The Day of AI Reckoning: Part III

QUESTION 

This article is the third installment of The Day of AI Reckoning. 

·       Part I was published on September 9, 2026.

·       Part II was published on September 16, 2026. 

The three-part article is a response to a reader who asked, in part: 

A day of reckoning is coming, I promise you. Where's the regulation? Who's liable if AI makes a mistake? Who will reimburse me if I get sued because of an AI error? Why is there no quality control? Why are all these FinTech companies springing up everywhere with AI promotions?  

Let's get real. I have no confidence that AI technology is a new type of economic revolution. Maybe it is for the billionaires who run it, since it seems to be a revolution in how fast they can become mega-billionaires.  

Will there be a day of AI reckoning in mortgage banking? 

COMPLIANCE SOLUTIONS  

AI POLICY PROGRAM FOR MORTGAGE BANKING™

Policies and procedures for artificial intelligence related to mortgage banking compliance administration, mortgage loan origination compliance, mortgage servicing compliance, and industry best practices—including, but not limited to, applicable federal and state banking and consumer lending laws. 

AI FOR MORTGAGE LOAN ORIGINATION - SALES MANUAL 

The Sales Manual provides 9 chapters covering the full AI-powered origination funnel in a 90-day implementation roadmap, with clear team roles and fair lending, FCRA, and advertising compliance guidance built into every chapter. Sample AI prompts, an ROI tracking framework, and a compliance glossary. 

RESPONSE 

In Part III, we discuss: 

• What is the scariest compliance problem for AI in mortgage banking?
• Why This Looks Like an Implosion, Not a Correction
• Future or Futuristic AI for Mortgage Banking

____________________________________ 

Order the free White Paper for Parts I, II, III

Distributed after publication of all three articles.

Orders distributed on a first-come, first-served basis.

____________________  

Here are a few recent related articles:  

The Day of AI Reckoning: Part II 

The Day of AI Reckoning: Part I 

Will AI Replace Me? 

AI Replaced Me 

AI Versus Humans: A Dialogue 

Overcoming the Fear of AI  

WHAT IS THE SCARIEST COMPLIANCE PROBLEM 

FOR AI IN MORTGAGE BANKING? 

The single scariest compliance problem would be the invisible, scaled fair-lending discrimination that nobody detects until it has already happened to thousands of borrowers – with no one accountable to fix it.

 

From speaking with many people in compliance about AI, the following outline is the scariest scenarios that keep them up at night!

 

·     Scary Scenario # 1: It's not a one-off error. It's the same error, repeated at scale, silently.

 

When a human underwriter makes a bad call, it's one file. When an AI model has a systematic bias baked in through training data, proxy variables, or model behavior nobody fully audited, it makes the same bad call on every similar file, every day, across every lender using that model or vendor. 

There is considerable concern that leading commercial LLMs could produce different outputs on identical loan applications that differ only by the applicant's race, and that the AI could systematically deny more loans and charge higher rates to minority applicants. If that happens, it won't be a bug that shows up once; it will be a pattern that compounds across the entire portfolio before anyone notices.

·     Scary Scenario # 2: It hides behind the appearance of objectivity.

A human loan officer who discriminates leaves some trace, like a comment or a pattern that a compliance officer can spot in a spot-check. I have said many times that compliance always leaves a trace.

 

However, what if it doesn't because it is systemically opaque?

 

An opaque model doesn't explain itself the same way. It can produce a facially neutral-looking decision (for instance, based on ZIP code, cash-flow patterns, education proxies, or "alternative data") that correlates tightly with race or national origin without ever using a protected class variable directly. That's exactly the mechanism regulators call algorithmic redlining, and it's genuinely harder to detect from the inside than old-fashioned discrimination was.

 

·     Scary Scenario # 3: The detection lag is long, and it usually surfaces from the outside, not the inside.

These patterns tend to get caught by analyzing years of HMDA loan-level data, aggregated and statistically analyzed, often by outside researchers, journalists, or plaintiffs' experts, not by the lender itself in real time.

 

By the time a lender finds out its AI tool has been systematically disadvantaging a protected class, it isn't one bad loan file to fix. It's a multi-year, multi-thousand-loan pattern with exposure under the Fair Housing Act, potential DOJ/CFPB enforcement, investor repurchase demands, and reputational damage – and it all happens at once!

 

·     Scary Scenario # 4: Here's the part that connects directly to what I discussed in Part I: no one is made whole.

The vendor whose model produced the pattern almost certainly disclaimed the output, denied indemnification for exactly this kind of claim, and is protected by an insurance policy that likely excludes AI-related losses.

 

The lender is left holding 100% of the liability for a pattern it didn't design and couldn't fully audit, in an examination area (viz., fair lending) where regulators have repeatedly made it clear that "the AI did it" is not a defense.

 

·     Scary Scenario # 5: One more twist that makes this scarier right now, not less:

 

The CFPB's April 2026 Regulation B (Final Rule) actually pulled back ECOA's disparate-impact "effects test," on the theory that ECOA doesn't authorize disparate-impact claims at all. That doesn't eliminate the risk because the Fair Housing Act's disparate-impact standard, upheld by the Supreme Court, still applies to mortgage lending, and state fair lending laws and, of course, GSE contractual requirements haven't gone anywhere. 

But it means one of the tools regulators and lenders themselves used to proactively catch these patterns early is being narrowed at the federal level, right as AI is making the underlying patterns harder to see in the first place. That combination, to wit, less proactive federal scrutiny, more opaque decisioning, and zero vendor accountability, is the scariest convergence in the whole picture. 

WHY THIS LOOKS LIKE AN IMPLOSION, 

NOT A CORRECTION 

So, pulling these threads together, AI presents as 

a regulatory apparatus rapidly closing the accountability gap,

a vendor market that structurally will not indemnify its own output,

a capital market unable to agree on what any of this is actually worth,

a public that is organizing, town by town, against the physical infrastructure the whole enterprise depends on, thus 

what emerges is not a single crisis but a convergence of them, each reinforcing the others. 

The deeper issue is one of category mismatch. Generative AI is a genuinely powerful tool for retrieval, drafting, and first-pass synthesis, a "sophisticated search engine," as you've stated, which is, in fact, not a bad description of where it delivers reliable value today. 

But mortgage lending is built on a different logic entirely: 

strict liability for fair lending outcomes, personal accountability for licensed professionals, documented human judgment behind every adverse decision, and a compliance infrastructure – like the GSEs, regulators, investors, and the courts – designed specifically to identify who is responsible when something goes wrong. 

A probabilistic tool that cannot explain its own reasoning with full reliability, whose vendor disclaims responsibility for its output by contract, and whose errors surface downstream in a QC file or an exam finding rather than in real time, is a poor structural fit for that environment, no matter how fluent its output reads. 

Lenders are discovering, one remediation engagement and indemnification claim at a time, that front-end speed and cost savings do not offset the liability, remediation cost, and regulatory exposure created on the back end. So far, the AI vendors selling the speed have declined to share in that exposure. 

In my view, until AI providers are willing to stand behind their output the way a licensed compliance professional does, with real indemnification, backed by real insurance, tied to real accountability, the burden of reliability will continue to fall on the institutions that use these tools and on the compliance professionals they call when the tools get it wrong. 

My conclusion in Part I holds fast: 

Lenders willing to lean more heavily on AI output and less on professionally accountable expertise are not eliminating risk; they are relocating it, uncompensated, onto their own balance sheet and regulatory record.  

FUTURE OR FUTURISTIC AI FOR MORTGAGE BANKING 

In Parts I, II, and III, I've discussed the regulatory convergence, the indemnification vacuum, the fair lending exposure, the capital markets wobble, and the infrastructure backlash. Here's where I think this actually lands, not just where the anxiety points.

 

Bottom Line

 

AI won't get expelled from mortgage banking. 

It will get fenced in.

 

Nobody credible in mortgage banking – and that means not the ABA, MBA, ACU, NAMB, CSBS, or other leading professional organizations and publications – is predicting AI disappears from mortgage lending.

 

What's converging instead, from every direction, is a specific functional role I have written about before, called "human-in-the-loop." That is where AI becomes an assistive layer that keeps expanding, sitting underneath a decisioning and accountability layer that stays firmly, and increasingly formally, human. Human-in-the-loop is not an aspiration; it is, in fact, the operating model regulators and institutions are actually converging on.

 

OBSERVATIONS ON THE FUTURE OF AI IN MORTGAGE BANKING

 

·     Productivity: Where AI keeps growing, largely unbothered!

It is fair to say that certain AI proclivities are here to stay, such as document intake and OCR; income and asset extraction from bank statements and pay stubs; fraud signal detection; servicing chatbots and delinquency prediction; marketing draft generation (with review); and internal workflow automation. These are lower-stakes, easily reviewed, high-volume tasks where AI's failure mode is or ought to be "caught in QC" rather than a "systemic discrimination lawsuit." From what I can tell, this is where the real productivity gains are landing today.

 

·     Human Override: Where AI keeps getting pushed back toward being an assistant, not the decider!

Final credit decisions, adverse action reasoning, and anything touching fair-lending-sensitive variables seem far off and unlikely – not impossible, but unlikely. The emerging industry consensus appears to be settling on explainability, pre-deployment fairness testing, human override authority, and audit-ready documentation to the extent that they are no longer optional Best Practices.

 

That is the case, at least for GSE governance mandates and CFPB expectations. If AI is going to be used, governance mandates will require a host of protective and testable functions. Potential systemic bias findings won't be the last of their kind, and each new one adds weight to the rule that a human must be the accountable decision-maker of record, not just a rubber stamp on the model's output.

 

·     Supervisory Wildcard: The politics of regulatory compliance doesn't change supervisory oversight as much as it seems it should.

The current federal administration's AI framework signals a lighter federal touch, such as not establishing a new AI regulator and the aggressive preemption of state AI laws, but even that weak framework leaves ECOA, Regulation B adverse action requirements, and model risk management standards fully intact and enforceable.

 

Deregulatory rhetoric at the federal level is real. Still, it targets new AI-specific bureaucracy, not unwinding the fair lending and safety-and-soundness obligations that already govern every AI use case in lending. I consider this a meaningful distinction, since the mortgage compliance minimum AI has to clear isn't moving much, even as the political conversation around AI regulation generally seems friendlier to industry.

 

·     AI Tools: Capital markets and infrastructure – slowing the pace, but not reversing direction.

If chip and software valuations keep correcting and data center buildout keeps hitting local resistance, AI compute stays more expensive and less available than the most aggressive forecasts assumed.

 

That doesn't kill mortgage AI, of course; however, it just makes lenders more selective about which AI investments they make, favoring proven, narrow, auditable tools over ambitious end-to-end autonomous systems. Or, put another way, it means less about "replacing the underwriter" and more about "giving the underwriter better tools and faster paperwork."

 

·     Liability: What central risk do mortgage lenders face in early-stage AI?

 

Here's how I see it:

 

1.   The more AI expands into the assistive layer, the more AI-touched files exist that someone accountable has to stand behind.

2.   The vendor market has shown no sign of closing the accountability and indemnification gap, because the fair lending and systemic bias risk is arguably uninsurable at the vendor level.

3.   That's not a temporary market inefficiency – it's structural.

4.   Which means interposing the human, professional, compliance role – the licensed, insured, accountable party for AI-assisted compliance output – is not being mandated while the technology matures. 

5.  Therefore, AI is being recklessly entrusted with unsupervised legal accountability. Legal and regulatory compliance support, regulations, supervision, and enforcement must lead the way. 

____________ 

This article, The Day of AI Reckoning: Part III, published on September 23, 2026, is authored by Jonathan Foxx, PhD, MBA, the Chairman & Managing Director of Lenders Compliance Group, founded in 2006, the first and only full-service mortgage risk management firm in the United States, specializing exclusively in residential mortgage compliance.