In our previous article, we explained why non-QM has struggled to scale despite rapid growth — and why the core issue isn’t underwriting itself, but the absence of a standard system. We also explored how advances in agentic AI finally make it possible to build that missing infrastructure, and why solutions that once felt out of reach are now viable.
Now let’s deep dive into what the AI agents LoanLight has built — and why we’re intentionally starting with pre-underwriting QA.
Why LoanLight Starts with Pre-Underwriting QA
Nearly every downstream issue in non-QM — stacked conditions, delayed closes, late-stage eligibility surprises, and repurchase risk — can be traced back to file quality at submission.
When files arrive incomplete or internally inconsistent, underwriting becomes reactive. Processors chase missing pages. Underwriters reconcile conflicting data. Disclosures are interpreted late instead of upfront. All of this work consumes time, introduces error, and compounds risk as volume grows.
LoanLight’s AI agents are designed to remove this friction upstream, ensuring files are complete, consistent, and ready before underwriting begins — so underwriters can focus on the work that actually matters.
At a high level, these agents work as follows:
Document Health Agent
Ensures all required documents are present, current, and complete. It flags missing, expired, soon-to-expire, or incomplete uploads (e.g., missing pages in bank statements) before they block underwriting.Data Validation Agent
Reconciles borrower data across the 1003, the LOS, and source documents. It surfaces mismatches in names, addresses, employers, identifiers, and entity information that would otherwise require manual “stare-and-compare” work.1003 Agent
Checks loan applications for structural completeness and logical consistency, using conditional logic to verify that all required fields, disclosures, and implied documentation align with program requirements for underwriting.Property Data Agent
Enriches files with third-party property data (ownership, valuation, listing, tax, lien) for the subject property and borrower residence. It performs early checks on occupancy, proximity, and valuation relationships to flag potential eligibility or risk concerns.
The Document Health, Data Validation, and 1003 agents are tightly connected and operate in sequence on the core loan file. They ensure the documents exist, the data reconciles, and the application logic holds together. The Property Data Agent pulls in external context and applies guideline-driven checks that would otherwise require manual research.
Together, they automate the initial underwriting activities that today bounce back and forth between brokers, processors, and underwriters. Instead of discovering basic issues mid-review, underwriters inherit files that are already clean, consistent, and contextually prepared — allowing them to focus on the judgment work that actually determines risk and eligibility.
AI Agents Deep Dive
Agent #1: Document Health Agent
Making sure every document is current, complete, and usable
Underwriters shouldn’t be discovering expired or incomplete documents halfway through a review—or worse, at clear-to-close.
The Document Health Agent continuously evaluates the loan file to ensure documents are valid for underwriting and closing.
It checks whether documents are:
Expired based on program-specific age rules (e.g., credit reports, pay stubs, appraisals)
At risk of expiring before the estimated closing date
Missing entirely, based on loan type and program requirements
Incomplete, including missing pages, partial statements, or broken document bundles
This includes handling complex document structures that routinely trip up manual review—such as multi-month bank statements, tax returns with attached schedules, or bundled W-2s from multiple employers.
Instead of underwriters finding these issues late, the agent flags them as soon as they appear so that there are fewer surprises downstream.
Agent #2: Data Validation Agent
Reconciling the 1003, LOS, and borrower documents
One of the most time-consuming parts of non-QM underwriting is “stare-and-compare” work—hunting through documents to see whether names, addresses, employers, or identifiers actually match the application.
The Data Validation Agent treats the 1003 as the source of truth and continuously cross-checks it against the borrower’s documents.
It builds a borrower profile from the application and scans the full document set to identify discrepancies such as:
Name variations across pay stubs, bank statements, IDs, or leases
Conflicting current or prior addresses
Employer or business name mismatches for self-employed borrowers
SSN / ITIN inconsistencies
Entity or account holder mismatches for assets
Even small discrepancies can create downstream friction or post-close risk. Catching them early allows teams to resolve issues before they turn into conditions—or worse, missed defects.
The goal isn’t to replace judgment. It’s to make sure judgment is applied to a file that already reconciles.
Agent #3: 1003 Agent
Ensuring the application is structurally complete and logically consistent
The 1003 is the foundation of every loan file. But reviewing it for completeness is still a manual, repetitive task—especially in non-QM, where conditional logic matters.
The 1003 Agent automatically reviews the application to confirm it’s structurally complete for underwriting.
It verifies:
All universally required fields are filled out
Conditional fields are completed when borrower responses require them
Disclosures and declarations imply the presence of supporting documentation
For example:
If the borrower is self-employed, required business details must be present
If additional properties are disclosed, the REO section must be complete
If gift funds or borrowed funds are declared, the file should reflect that
If declarations indicate elevated risk (e.g., bankruptcy, liens, concurrent credit), those responses are clearly surfaced for review
Instead of underwriters discovering missing application elements mid-review, the file arrives structurally sound and ready to evaluate.
Agent #4: Property Data Agent
Automating early property research and collateral context
Manual property research is another quiet time sink in non-QM underwriting. Underwriters routinely jump between Zillow, public records, and internal tools to piece together collateral context.
The Property Data Agent automates this work by enriching the file with third-party property intelligence for both the subject property and the borrower’s current residence.
It pulls and consolidates:
Listing status, valuation estimates, and price history
Ownership records, tax assessments, and lien data
Property characteristics and historical transactions
It also flags early risk indicators, such as:
Rapid or unusual price changes
Owner-occupancy inconsistencies
Second homes located unusually close to the primary residence
Investment properties valued higher than the borrower’s primary home
Properties subject to liens or unusual ownership structures
This context doesn’t replace appraisal or underwriting judgment—but it gives teams earlier visibility into issues that would otherwise surface late.
Looking Ahead
LoanLight is an intelligence layer for non-QM — one designed to standardize file quality, eligibility, and ultimately liquidity across the loan lifecycle.
We’re starting with pre-underwriting QA because nothing downstream works without clean, decision-ready files. When quality is established upfront, eligibility becomes consistent. When eligibility is consistent, investor confidence increases. And when confidence compounds, liquidity follows.
The four agents described here represent the first step in that system. Over time, LoanLight will extend this intelligence across underwriting, closing, post-close, and capital routing — externalizing complexity that has historically lived in manual workflows and institutional memory.
This is just the beginning.
If you’d like to see the AI agents in action, schedule a demo.
