Snapshot: Tidalwave Leads in Mortgage Underwriting Benchmark
In a timely, first-of-its-kind benchmark released this week, Tidalwave’s mortgage-trained SOLO AI outperformed a widely used general model on a 90-question underwriting test. The study reports an overall accuracy of 84% for SOLO, compared with 71% for the competing model, signaling a potential shift in how lenders vet loan applications and assess risk.
The results come as lenders grapple with increasing demand for faster, more accurate decision-making while maintaining strict privacy standards. The benchmark measures the AI agents on typical underwriting questions, ranging from payroll validation and bank statement review to detection of potentially foreign-sourced funds.
Who Conducted the Test
The benchmarking is a collaboration between mortgage tech firm Tidalwave and researchers affiliated with Columbia University. The team evaluated Tidalwave’s SOLO agent against Claude 4.5 from the AI company Anthropic, using the same 90 underwriting prompts used by loan officers during origination.
Researchers emphasized that the comparison focused on the AI’s ability to interpret structured mortgage data, not just free-text notes. The test simulated common lender workflows to gauge practical performance in real-world scenarios.
Where SOLO Outperformed—and Where It Didn’t
- Yes-or-no compliance checks: SOLO achieved 95% accuracy, a dramatic leap over Claude 4.5’s 42% in this category. These checks flag payroll mismatches, undisclosed debts, and suspicious transactions.
- Transaction identification: SOLO scored 83% versus 80% for Claude 4.5, a closer margin but still meaningful for workflow accuracy.
- Account verification: Claude 4.5 led with 86% accuracy, while SOLO logged 67% in this area, highlighting differences in how each model handles verification data.
The study notes that SOLO’s edge in compliance stems from a design that omits personally identifiable information (PII) during processing, a feature touted by Tidalwave as a privacy safeguard for mortgage data.
Privacy and Data Handling: A Core Advantage
Diane Yu, co-founder and CEO of Tidalwave, described SOLO’s privacy-first approach as central to the benchmark outcome. In her remarks, she underscored that stripping PII before analysis helps mitigate privacy risks inherent in mortgage origination.
“Our design choice to remove sensitive identifiers before any AI processing aligns with rising regulatory expectations and consumer protections,” Yu said. “This isn’t just about accuracy; it’s about responsible AI used in a highly regulated?industry.”
By training SOLO on structured mortgage datasets—such as Uniform Loan Application Dataset (URLA) files—and linking it to underwriting systems used by major mortgage agencies, Tidalwave argues the model has a clearer, more actionable interpretation of loan data than general LLMs that treat files as plain text.
What This Means for Lenders
Industry observers say the findings could influence how lenders deploy AI in origination and risk assessment. The accuracy gaps in high-stakes categories like compliance could translate to fewer manual reviews, faster loan approvals, and improved audit trails—provided privacy safeguards remain in place.
“If a lender can lean on an AI assistant that inherently respects borrower privacy while delivering higher compliance accuracy, that’s a meaningful advantage,” said a senior risk executive who asked not to be named. “The question now is how quickly such systems scale across diverse loan products and geographies.”
Implications for Privacy, Regulation, and the Road Ahead
The benchmark arrives at a moment when financial regulators are scrutinizing AI use in consumer lending. Privacy-by-design features, like PII minimization, are increasingly seen as essential to meeting evolving guidelines and consumer expectations.
Beyond performance metrics, the study highlights a broader shift in the AI landscape: tailored, domain-specific models that integrate with existing underwriting pipelines can outperform general-purpose tools on tasks that require structured data interpretation and regulatory awareness.
What’s Next for Tidalwave and the Industry?
Tidalwave indicated plans to extend SOLO’s reach to additional mortgage products and to run larger, multi-institution pilots to validate the model’s performance across varied loan portfolios. Researchers from Columbia say they aim to expand the dataset with more real-world underwriting scenarios, including risk-based pricing and post-close compliance checks.
Analysts expect other AI developers to respond with domain-adapted models. The broader question for the market is whether lenders will widely adopt these tools, balancing speed and accuracy with privacy, governance, and explainability concerns.
In the end, the findings reinforce a simple, timely takeaway for the mortgage industry: tidalwave tops general models when it comes to underwriting accuracy in a privacy-conscious framework, marking a notable milestone in the push for smarter, safer AI-assisted lending.
Note: The headline result—“tidalwave tops general models”—captures the core takeaway that emerged from the pairing of Tidalwave’s SOLO AI with Columbia University researchers in this benchmark context.
Discussion