Layla Zulfa
The sight of for-sale notices priced in U.S. dollars plastered across windows in Buenos Aires’ fashionable Palermo district is no longer a surprising anomaly. It is a stark, tangible illustration of a profound deficit in institutional trust. While this may seem like a scene from a travelogue, it serves as a potent metaphor for a critical challenge facing the mortgage industry, an industry that fundamentally operates on a complex edifice of delegated trust. From the Government-Sponsored Enterprises (GSEs) trusting lenders, to lenders trusting loan officers, and loan officers trusting borrowers, this chain of confidence is historically underpinned by robust repurchase and warranty structures and meticulous documentation. When a loan defaults, this evidentiary architecture allows for examination, fault assignment, and the enforcement of accountability through repurchase demands.
The Global Financial Crisis (GFC) of 2008 served as a brutal exposé of what happens when this delegated trust outpaces the solidity of its evidentiary foundation. The ensuing wave of loan repurchase demands revealed that the delegated trust was far more extensive than the documentation could adequately support, resulting in tens of billions of dollars in costly settlements. Beyond these headline-grabbing financial repercussions, the GFC also precipitated less visible, yet equally painful, shifts. A policy implemented at Citimortgage, for instance, mandated the re-underwriting of all correspondent-sourced loans, despite the fact that the overwhelming majority of problematic loans originated through capital markets. While settlements were viewed as a necessary cauterization of risk, the re-underwriting mandate catalyzed long-term strategic adjustments that the industry continues to grapple with today.
AI Presents a Structural Trust Problem, Not a Disciplinary One
The advent of Artificial Intelligence (AI) introduces a new, more fundamental trust problem for the mortgage sector. Unlike traditional human underwriters or even deterministic, rules-based engines, AI systems do not inherently generate the kind of process records that are readily reconstructible for audit purposes. AI outputs are often characterized by a non-deterministic reasoning process – the specific weighting of inputs that leads to a conclusion may not be preserved in an auditable format. This means that the same inputs, processed at different times, can yield different outputs, directly challenging the long-held assumption that loan decisioning processes are provable.
In practical terms, this opacity can leave lenders in an untenable position. An AI system might deny a loan application, but be unable to articulate the specific reasons for that denial. This inability to explain the decision process leaves lenders exposed to a cascade of potential repercussions: from disgruntled borrowers demanding clarity, to regulators scrutinizing compliance with fair lending laws, and the GSEs initiating repurchase demands due to a lack of demonstrable underwriting integrity. This is not a minor flaw to be rectified with enhanced record-keeping; it is an intrinsic characteristic of the AI systems currently being deployed at scale within the industry.
The scope of this exposure is further amplified by the increasingly layered nature of the mortgage technology stack. Lenders now rely on a network of AI-powered vendors for various critical functions, including point-of-sale (POS) systems, Loan Origination Systems (LOS), Automated Valuation Models (AVMs), fraud detection algorithms, and income verification tools. Each of these systems makes independent judgment calls that feed into the subsequent stages of the loan process. Crucially, no single lender possesses complete visibility into the entirety of this AI-driven chain. Consequently, when a loan ultimately defaults or experiences issues, pinpointing the specific AI system that introduced the error, and determining whether its decision-making process is auditable, can become an insurmountable challenge.
A mortgage process distributed across four or five "black-box" AI vendors is inherently indefensible under a repurchase and warranty framework. Unlike traditional rules-based systems, where vendor logic could be contractually stipulated and subjected to examination, the outputs of AI systems are inherently variable. Even the vendor may be unable to reconstruct the precise reasoning behind a particular decision, mirroring the lender’s own limitations. This creates a critical gap: the accountability architecture falters at crucial junctures in the technology stack. The accumulated risk is accumulating in areas where oversight is nascent or entirely absent.
Returning to the Analogy of Buenos Aires
The economic realities in Argentina, particularly the widespread use of U.S. dollars for property transactions, underscore the profound impact of a trust deficit. This practice is not a matter of preference but a necessity, driven by the volatility of the Argentine peso, which cannot be relied upon to retain its value between the signing of a contract and the finalization of a sale. The most trust-intensive transaction an individual typically undertakes – the purchase of a home – has been fundamentally restructured to operate in the absence of robust institutional trust in the local currency. This demonstrates a critical principle: markets do not cease to function when trust infrastructure degrades; they adapt and mutate.
The mortgage industry, too, is susceptible to this pressure. When confidence in the underlying accountability architecture erodes, capital becomes more cautious, more expensive, and slower to deploy. The solution cannot be to simply layer more paperwork onto systems that lack transparency. Instead, the response must be architectural, addressing the fundamental design of these systems rather than attempting to retroactively reconstruct their outputs.
The Harrods Conundrum: A Cautionary Tale of Lost Confidence
A poignant, albeit metaphorical, illustration of this principle can be found just two blocks from the author’s hotel in Buenos Aires: the former Harrods department store. This once-prestigious establishment, the only Harrods to operate outside the United Kingdom, shuttered its doors in 1998 and has remained vacant for nearly three decades, a prominent scar on a prime commercial corridor of a major global city. While the building’s facade remains largely intact, and its fundamental structure is sound, what is demonstrably absent is the architecture of trust necessary for an enterprise to confidently commit its resources. Capital may be available, and demand might exist, but without the assurance of verifiable integrity and accountability, investment withers.
The mortgage industry faces a similar existential threat if it mismanages the integration of AI. A superficially polished facade of regulatory compliance can conceal a hollow interior where the fundamental ability to verify trust has quietly eroded. The true vulnerabilities lie at the interfaces between systems – in the chain of judgment calls that flow from one vendor’s AI to another’s. In this complex ecosystem, no single AI model bears responsibility for the aggregate output, and no single contract can fully encapsulate the entirety of the decision-making process.
When AI logic is deeply embedded within third-party vendor systems that lenders neither own nor can fully audit, it exists largely outside the established governance programs that the industry has painstakingly built over years. This creates a significant risk: the potential to erect impressive digital storefronts while leaving the fundamental structural seams unmapped and unprotected.
The Path Forward: Rebuilding Trust Through Architectural Innovation
Despite the daunting challenges, the path forward for rebuilding trust in AI-driven mortgage decisioning is discernible, though by no means simple. Initiatives like MISMO’s Framework for Responsible AI in the Mortgage Ecosystem (FRAME) represent a crucial recognition that the industry must proactively own its response to this evolving landscape. The next phase of this effort must extend beyond the internal workings of individual AI systems to address the critical interactions between these systems.
Effective governance in a multi-vendor AI environment necessitates a fundamental shift in organization, centering on the entire decisioning workflow. This means mapping the complete sequence from initial data input through to the final output, encompassing every system that touched the loan application. Achieving this requires a minimum of three key strategic adjustments:
-
Establishing Verifiable AI Chains: Lenders must demand and develop mechanisms to ensure that the outputs of AI systems can be traced and audited across the entire vendor stack. This involves creating a continuous, auditable log of AI-driven decisions and their contributing factors, from the initial point of data entry through to the final underwriting decision. This requires vendors to adopt more transparent and auditable AI architectures.
-
Defining Shared Accountability: The current model of fragmented responsibility, where each vendor operates in a silo, is unsustainable. A new framework must define clear lines of shared accountability among all parties involved in the AI-driven decisioning process. This could involve contractual agreements that explicitly outline the responsibilities of each vendor in ensuring the integrity and auditability of their AI models.
-
Investing in Explainable AI (XAI): The industry must prioritize the adoption and development of Explainable AI technologies. XAI aims to make AI decision-making processes transparent and understandable to humans, allowing for clear explanations of why a particular decision was made. This is critical for both internal review and for communicating decisions to borrowers and regulators.
The institutions that proactively build this robust architectural infrastructure will be the ones that command the trust of capital providers, the GSEs, and ultimately, the courts, especially when the inevitable first major AI-related repurchase wave arrives. The mortgage industry has already experienced the immense cost of reconstructing accountability after a crisis. The critical question now is whether it possesses the foresight and the will to proactively address this challenge, thereby safeguarding its future and the confidence upon which it is built.
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.
