The Mortgage AI Field Guide
Page 3 of 11
Verification of Employment With an AI Agent: The Written VOE, the Day-of-Closing Verbal, and Why The Work Number Is a Credit Report
VOE reads like a phone call and is actually three separate controls with three separate failure modes. The written verification, the verbal within ten business days of the note, and the reverification when something changes. Where an AI agent runs each one, what Fannie Mae B3-3.1-07 actually requires, and why treating a database VOE as free of FCRA is the mistake that shows up in a dispute.
The Large Deposit Question: Verifying Assets With an AI Agent Under Fannie Mae B3-4.2-02, and Where Asset Sourcing Meets the Bank Secrecy Act
Asset verification looks like adding up account balances and is actually a sourcing investigation. The 50 percent large-deposit rule, funds that have to be the borrower's own, gift documentation, and the point where an unexplained deposit stops being an underwriting condition and becomes a source-of-funds question a regulated lender cannot ignore. Where an AI agent reads bank statements, what it flags, and the line between an eligibility problem and an AML problem.
Reading the AUS Findings: Where an AI Underwriting Agent Adds Value, Where SR 11-7 Draws the Line, and Why the Model That Touches Credit Is Governed Differently
The automated underwriting system already returned Approve/Eligible. The work that follows, clearing the findings, reconciling the conditions against the file, and deciding what the AUS could not see, is where cycle time and defects live. Where an AI agent operates on DU and LPA findings, why an agent that influences a credit decision falls under model risk management, and the governance we run so the agent is a documented, tested model and not a black box in the underwriting path.
The Loan Is Not Sold Until It Clears Suspense: An AI Agent on the Investor Delivery Desk and the Reps and Warranties Behind Every Sale
Origination gets the applause and secondary marketing carries the cost. Purchase suspense, trailing documents, data mismatches between the loan file and the delivery data, and the representations and warranties that make the seller liable for a defect long after the loan funds. Where an AI agent works the delivery desk, how it reconciles the file against the delivery data before the loan ships, and why the reps a lender makes at delivery are the reps an AI cannot make on its own.
The Payoff Desk Is a Compliance Surface: Quoting Payoffs, Reinstatements, and Partial Payments With an AI Agent Under Regulation Z 1026.36(c)
A payoff quote looks like a lookup and is actually a per-diem calculation with a statutory delivery clock, a good-through date the borrower relies on, and a payment-application rulebook underneath it. Where servicers get the numbers wrong, why suspense accounts and late-fee pyramiding turn cashiering into a Regulation Z problem, and what an AI agent on the payoff desk is allowed to compute versus what stays with the servicer.
PMI Cancellation and Termination Is a Date Problem: How an AI Servicing Agent Tracks the Homeowners Protection Act Without Missing the 78 Percent Line
The Homeowners Protection Act runs on dates and percentages a servicer computes from the amortization schedule, not from the current balance. Automatic termination at 78 percent, borrower-requested cancellation at 80 percent, and a final termination at the midpoint that has no LTV test at all. Where servicers miss the line, why the errors are systematic rather than random, and what an AI agent that recomputes the schedule every day actually watches.
AI in Appraisal Review: Reading the Report, Triggering the ROV, and Never Touching the Value
Collateral review is where AI can help most and overstep worst. How to build an appraisal-review agent that flags quality and bias risk, routes reconsideration-of-value requests under the 2024 interagency guidance, and leaves the opinion of value to a licensed human.
Conditions Clearing Is the Slowest Part of the Loan, and the Best Place to Put an AI Agent
The gap between conditional approval and clear-to-close is where cycle time goes to die. How an AI processing agent clears conditions inside the LOS without breaking the Reg B incompleteness clock or the TRID redisclosure rules.
Mortgage QC at 100% Coverage: What Changes When AI Reviews Every Loan Instead of a 10% Sample
GSE quality control rules were written around sampling because full review was impossible by hand. AI removes that constraint. What a full-population pre-funding and post-close QC program looks like, and where the Fannie and Freddie requirements still bind.
AI Income Calculation for Self-Employed Borrowers: The Part of Underwriting Where the Math Has to Be Auditable
Self-employed income is the hardest number in a loan file and the easiest one to get wrong. How to build an AI income engine that matches Fannie Mae Form 1084, holds up under ATR/QM, and carries an audit trail an underwriter and a model validator both trust.
Where AI Actually Takes Cost Out of a Mortgage: The Touches Worth Removing, the $11,000 Loan, and the Controls You Cannot Automate Away
The fully loaded cost to originate a mortgage has run above $11,000 per loan in recent MBA reporting, and most of that is human time. The instinct is to point AI at the whole process and watch the cost fall. That is the wrong model, because some touches are cost to be removed and others are controls that exist on purpose. A production leader's guide to which touches AI should take, which it should assist, and which have to stay human no matter what the cost pressure says.
Calculating Self-Employed Income With an AI Agent: The 1084 Cash-Flow Analysis, the 4506-C Transcript, and the Reasonableness Call the Underwriter Owns
Self-employed income is where mortgage underwriting is slowest, most inconsistent, and most exposed to fair-lending risk, because two underwriters can read the same tax returns and reach different qualifying income. An AI agent can run the Form 1084 cash-flow analysis the same way every time and document every add-back to its line on the return. What it cannot do is make the reasonableness determination Fannie assigns to the underwriter, and building the agent so it stops at that line is the whole design problem.
You Ain't Seen Nothin' Yet
- Any loan type, any agency guideline or custom investor overlays.
- Every finding cited to the guideline or document it came from