AI Mortgage Assistants Struggle with Accuracy, Study Finds

Researchers have found that AI mortgage assistants are prone to errors when it comes to identifying foreign deposits in bank statements. In a recent study, three leading general-use AI models were tested on their ability to flag transactions of foreign origin. The results showed that the models struggled significantly with this task, highlighting concerns about the accuracy and bias of these systems.

When researchers asked the AI models which deposits could be of foreign origin, they found that transactions connected to English names were flagged only 13.3% of the time. However, when the deposit came from a non-English name, the number jumped to 77%. This disparity suggests that the models are relying too heavily on linguistic cues rather than actual evidence.

This issue is not just about AI’s ability to provide accurate information; it also raises concerns about bias and fairness in mortgage lending. Matthew Toles, a Columbia University doctoral student and co-author of the study, notes that ‘you aren’t determined whether you’re foreign-based on how your name sounds.’ This means that borrowers from non-English speaking countries may be unfairly flagged as high-risk or denied loans.

The study’s findings are particularly concerning given the widespread adoption of AI mortgage assistants in the industry. More than 80% of lenders were evaluating the technology as of June, according to a survey by The Mortgage Collaborative, while 17% had already deployed it in live production workflows. This rapid adoption has raised concerns about the lack of standardization and testing for these systems.

Diane Yu, co-founder and CEO of Tidalwave, a mortgage technology company that collaborated with Columbia researchers on the work, notes that ‘everybody is using AI, but nobody really understands how to use AI in compliance and with the correct guardrail.’ This highlights the need for standardized testing and evaluation methods to ensure that these systems are used correctly.

The study introduces MortarBench, an open-source benchmark designed to measure the accuracy of AI mortgage assistants. The benchmark includes a set of common and useful tasks in loan origination, such as identifying payroll deposits, large transactions, and jointly held accounts. By using this benchmark, lenders can compare the performance of different models and identify areas for improvement.

The researchers tested three leading general-use AI models - Gemini 3.1 Pro, GPT-5.5, and Claude Sonnet 4.6 - on a set of tasks from MortarBench. The results showed that even these strong models struggled with accuracy, particularly when it came to identifying specific transactions in bank statements.

Zhou Yu, an associate professor at Columbia and study co-author, compares the task of finding specific transactions to ‘finding the needle in the haystack.’ However, the researchers found that the AI models often made the opposite mistake, pulling in transactions that didn’t belong. This highlights the need for more accurate and reliable methods for identifying relevant information.

The most common problem with the AI models was misclassification, where they incorrectly identified a transaction as belonging to a certain category. For example, Gemini counted a personal loan as a buy-now-pay-later transaction, while GPT-5.5 assumed all wire transfers were international. These errors demonstrate that simply finding relevant information is not enough; the systems must also accurately classify and interpret it.

Matthew Toles notes that these scores are for ‘naive use of foundational models,’ which means they are using a basic model without any additional training or fine-tuning. He cautions that major industry players typically develop their own proprietary models, which perform better on these tasks. For example, Tidalwave scored 95% on yes or no questions in a separate benchmark test.

The study’s findings have significant implications for the mortgage lending industry, particularly when it comes to AI adoption and standardization. As lenders look to automate processes and reduce costs, they must also ensure that their systems are accurate and reliable. The MortarBench study provides a much-needed tool for evaluating these systems and identifying areas for improvement.

The results of this study could have far-reaching consequences beyond the mortgage lending industry. As users increasingly rely on AI models for financial decision-making, errors and biases can carry higher stakes. Matthew Toles warns that ‘we don’t have visibility into what these models are doing, what companies are doing with them, and what the outcomes are.’ This highlights the need for greater transparency and accountability in AI development and deployment.

The study’s authors emphasize the importance of asking lenders about their use of AI mortgage assistants and whether they have undergone independent evaluation. Borrowers should also be aware that these systems may not always provide accurate or unbiased information. As Diane Yu notes, ‘you should ask those questions’ and ‘be very careful.’

The MortarBench study provides a valuable resource for evaluating the accuracy and reliability of AI mortgage assistants. By using this benchmark, lenders can compare their models to industry standards and identify areas for improvement. This could help mitigate systemic risks associated with widespread adoption of these systems.

In conclusion, the findings of this study highlight the need for greater standardization and testing in the development and deployment of AI mortgage assistants. As the industry continues to adopt these technologies, it is essential that we prioritize accuracy, reliability, and fairness.