According to HousingWire, artificial intelligence is reshaping mortgage lending by automating time-consuming tasks and processing larger volumes of data, but the industry needs to approach implementation carefully since errors can harm borrowers, lenders, and investors. Julia Curran, who brings over four decades of mortgage experience and now leads residential AI products at SitusAMC, emphasizes that mortgage professionals need to understand that different AI models have different capabilities and weaknesses. When evaluating AI vendors, companies should involve their subject matter experts rather than simply adopting whatever technology is available, as Curran's team found that most of the vendors they tested could not handle the complex requirements of secondary market work.
The selection of which AI model to use should depend entirely on the specific task at hand, Curran explains. One model might excel at extracting data from documents while another performs better at summarizing information or working with tables. These varying strengths come with different costs, so lenders need to understand how to manage the whole ecosystem of model selection, accuracy needs, and expenses if they want to optimize both performance and spending.
Rigorous testing is essential before any AI gets deployed in a mortgage workflow, according to Curran. Companies need to test against the broadest range of loan types and borrower situations they can imagine, using databases of loans that have already been reviewed by experienced people to establish what the correct answer should be. For example, income verification works very differently depending on whether a borrower has a single job with a pay stub or multiple income sources requiring tax returns and partnership documentation. Even after AI goes live, companies cannot stop testing because AI models change and need regression testing to ensure they continue producing reliable results. Bias testing matters critically too, as AI systems can make decisions based on information that humans might not expect them to use.
Curran draws a clear line about where AI should and should not be used in mortgage operations. She sees significant opportunity to automate data entry and information extraction across loan origination systems, pricing and origination systems, and servicing platforms. However, she believes credit decisions should always remain with a human being because lending guidelines often involve compensating factors and circumstances that require judgment rather than just numbers. AI can certainly support decision making by flagging issues or organizing information, but the final approval or denial should rest with a person.
One meaningful area where AI can reduce risk involves improving due diligence in securitizations, where traditional loan sampling can create blind spots. Historically, when a pool of loans gets reviewed through sampling before securitization, the highest-rated loans get selected for the deal while lower-rated loans get excluded. This can make the overall pool appear stronger than it actually is when rating agencies only see the grades from the reviewed sample. Using AI to review all loans in a pool, not just a sample, combined with targeted human review of any loans flagged by the AI, can provide a much more complete picture of actual credit quality.
What I am seeing locally here in the Bay Area and throughout the East Bay is that our lending community is still in the early stages of understanding how to properly implement AI in their operations. The firms that will thrive are those that take the time to test thoroughly, involve their experienced loan officers and compliance teams in the process, and maintain human oversight of major decisions. Lenders and brokers who cut corners with AI implementation or who trust it too completely without verification are taking on real risk that could surface later in loan performance or investor disputes.
