraining do exist: Pawelczyk and co-authors' probabilistically robust recourse, presented at ICLR 2023, is the clearest of them, and it buys robustness with computational cost. Until that cost is low and the tooling is stable, counterfactuals belong in a loan officer's hands as a conversation tool, particularly in SME and commercial lending where the client is actively managing a business. They do not belong in a retail app as prescriptive advice.

R5 is the mandate — and Vietnam gets there by a different road

Three things must exist together at the top rung. Two of the three are architecture; none of the three is a model.

In Europe, that top rung has a clean statutory home: Article 14 of the AI Act requires human oversight of a high-risk system, and creditworthiness evaluation is a high-risk system. Vietnam reaches the same operating requirement by a different road, and the road matters — because a programme justified on the wrong provision does not survive its first legal review.

Do not import the European justification. You do not need it, and the reason is the part most commentary on Vietnamese banking has missed: Vietnam already imposes an algorithmic-explanation duty, and it has been in force since 1 January 2026.

It is not in the Law. It is in the implementing decree. Decree 356/2025/NĐ-CP, Article 10(3) requires the controller to notify the data subject of automated personal-data processing, to explain the operating principles of the algorithm and its effects on the subject's rights and legitimate interests, and to provide options for the subject to exercise a right not to participate. Article 10(5)(b) adds the controller's duty to account to the data subject. The Law's own Article 30(4) sits behind it, requiring that AI processing of personal data be risk-classified — a duty on the controller, not a right for the subject, which is why the decree is where the individual entitlement actually lives.

Read that against this article's subject and the asymmetry is the opposite of the one usually claimed. On explanation, Vietnam is ahead of where most banks assume it is. What Vietnam does not have is the other half of GDPR Article 22: there is no prohibition on solely-automated decisions, and no right to demand human review. That is not an inference from silence in one place — the Vietnamese terms for intervention (can thiệp) and for reconsideration or review (xem xét lại) do not appear anywhere in either the Law or the Decree. So a customer in Vietnam can require you to explain the algorithm and can opt out of the automated processing; they cannot require a human to re-decide.

Now read the rest against a credit-decisioning stack. A customer who objects to the processing, or withdraws the consent that Article 27(1)(b) requires before you may score them at all, has to be answered by something — and an automated decisioning line with no human tier behind it has nothing to answer with. Objection intake needs somewhere to route. Restriction needs a path that is not "the model runs anyway". The explanation duty needs a system that can state the algorithm's operating principles to a customer, not only to a supervisor. The impact assessment has to describe a control environment that exists.

Decree 356/2025 Article 5 then puts a clock on it: two working days to acknowledge a request to withdraw consent, restrict or object, fifteen days to implement it, twenty if a processor or third party must act, one extension of at most fifteen days with the burden on the controller to prove the extension was necessary and reasonable. One caveat that matters more than the numbers: Article 5(2) excepts processing that does not require consent under PDPL Article 19, and a good deal of bank credit processing may sit on a non-consent basis. Whether that narrows the objection route for credit decisioning is a question for Vietnamese counsel, not for a framework — do not build a service-level commitment on it without that advice.

That is the R5 lane, reached from data-subject rights rather than borrowed from Europe's oversight article. The lane is the same lane: objection intake, a human underwriter tier working from the same risk parameters, and an audit log wired into the customer-service system. Build it in 2026 on the PDPL's own terms, and check the enacted text with your own counsel before you write the justification into a board paper — the provisions are what the argument stands on, and this article names the rights rather than their article numbers for exactly that reason.

Proof: the case, read precisely, and what the evidence does not support

Honesty about the evidence base is part of the argument here, since the subject is telling applicants the truth.

The case. In March 2021 the New York State Department of Financial Services published its Report on Apple Card Investigation, opened after public allegations that the algorithm behind the card — underwritten by Goldman Sachs — gave women lower credit limits than their husbands. Using underwriting data for what the report calls "nearly 400,000 New York applicants", covering applications from the card's launch to the date of the initial complaints, the Department's Consumer Examinations Unit ran regression analysis and did not find evidence of unlawful disparate treatment, and did not find evidence of disparate impact fair lending violations. The report's own summary of its exhaustive review is that it "did not produce evidence of deliberate or disparate impact discrimination but showed deficiencies in customer service and transparency." Note the epistemics: a regulator finding no evidence of a violation is not the same as a proof of fairness, and this article does not upgrade it into one. On the statistics, the model was cleared.

The finding that mattered. The second half of that sentence is the finding this framework is built on, and it is the Department's own conclusion rather than a consumer's complaint. One detail in the report is the thesis in miniature: at the Department's request, the bank explained its credit decisions for every individual who had complained, and "in each instance, the Bank was able to identify the factors that led to the credit decisions, such as credit score, indebtedness, income, credit utilization, missed payments" — none of them an unlawful basis. The reasons existed. The bank could produce them for a regulator on request and could not produce them for the customer on the phone. The report separately records the consumer's complaint that "an Apple Card customer service agent could not explain these algorithms or the basis for the difference in credit limits", and criticises the appeals design in its own voice: "Requiring account holders to wait six months to appeal credit terms frustrated applicants." The report also records that the bank eliminated that six-month waiting period in the aftermath of the November 2019 media scrutiny — so read the six months as the design that produced the grievance, not as a practice still standing when the report published.

The bill. On 23 October 2024 the Consumer Financial Protection Bureau ordered Apple and Goldman Sachs to pay over 89 million dollars — at least 19.8 million dollars in redress and a 45 million dollar civil penalty from Goldman Sachs Bank USA, a 25 million dollar civil penalty from Apple. The redress figure is "not less than" 19.8 million in the order's own words, which is why the Bureau says "over 89 million" and never names a total.

What the order was actually about matters more than the number, and it is wider than the usual summary. Two things, not one. The first is dispute handling: Apple failed to forward transaction disputes to Goldman, and Goldman missed Regulation Z's acknowledgment, investigation and resolution requirements, reported disputed amounts adversely before finishing billing-error resolution, and held consumers liable for claimed unauthorised use before a reasonable investigation. The second is the marketing of interest-free financing — both firms were found to have misled customers into expecting that Apple device purchases on the card would be automatically enrolled in Apple Card Monthly Installments, and Goldman was additionally found to have misrepresented how refunds applied to those balances. Apple's conduct drew an abusiveness finding for material interference with consumers' understanding of enrolment, which is the CFPA-specific standard with no analogue in the older statutes and the most interesting count in the document for a bank reader.

The findings were made under the Consumer Financial Protection Act and the Truth in Lending Act. No ECOA, no Regulation B, no fair-lending count, no model-bias count. That is the citation behind this article's insistence that the order was not about the model — it is not an inference from the press release, it is the statutory basis on the face of the orders. The Bureau's order and the Department's report are three and a half years apart, from two different regulators, pointing at the same operating model.

One shelf-life note, because this article is about not citing dead instruments. Everything above is stated in the past tense deliberately: the Bureau terminated Apple's consent order on 22 September 2025, expressly waiving any alleged non-compliance, after Apple had paid the 25 million dollar penalty. The order as issued, and the money as ordered, remain matters of record. Apple is not under that order now, and nothing here should be read as saying otherwise. Goldman's order, on the evidence available, remains in force.

A sound model. A failed operating model. A bill three and a half years later — not for the model, and not for the explanation, but for what happened after the customer picked up the phone.

Be careful with that sequence, because the obvious reading of it is wrong. The 89 million dollars was not the price of unexplainability; no regulator has yet charged anyone for that. What the case supports is narrower and still damaging: the same operating-model weakness that leaves an agent unable to explain a decision is the weakness that mishandles the dispute which follows, and the second one is demonstrably expensive.

What the record supports: the Department's finding and the Bureau's order as described above; the statutory and regulatory adverse-action duty in ECOA and Regulation B; the two circulars' reading of that duty, and their withdrawal; the EU AI Act's high-risk classification of creditworthiness evaluation with fraud detection carved out, and its 2 December 2027 application date as amended by Regulation (EU) 2026/1744; the Vietnamese statutes and decree and their commencement dates, including the Article 10(3) automated-processing explanation duty; the demonstrated adversarial vulnerability of post-hoc explainers; recourse invalidation under retraining.

What it does not: there is no publicly documented enforcement case anywhere for vague or inaccurate adverse-action reasons generated by an AI model. The enforcement that exists attaches to servicing and disputes. And no bank has published an appeal-overturn rate, an objection take-up rate, or a cost per human review.

That last gap matters more than it looks, because it means the strongest objection to R5 — mandatory human review destroys the economics of automation — cannot be answered with a number. It can only be answered with mechanics: human-in-the-loop is not human-does-everything. If R3 is engineered properly, the reviewer validates rather than re-underwrites, and the cost of the lane is a function of how good the reasons are. Which returns the whole argument to the floor.

How to apply this — 5 steps

Read this as a credit-committee agenda for the next cycle, not a transformation programme.

1. Ban unconstrained black-box models from primary retail credit decisioning. Mandate monotonically constrained gradient boosting with exact TreeSHAP, or an equivalent interpretable-by-construction stack. This secures R3 mathematically and removes the adversarial-explainer exposure at the point where it is cheapest to remove — model selection.
2. Run the frontline delivery test this week. Ten denials, one agent, one screen. Fix what they cannot explain. This is the only step on the list that produces evidence by Friday.
3. Build the R5 lane during 2026. Objection intake, a human underwriter tier, and an immutable audit log the reviewer and the supervisor can both read — justified on the PDPL's objection, restriction, withdrawal-of-consent and impact-assessment duties, which are the provisions your counsel can actually point at.
4. Do not automate counterfactual recourse. Give it to officers as a conversation tool; keep it out of the app as prescriptive advice. Revisit when robust-recourse methods are cheap and stable, and not before.
5. Quarantine generative AI from final decisioning. Excellent for summarising a credit memo; unusable as the thing that says no, because the causal chain cannot be proven to a supervisor — and a reason you cannot prove is a reason you cannot defend.

Underneath all five is one filing task that is easy to forget and dated: the impact assessment for every profiling and automated-decisioning model, within 60 days of first processing.

Risks and caveats

This framework describes a problem a bank may not have yet. If underwriting is still mostly scorecard-and-officer, the ladder reads as future tense — while the PDPL's obligations landed in January regardless. The sequencing advice is therefore unglamorous: build the R5 lane because the law requires it, and do not buy a black box you would then have to explain.

On the legal record, three limits are load-bearing. First, the two CFPB circulars named above are withdrawn, applicable 12 May 2025, and are cited here as history rather than as operative guidance; the duty they described lives in ECOA and Regulation B, which are unchanged, and a US regulator's reading of a US statute is in any case persuasive rather than binding in Vietnam. Second, the EU date has already moved once and this series got it wrong in both directions before getting it right, so the provenance is worth stating. As enacted, Article 113 put Annex III high-risk obligations at 2 August 2026. Regulation (EU) 2026/1744 replaced that provision and moved them to 2 December 2027; it entered into force on 27 July 2026, six days before the original date would have bitten. The amended date is a hard one, not conditional on standards availability — an earlier draft of this article followed the Commission's proposal-era framing and said otherwise. What has not moved is Article 86 and the Article 50 transparency duties, which apply from 2 August 2026; whether Article 86 reaches a credit decision while Chapter III is still deferred is this article's reading, not a settled point. Third, this article names the PDPL rights it relies on in words and, where it cites an article number, cites one that has been checked against the enacted text. The rights list is Article 4(1) — withdrawal of consent at point (b), and objection, restriction and deletion bundled together at point (d) — and the impact-assessment duty is Article 21(1). One residual caveat, stated because it is the honest one: the only official source is a digitally signed scan, read by OCR, and Vietnamese OCR does not reliably separate d from đ. The point sequence, the legible final point and an independent English translation all agree, which is why the numbers appear here at all; a reader relying on them for a board paper should still put one human eye on the promulgated text. Read Article 35 of Law 134/2025/QH15 the same way — a transition for pre-existing systems, on this series' reading, and worth checking against the primary text.

On the technical record, two limits. The statutory duty here is old and stable, even as the supervisory guidance around it churns — but the practice base is thin. The one deployment this article names is a software company's small-business lending arm, not a regulated retail bank, and no bank has published a peer-reviewable account of a constrained, exactly-attributed credit model with its discrimination metrics and its adverse-action notices side by side. And an earlier draft of this article put a specific AUC cost on constrained boosting. That figure is withdrawn: it had no source, and the named deployment reports the opposite sign — monotonic constraints improving K-S by about 8% and ROC-AUC by about 3%. The honest claim is that on the only published numbers the constraint is not a cost, measured in one place, on one book.

And a US rule is not a plan — which this article has already demonstrated with its own citations rather than needing a footnote for it. Two circulars that stated the US position on algorithmic adverse action were withdrawn by the agency that wrote them, fifteen months before this was published, without a word of the statute changing. The same instability shows up elsewhere, and the open-banking rule is the cleanest example available: the Bureau's Section 1033 rule was finalised in October 2024 (89 FR 90838, 18 November 2024), reopened by the Bureau's own advance notice in August 2025 (90 FR 40986), and then preliminarily enjoined on 29 October 2025 in Forcht Bank, N.A. v. CFPB pending completion of that reconsideration, with the consolidated appeals held in abeyance in the Sixth Circuit since March 2026. Finalised, self-reopened, enjoined and parked, inside twenty-two months — and preliminarily enjoined is not struck down, which is itself the kind of distinction that gets lost in a board paper. Europe moved a date by sixteen months six days before it was due. Guidance moves; so do the instruments that look most settled. Build against duties that are in force in your own jurisdiction, and date-stamp the ones that are not.

This is independent thought leadership, not affiliated with any current or past employer, and not a substitute for your own legal, compliance and model-risk review. Regulatory and banking references anywhere in this series use publicly disclosed primary sources only.

Where this fits in the series

The GenAI Assurance Grid asked whether a bank can assure a model it cannot open. This week asks a different question with a different creditor: not what the bank owes its supervisor, but what it owes the person the model refused. Those are separate duties, and the second one is answered in call centres and appeal queues rather than in model validation. Last week's Deposit Loyalty Gate looked outward at the risk your customer's model creates against your funding base; this week turns back inward, to the risk your own model creates against the customer. The R5 lane is built inside the Trust pillar of The AI-Augmented Bank's five-pillar operating model, and a constrained-model mandate is exactly the kind of rule that belongs in the pre-deploy checkpoint described by the 4 Stone Guardians governance gate. Whether a reason code can be accurate at all depends on the meaning layer in The AI-Ready Bank's five-layer data-readiness map — a model cannot name the factor it scored if the bank cannot name the field.

Before your next credit committee

Ask one question, and do not accept a model-performance answer to it: for the last thousand denials, can we say why — to the person we denied, in words that reach them, with a human who can be asked to look again?

If the answer involves the word "complex", the bank is standing on a rung that no longer holds weight.

Every week, we send one architecture-grade framework like this one to leaders governing AI in banking, in The AI Architect Letter — one issue a week, no hype.

Sources & note. Apple Card: New York State Department of Financial Services, Report on Apple Card Investigation, March 2021 — regression analysis by the Department's Consumer Examinations Unit on underwriting data for "nearly 400,000 New York applicants" from the card's launch to the date of the initial complaints, finding no evidence of unlawful disparate treatment and no evidence of disparate impact fair lending violations, and concluding that the review "did not produce evidence of deliberate or disparate impact discrimination but showed deficiencies in customer service and transparency"; the report records that at the Department's request the bank was able to identify the factors behind each complainant's decision, records the consumer's complaint that a customer service agent could not explain the basis for the difference in credit limits, criticises that "requiring account holders to wait six months to appeal credit terms frustrated applicants", and records that the bank eliminated that waiting period in the aftermath of the November 2019 media scrutiny. Consumer Financial Protection Bureau consent orders of 23 October 2024, File Nos. 2024-CFPB-0012 (Apple Inc.) and 2024-CFPB-0011 (Goldman Sachs Bank USA) — Apple, a $25m civil money penalty under CFPA §1055(c) (order §IX ¶52); Goldman, redress of "an amount not less than $19.8 million" (§IX ¶73) plus a $45m civil money penalty (§X ¶82). "Not less than" is the order's own wording on the redress, which is why the Bureau frames the total as "over $89 million" and never names an exact figure. Apple's counts: unfair acts in failing to forward transaction disputes; deceptive acts as to automatic ACMI enrolment on Apple device purchases; and abusive acts through material interference with consumers' understanding of ACMI enrolment (12 U.S.C. §§5531(a), (d)(1), 5536(a)(1)(B)). Goldman's counts: six dispute-handling limbs under Regulation Z (12 C.F.R. §1026.13(c)–(f) and Supp. I Comment 12(b)) with a derivative CFPA violation, plus deceptive acts as to ACMI enrolment and as to the application of refunds to ACMI and non-ACMI balances. Statutory basis is the Consumer Financial Protection Act and the Truth in Lending Act — no ECOA, no Regulation B, no fair-lending count, no model-bias count, which is the citation behind this article's insistence that the order was not about the model. The Apple consent order was terminated by the Bureau on 22 September 2025 (File No. 2024-CFPB-0012, Document 3), expressly waiving any alleged non-compliance, after Apple had paid the $25,000,000 penalty; every statement about it above is in the past tense for that reason. No termination order for Goldman Sachs Bank USA was found in the Bureau's published document index, so its order — including the restriction on launching a new credit card absent a credible compliance plan — is taken to remain in force; that is absence of evidence from one index rather than proof, and nothing in this article rests on it. A separate $25 million figure appears in both orders as a liquidated-damages term between Apple and Goldman for launch delay; it is a commercial term between the two companies and is not the civil money penalty. United States adverse-action duty: 15 U.S.C. 1691(d)(2) and (d)(3); 12 CFR 1002.9(b)(2) ("must be specific and indicate the principal reason(s) for the adverse action"); the Official Interpretations to Regulation B (the specific reasons disclosed "must relate to and accurately describe the factors actually considered or scored by a creditor", and a creditor need not describe how or why a factor adversely affected an applicant); 12 CFR 1002.2(c), which includes refusing to increase a credit limit within adverse action. CFPB Circular 2022-03, released 26 May 2022, 87 FR 35864 (14 June 2022) — ECOA and Regulation B "do not permit creditors to use complex algorithms when doing so means they cannot provide the specific and accurate reasons for adverse actions", with its footnote 1 on validating the accuracy of post-hoc approximations. CFPB Circular 2023-03, released 19 September 2023, 89 FR 27361 (17 April 2024) — creditors "may not rely on the checklist of reasons provided in the sample forms" where those reasons do not specifically and accurately indicate the principal reasons. Both circulars were withdrawn by the Bureau, applicable as of 12 May 2025, 90 FR 20084, and are cited above as history — an articulation of the statutory and regulatory duty, not operative guidance. European Union: Regulation (EU) 2024/1689, Annex III point 5(b) (creditworthiness evaluation of natural persons as high-risk), Article 14 (human oversight), with systems used to detect financial fraud carved out — none of which was amended. Regulation (EU) 2026/1744 (Digital Omnibus on AI), adopted 8 July 2026, OJ L 2026/1744 of 24 July 2026, in force 27 July 2026 — Article 1(40)(b) replaced Article 113's third paragraph point (c) and moved Annex III high-risk obligations to 2 December 2027 (Annex I products to 2 August 2028). The amended date carries no condition as to standards availability: recital 40 gives delayed standards as the reason and says the date "is set to" 2 December 2027, with no trigger, no Commission decision and no earlier-of mechanism in the operative text; the Commission's November 2025 proposal was conditional, the enacted regulation is not. Article 6(5), on Commission classification guidelines, is expressly excepted from the deferral and still applies from 2 August 2026. Article 50's transparency obligations were not deferred either: Chapter IV is not referenced in the amending regulation, Article 113's second paragraph is unamended, and Article 50 is touched only at paragraph 7 on codes of practice — though new Article 111(4) gives synthetic-content systems placed on the market before 2 August 2026 until 2 December 2026 to comply with Article 50(2). Article 86 (right to explanation of individual decision-making) sits in Chapter IX Section 4, was not amended and was not carved out of Article 113, so it applies from 2 August 2026, as do the Article 50 transparency obligations; whether Article 86 reaches a credit decision before Chapter III applies is this series' reading of the text and not a resolved question. Vietnam: Law 91/2025/QH15 (Personal Data Protection), passed 26 June 2025, in force 1 January 2026 — Article 4(1) rights, with withdrawal of consent at point (b) and provision, deletion, restriction and objection bundled at point (d); Article 21(1) impact-assessment duty, one original to the specialised authority within 60 days of first processing, refreshed every six months under Article 22(1), with exemptions only at Article 21(6) and the size-based carve-outs at Article 38(2)–(3) that no bank meets; Article 27(1)(b), barring use of credit information to score, rank or assess creditworthiness without the subject's consent; Article 30(4), requiring risk classification of AI processing of personal data. Decree 356/2025/NĐ-CP, signed 31 December 2025, in force 1 January 2026, detailing the PDPL — Article 10(3), the duty to notify the data subject of automated personal-data processing, explain the algorithm's operating principles and its effects on the subject's rights and legitimate interests, and offer a right not to participate; Article 10(5)(b), the controller's duty to account to the subject; Article 8(2), the finance-sector consent-disclosure list; Article 5, the response clocks for rights requests, subject to its own exception for processing that does not require consent under PDPL Article 19. The DPIA duty is stated as a duty in its own right and not as a consequence of any high-risk or profiling designation, which the enacted text does not support. There is no prohibition on solely-automated decisions and no right to demand human review in either instrument: can thiệp (intervention) and xem xét lại (review) do not occur in either text. Vietnamese primary sources are the Government's digitally signed promulgation PDFs read by OCR and cross-checked against an unofficial English translation; the Article 4(1) point letters rest on the point sequence, the legible final point and that translation agreeing, and OCR does not reliably distinguish d from đ. Law 134/2025/QH15 (Artificial Intelligence), in force 1 March 2026, Article 35 allowing pre-existing systems in finance 18 months, approximately 1 September 2027 — a transition for pre-existing systems, not a sector-wide exemption. All Vietnamese readings above are this series' reading of the primary text rather than a regulator's published interpretation. Explainability: Slack, Hilgard, Jia, Singh and Lakkaraju, "Fooling LIME and SHAP: Adversarial Attacks on Post hoc Explanation Methods", AIES 2020 — a technical vulnerability under an adversarial-deployment threat model; the legal consequence drawn from it above is this framework's inference and has not been adjudicated. Tabular models: Grinsztajn, Oyallon and Varoquaux, "Why do tree-based models still outperform deep learning on typical tabular data?", NeurIPS 2022 Datasets and Benchmarks Track — the published title includes "typical", which the arXiv title omits. Deployment: Wang, Lesner, Ran, Rukonic, Xue and Shiu (Intuit Inc.), "Using Small Business Banking Data for Explainable Credit Risk Scoring", IAAI-20, AAAI 2020, pp. 13396–13401 — monotonic XGBoost with weight-of-evidence binning and exact TreeSHAP, deployed by QuickBooks Capital to assess incoming loan applications since July 2019, outperforming the scorecard by 7% in K-S (0.437 vs 0.410), with the paper's ablation reporting monotonic constraints improving K-S by about 8% and ROC-AUC by about 3% on average, and its neural network scoring 0.385. Its SHAP-to-adverse-action-code mapping and its documented "Monotonic SHAP Pitfall" are cited above; its attribution of the US explanation duty to the Fair Credit Reporting Act is not adopted — for adverse-action reasons in credit the operative duty is ECOA and Regulation B. An earlier draft of this article asserted a specific AUC cost for constrained boosting; that figure had no source, is contradicted in sign by this deployment, and is withdrawn. Recourse invalidation: Upadhyay, Joshi and Lakkaraju, "Towards Robust and Reliable Algorithmic Recourse", NeurIPS 2021; Rawal, Kamar and Lakkaraju, "Algorithmic Recourse in the Wild", arXiv 2020 (preprint, no peer-reviewed venue on record). Robust recourse: Pawelczyk, Datta, van-den-Heuvel, Kasneci and Lakkaraju, "Probabilistically Robust Recourse", ICLR 2023 — robustness spanning model shift and action perturbation. Guyomard, Fessant, Guyet, Bouadi and Termier, "Generating Robust Counterfactual Explanations" (CROCO), ECML PKDD 2023, addresses robustness to perturbation in the recourse action rather than to model shift, and is not cited above as an answer to retraining-driven invalidation. The Bureau's Section 1033 open-banking rule: finalised 89 FR 90838 (18 November 2024), reopened by the Bureau's advance notice at 90 FR 40986 (22 August 2025), preliminarily enjoined 29 October 2025 in Forcht Bank, N.A. v. CFPB, No. 5:24-cv-00304 (E.D. Ky.), consolidated on appeal in the Sixth Circuit (Nos. 25-6164, 26-5005) with briefing held in abeyance since 30 March 2026 — preliminarily enjoined, not vacated. This is independent thought leadership, not affiliated with any current or past employer, and not a substitute for your own legal, compliance and model-risk review.