With Fannie Mae’s artificial intelligence governance requirements set to take effect Aug. 6, mortgage technology provider TRUE has released a guide urging lenders to examine whether they can trace AI-generated loan data from its source document through validation, human review, and final entry into the loan origination system.
The executive guide addresses an increasingly important question for mortgage bankers adopting AI: Can the lender reconstruct and defend the path each data point took before it entered the loan file?
TRUE argues that extraction accuracy and model confidence alone are no longer enough for lenders deploying large language models across mortgage operations. Because LLMs generate probabilistic outputs, lenders also need evidence showing where extracted information originated, how it was independently validated, whether a user changed it, and how the final value was mapped into the lender’s system of record.
“Fannie Mae isn’t asking lenders to slow their adoption of AI,” TRUE CEO Steve Butler said. “It’s reinforcing that lenders remain responsible for understanding where AI-generated information came from, how it was validated, what changes were made, and how that data ultimately entered the loan file.”
Fannie Mae issued Lender Letter LL-2026-04 on April 8, establishing an AI and machine-learning governance framework for seller/servicers that use the technology in loan origination or servicing. The requirements become effective 120 days after publication.
The lender letter calls for policies and procedures governing the development, implementation, use, and maintenance of AI and machine-learning systems, along with processes for identifying, measuring, monitoring, and managing the associated risks.
It also requires lenders to govern vendor and subcontractor use of AI under standards no less protective than their own. If requested, seller/servicers must disclose to Fannie Mae the types of AI they use, how the systems are used, and the safeguards in place.
The requirements have already pushed AI governance from a policy discussion into a more immediate procurement and operational issue for mortgage companies. Freddie Mac’s related AI governance requirements took effect March 3.
Following The Data, Not Just The Model
TRUE’s guide introduces two company-developed concepts: “Trust Architecture,” a framework that combines probabilistic AI with deterministic validation, and “Chain of Trust,” the sequence of evidence that accompanies a data element from its source document through acceptance into the loan file.
Those terms and the specific technical controls described by TRUE are not requirements established by Fannie Mae. Rather, TRUE is positioning its framework as a way for lenders to support their governance, monitoring, and disclosure responsibilities.
Under TRUE’s framework, each extracted field remains connected to the original borrower document, including its page location and bounding-box coordinates. The platform also preserves the original LLM-generated output and the separately validated value presented to the user.
TRUE said its Mortgage Operations Service, or MOS, records user modifications, approvals, and values mapped into ICE Mortgage Technology’s Encompass loan origination system. That creates an audit trail extending from the source document through AI interpretation, independent validation, human intervention, and the final LOS update.
The distinction is significant for lenders whose existing AI oversight may focus primarily on whether a system produces accurate results. TRUE’s guide argues that lenders also need to prove why a particular value was accepted and how it reached the loan file.
“Trust Architecture wasn’t created to improve AI accuracy,” Butler said. “It was created to ensure lenders can trust, explain, and defend AI-generated data long after the loan closes.”
Marc Hernandez, founder and CEO of Guideline Buddy, said the ability to document the complete path from a borrower document to validated loan data gives the company greater visibility into its automation.
“TRUE’s ability to document the complete path from borrower document to validated loan data gives us the visibility we need to confidently expand automation while supporting our governance strategy,” Hernandez said.
What It Means
The approaching Fannie Mae deadline raises a more precise issue than whether a lender has adopted an AI policy. Mortgage bankers may need to determine whether their production systems can supply the evidence that policy promises.
That could mean examining whether an AI vendor preserves source-level documentation, records the model’s original response, separates AI output from independently validated data, logs human changes, and connects the final accepted value to the appropriate LOS field.
NMP previously reported that the new GSE requirements are pushing lenders to conduct deeper audits of their AI vendors, with conventional technology reviews potentially falling short of the information needed to evaluate AI-specific risks.
The practical test now is not simply whether a lender or its vendor can explain how an AI product generally works. It is whether the lender can reconstruct what happened to a specific piece of borrower information on a specific loan.
For mortgage bankers expanding AI use across document processing, quality control, underwriting support, and servicing, that traceability could become an important part of internal governance, investor due diligence, audit preparation, and vendor selection.
The Aug. 6 implementation date does not require lenders to adopt TRUE’s architecture. It does, however, leave little time for seller/servicers to determine whether their existing policies, vendor contracts, and operational records can support the level of accountability their AI programs require.
*This article was primarily written by a human author. AI tools were used in a limited capacity for research assistance or light editing.
