The architecture of trust in the age of AI

By the time my morning stroll took me past a window of USD-priced for-sale notices for flats in the fashionable Palermo district of Buenos Aires, I was no longer surprised. This is what a deficit of trust looks like. But what gives? You’re reading HousingWire and not a travel blog.

The common thread, aside from real estate, is that our industry runs on a broadly delegated form of trust. The GSEs trust the lenders. The lenders trust the loan officers. The loan officers trust the borrowers. The rep and warrant structure and paper trail make that delegation enforceable. When a loan goes sideways, the chain can be examined. Fault can be assigned. Repurchase demands can be issued.

The Global Financial Crisis (GFC) exposed what happens when this delegated trust outpaces documentation. The repurchase wave stressed that delegation was more than its evidentiary architecture could hold, failing expensively in tens of billions in settlements. 

It also manifested in smaller ways, like the policy to re-underwrite all correspondent-sourced loans when I was at Citimortgage, even though the overwhelming share of bad loans came through capital markets. Less headline-grabbing, but more painful. While settlements are greeted as cauterization of risk, re-underwriting produced long-run changes in strategy the industry is still wrestling with today.

AI presents a trust problem. Not about discipline. About architecture.

This brings us to the matter of AI (of course). AI systems don’t produce the kind of process records a human underwriter or rules-based engine does. They produce outputs with reasoning – how the system weighs inputs to generate a conclusion – that is not preserved in a form audits can readily reconstruct. The same inputs can produce different outputs on different days. Our rep and warrant framework assumes process can be proven. Non-deterministic AI turns that assumption on its head.

In plain terms: the system may not know why it said no. Which means you can’t tell the borrower why you said no. Which means you are exposed – to the borrower, to the regulator and to the GSE on repurchase – in ways that traditional systems never imagined.

This is not a flaw to be patched with better bookkeeping. It is a structural property of the systems the industry is now deploying at scale.

And the exposure extends beyond models a lender controls. The mortgage stack now runs on a layered set of AI vendors – POS, LOS, AVM, fraud detection, income verification – each making judgment calls that feed the next. No single lender has full visibility into that chain. When a loan goes wrong, the question of which system introduced the error, and whether it can be audited, may be unanswerable. 

A process distributed across four or five black-box vendors is not defensible under a rep and warrant framework. Unlike rules-based systems, where vendor logic could be contractually specified and examined, AI vendor outputs are inherently variable – the vendor may not be able to reconstruct the reasoning any more than the lender can.

The accountability architecture stops at the wrong point in the stack. The exposure is accumulating where nobody is looking.

Back to Buenos Aires

A significant share of Argentine property transactions gets conducted in USD, outside the banking system. Not as preference but as necessity, since the peso can’t be trusted to hold value between contract signing and closing. The most trust-intensive transaction most people ever make has been restructured entirely around the absence of institutional trust.

The market doesn’t stop when trust infrastructure degrades. It mutates.

The mortgage industry is not immune to the same pressure. When confidence in the accountability architecture erodes, capital becomes more cautious, more expensive and slower to deploy. The response cannot be more paperwork layered on top of systems that can’t explain their own decisions. It has to be architectural – embedded in how systems are built, not reconstructed from what they produce.

The Harrods conundrum

Two blocks from my hotel sits the only Harrods ever to operate outside the UK. It closed in 1998, never to reopen, nearly thirty years of vacancy in a prime commercial corridor of a major world city. The facade remains mostly intact. The bones are there. What is missing is not capital, nor is it demand. It is the architecture of trust required for an enterprise to confidently commit.

Mismanage the AI opportunity, and the mortgage industry faces a similar fate: a polished storefront of compliance hiding a hollow interior where the ability to verify trust has quietly expired.

The exposure lives at the interfaces – in the chain of judgment calls that flows from one vendor system to the next, where no single model owns the aggregate output and no single contract captures the full decision. AI logic embedded in vendor systems the lender neither owns nor can fully audit sits substantially outside the governance programs the industry has spent years building.

The true exposure lives at the interfaces. It resides in the chain of judgment calls flowing from one vendor system to the next, where no single model owns the aggregate output, and no single contract captures the full decision. When AI logic is deeply embedded in third-party systems that lenders neither own nor can audit traditionally, it sits substantially outside the governance frameworks the industry spent years building. We risk erecting beautiful digital facades while leaving the structural seams unmapped.

The path forward is clear, if not easy

Efforts like MISMO’s FRAME (Framework for Responsible AI in the Mortgage Ecosystem) initiative represent genuine recognition that the industry needs to own the response. The next phase has to go further, addressing what happens between systems, not just within them. 

Governance in a multi-vendor AI environment has to be organized around the decisioning workflow, the full sequence from input to output across every system that touched the loan. Getting there requires, at minimum, three shifts: 

1. Naming an owner for the aggregate decisioning chain, not just individual tools. 
2. Vendor contracts that document how each system’s outputs interact with those around it
3. Workflow-level logging that makes the question of where a decision was made answerable from the lender’s own records, not reconstructed from vendor files after the fact.

The institutions that build this infrastructure first will be the ones capital, the GSEs and the courts trust to extend delegation to when the first major AI-related repurchase wave arrives.

The mortgage industry already knows what it costs to reconstruct accountability after the fact. The question is whether it will recognize this as the moment to get ahead of it.

Marvin Chang is the Associate Director of the FinTech program at Duke University’s Pratt School of Engineering and the Principal of Mercer Knoll Strategies. 
This column does not necessarily reflect the opinion of HousingWire’s editorial department and its owners. To contact the editor responsible for this piece: zeb@hwmedia.com.