The mortgage industry is currently navigating a period of profound technological transition. As artificial intelligence continues to permeate every facet of the financial sector, the promise of automating labor-intensive processes, analyzing vast datasets, and reducing manual overhead across the loan lifecycle has moved from a theoretical future to an operational necessity. However, in an industry where a single piece of inaccurate information can trigger a cascade of consequences for borrowers, lenders, and secondary-market investors, the implementation of AI requires a level of scrutiny that far exceeds standard software deployment.
In the high-stakes environment of mortgage banking, "good enough" is a dangerous threshold. The complexities of regulatory compliance, credit risk, and investor requirements mean that the deployment of AI cannot simply be a matter of selecting a popular large language model (LLM) and setting it to work. Instead, it requires a sophisticated understanding of the technology’s limitations and a steadfast commitment to human-in-the-loop oversight.
Julia Curran, the senior managing director of residential AI products at SitusAMC, stands at the center of this technological evolution. With a career spanning more than four decades, Curran’s experience touches every corner of the industry, including origination, secondary markets, servicing, and technology. Her perspective highlights a critical reality: while AI offers transformative potential, its success is entirely dependent on the rigor of its testing and the depth of the subject-matter expertise guiding its development.
Not all mortgage AI is built the same
The term "artificial intelligence" is often used as a catch-all phrase, suggesting a monolithic technology that functions uniformly across different sectors. In reality, the AI landscape is a fragmented ecosystem of models, each with distinct architectures, strengths, and weaknesses. For mortgage leaders, understanding these nuances is the first step toward a successful digital transformation.
According to Curran, the mortgage market is a highly specialized niche, and the tasks required of AI are exceptionally complex. The industry is not merely looking for tools that can summarize servicing notes or extract basic data points; it requires systems capable of navigating intricate legal documents, financial statements, and complex regulatory frameworks.
"Our market is very niche, and what we need AI to do is complex," Curran explains. She notes that during the evaluation process at SitusAMC, the team conducted proof-of-concept testing with 27 different vendors. The results were telling: 26 of those vendors were unable to perform the specific, high-level tasks necessary to make AI a viable tool for secondary market reviews. This stark failure rate underscores the gap between general-purpose AI and the specialized tools required for the mortgage sector.
For executives and decision-makers, this means that model selection must be driven by specific use cases rather than brand recognition. Different models carry different costs and performance metrics; some may excel at identifying documentation, while others are better suited for interpreting complex tables or summarizing qualitative data. Managing this process requires a strategic balance between cost-efficiency and accuracy, necessitating a deep understanding of credit usage and model limitations.
Crucially, this selection process cannot happen in a vacuum. Curran emphasizes that subject-matter experts—the people who have spent years in the trenches of mortgage banking—must be involved in the evaluation of AI vendors. To automate a task accurately, the technology must mirror the logic and judgment of an experienced professional. While an AI might extract data correctly in a technical sense, only a human expert can determine if it has extracted the right information for a specific, often nuanced, situation.
Accuracy is only the first test
In many industries, a 90% accuracy rate might be considered a success for an AI pilot. In the mortgage industry, such a margin of error is unacceptable. Rigorous testing is the bedrock of AI implementation, and that testing must account for a staggering variety of real-world scenarios.
To ensure reliability, Curran advocates for the use of "ground truth" databases—repositories containing hundreds or thousands of loans that have already undergone exhaustive manual review. These databases serve as the gold standard against which AI performance is measured. By running an AI model against these verified loans, developers can see exactly where the technology aligns with human judgment and where it falters.
The complexity of income verification provides a perfect example of why this variety is necessary. A standard AI might easily process a straightforward W-2 and a single pay stub for a salaried employee. However, the mortgage world is rarely that simple. A single borrower might have six different income sources, complex tax returns, and K-1 forms from various business interests. A truly effective AI must be able to handle all these variations with the same level of consistency and accuracy as a senior underwriter.
Furthermore, the testing process does not end once a model is deployed. The nature of AI—particularly models hosted by third-party providers—is that they are constantly being updated and refined. This necessitates "regression testing," a continuous process of ensuring that new updates to the model do not break existing prompts or lead to unexpected changes in output. If a model’s underlying logic shifts, the results it produces for a specific loan type could change, potentially introducing errors into a live workflow.
Beyond technical accuracy, bias testing has emerged as a critical component of the development cycle. AI models can inadvertently introduce bias based on proxy data or unexpected correlations within the training sets. Without strict controls and constant monitoring of what the model is and is not allowed to consider, lenders risk running afoul of fair lending laws and ethical standards.
Efficiency has limits when judgment is involved
The primary driver of AI adoption is efficiency—the ability to do more with less. In the mortgage lifecycle, there are numerous areas where AI can meaningfully reduce manual labor. Data entry is the most obvious candidate. Manually keying information into a Loan Origination System (LOS), Point of Sale (POS), or servicing system is a labor-intensive process prone to human error. AI’s ability to accurately extract and port this data can streamline operations and allow staff to focus on higher-value tasks.
However, Curran draws a firm line when it comes to the final credit decision. While AI can support the decision-making process by organizing data and flagging potential issues, it lacks the nuanced judgment required to sign off on a loan.
"I don’t believe AI should make the final credit decision," Curran asserts. Credit decisions are not purely mathematical; they often involve "compensating factors"—contextual details that might explain a dip in credit score or a gap in employment that a rigid algorithm might miss. Guidelines vary between investors, and there are often circumstances that require a human to weigh risk against the specific story of the borrower. In these instances, AI serves as an incredibly powerful assistant, but the final accountability must remain with a human professional.
AI could expose what loan sampling misses
One of the most significant potential benefits of AI lies in its ability to transform due diligence in the secondary market. Historically, the process of securitization has relied on loan sampling. Because it was too expensive and time-consuming to manually review every single loan in a massive pool, firms would review a representative sample—often around 30% to 40%—to estimate the risk of the entire pool.
Curran points out that this traditional sampling method contains inherent blind spots that can lead to an understatement of risk. In a typical securitization, loans are graded from A to D. Often, only the highest-rated loans (A and B) are selected for the final pool, while C and D grades are excluded. However, if the initial sample showed a high rate of C and D grades, that risk profile should theoretically apply to the unreviewed portion of the pool as well.
For example, in a pool of 1,200 loans, if a 600-loan sample reveals a 33% rate of C or D grades, the remaining 600 unreviewed loans likely share that same risk profile. If an issuer then creates a deal using only the "clean" loans from the sample plus the unreviewed loans, the rating agency might only see the high grades of the reviewed portion, potentially missing the systemic risk present in the unreviewed half.
AI offers a solution to this transparency gap. By combining traditional third-party reviewer (TPR) audits with AI-driven reviews, firms can achieve 100% coverage. AI can extract data and run it through compliance and credit calculation engines for every single loan in a pool. If the AI flags an issue in a loan that wasn’t part of the manual sample, a human expert can then step in to perform a targeted review. This hybrid approach ensures that no "D-grade" loans slip through the cracks, providing a more accurate risk assessment for rating agencies and investors alike.
Why quality matters more than speed or price
As the industry moves toward broader AI adoption, a divide is forming between firms that prioritize long-term value and those that prioritize speed to market. Curran expresses concern over the lack of industry-wide standards for evaluating AI models. In the current "gold rush" for automation, vendors can release models quickly, claiming high accuracy without having undergone the rigorous testing that firms like SitusAMC require.
"What scares me most is that there’s no industry-wide way to determine whose AI models or data sets are actually better," Curran says. Without standardized benchmarks, it is difficult for lenders to know if the answers they are getting are truly accurate.
The stakes of this uncertainty are high. If an inaccurate AI model is used in the secondary market to evaluate loan quality, it could lead to loans being rated more highly than they deserve. These overvalued assets are then sold to investors, creating a ripple effect of risk that can compromise the stability of the market.
For Curran, the path forward is defined by a commitment to quality over convenience. At SitusAMC, the threshold for releasing an AI capability into a live workflow is a minimum of 95% accuracy—a benchmark that can take months of parallel testing and "soft launches" to achieve. Even after a tool goes live, the process of manual spot-checks and regression testing must continue indefinitely.
The integration of artificial intelligence into the mortgage industry is not a race to be won by the fastest or the cheapest. It is a meticulous process of blending cutting-edge technology with decades of human experience. For the lenders and investors who get it right, AI promises a future of unprecedented clarity and efficiency. For those who cut corners, the risks to the borrower, the lender, and the broader financial ecosystem are too great to ignore. Accuracy, supported by human judgment, remains the only viable currency in the mortgage lifecycle.
