Mortgage companies have been putting AI policies in place. Keeping watch over the tools after they go live is proving less consistent, particularly among smaller lenders.
A new survey by the American Association of Residential Mortgage Regulators (AARMR), the Mortgage Bankers Association and Boston Consulting Group found that 87% of respondents had written AI policies or standards, but 58% reported ongoing monitoring for issues such as accuracy and model drift. Among respondents that originated fewer than 5,000 loans in 2025, 40% reported ongoing monitoring, compared with 80% of those that originated at least 25,000 loans.
Those findings land just two weeks after the Conference of State Bank Supervisors (CSBS) released an AI Supervisory Framework for state examiners reviewing banks and state-licensed nonbank financial institutions. Its Core Examiner Guide includes procedures for reviewing whether a company tests or monitors an AI tool’s performance and risks over time, including when its data, vendor, or operating conditions change.
The survey and the examiner guide are separate efforts. Together, they show where an emerging oversight question meets current mortgage company practices: Can a lender demonstrate how it checks an AI tool after approving it for use?
The Gap Grows At Smaller Lenders
The survey was conducted online from April through July and included 31 institutions, 29 of them independent mortgage banks. Participants accounted for approximately 40% of annual U.S. mortgage origination volume. The smaller-lender group contained 10 respondents, so its results should be read as a finding within this sample, not an estimate for all small mortgage companies.
The size difference extended beyond monitoring. Regular AI reporting to a board or equivalent oversight body was reported by 10% of the smaller-lender group, compared with 80% of the larger-lender group. Across all 31 respondents, 81% reported human review of AI and 74% reported third-party or vendor controls.
Regulatory and compliance uncertainty was the most commonly cited barrier to scaling AI, selected by 59% of the 29 respondents who answered that question. Unclear return on investment followed at 45%. Meanwhile, the survey found AI use concentrated in employee productivity and document-heavy work, with less evidence of significant, measurable gains in outcomes such as origination cost or customer experience.
What An Examiner May Request
The CSBS guide gives lenders a more concrete view of how state examiners could review AI use. Depending on the company’s tools and risks, its document request list includes AI policies, an inventory of systems and use cases, vendor oversight records, monitoring reports, testing materials and samples of AI-assisted communications with consumers. It also directs examiners to look for AI features embedded in third-party products — a relevant point for lenders that buy technology rather than build it themselves.
A separate CSBS supplement for nonbanks addresses whether companies monitor vendor AI performance and material changes after initial due diligence. For customer-facing uses, it points examiners toward issues including accuracy, notices, data handling and consumer outcomes. That reaches beyond an AI system making a final credit decision: a tool that helps draft borrower communications or supports eligibility reviews can also warrant oversight appropriate to its use.
The CSBS framework is discretionary. It creates no new legal obligation, and each state agency decides how extensively to use it. Separately, Fannie Mae’s AI governance requirements and MISMO’s vendor governance work have already put lender and vendor oversight on NMP’s coverage map. The new survey adds evidence of where participating lenders report that oversight is strongest — and where it thins out after deployment.
