FIELD GUIDE
The Mortgage AI Field Guide
Long-form working notes on putting AI agents on regulated mortgage workflows — regulation by regulation, desk by desk. Written for the people who have to answer for the output.
Where to start
Three routes through the guide, depending on which side of the loan you own. Everything else is below, newest first.
Start here: underwriting and closing
Where the cost sits, and which parts of it a rule can actually take over. Self-employed income, conditions, the TRID tolerance buckets and QC at full coverage.
- Calculating Self-Employed Income With an AI Agent: The 1084 Cash-Flow Analysis, the 4506-C Transcript, and the Reasonableness Call the Underwriter Owns
- Conditions Clearing Is the Slowest Part of the Loan, and the Best Place to Put an AI Agent
- The Three Tolerance Buckets: Balancing the Closing Disclosure Against the Loan Estimate With an AI Agent
- Mortgage QC at 100% Coverage: What Changes When AI Reviews Every Loan Instead of a 10% Sample
- AI Agents Inside the LOS: The System-of-Record Boundary, the Write-Back Discipline, and the Audit Trail That Survives an Exam
Start here: compliance and model risk
The questions your risk committee will ask about an AI vendor, answered in long form: SR 11-7, third-party risk, grounding, prompt injection and the 36-hour rule.
- Model Risk Management for AI Agents: An SR 11-7 and NIST AI RMF Playbook
- Third-Party Risk Management for AI Vendors: An Interagency TPRM Playbook for Banks
- Preventing AI Hallucinations in Bank Customer Service: A Grounding and Citation Architecture
- Prompt Injection Defense for Banking AI Agents: Threat Model, Controls, and a Red-Team Cadence
- The Interagency 36-Hour Computer-Security Incident Notification Rule Applied to Bank AI Systems
Start here: servicing and the phone
What governs a borrower conversation before anyone thinks about AI. Reg F, TCPA, the SAFE Act licensing line, loss mitigation and caller verification.
- Regulation F for AI Voice Debt Collection: The 7-in-7 Ledger, the Limited-Content Message, and the Per-Debt Architecture
- TCPA Compliance for AI Voice Agents in Mortgage and Bank Outreach: A 2026 Field Guide
- The SAFE Act Line for AI Mortgage Assistants: When Quoting a Rate Becomes Loan Origination and What NMLS Cares About
- AI Agents for Mortgage Loss Mitigation: A Regulation X Playbook for Servicers
- Voice Cloning and the End of Voice Biometrics as a Sole Factor: A Caller-Verification Architecture for Banks
Latest posts
The Transfer an AI Servicing Agent Cannot Refuse to Discuss: Garn-St Germain, the Due-on-Sale Clause, and the Assumption Request the Note Did Not Anticipate
A servicer's reflex when a property changes hands is to protect the lien and enforce the due-on-sale clause. Federal law forbids that on a specific list of transfers: the Garn-St Germain Act and 12 CFR 591.5(b)(1) bar acceleration on a death, a transfer to a relative or a spouse, a divorce settlement, or a move into the borrower's own living trust. An agent trained to protect the lien will decline to talk to the widow who just inherited the house, which is precisely the unlawful move. Here is how we route these calls, the determination a human keeps, and why 'not on the note' can never mean 'no.'
When the Model Writes the Ad: Mortgage Marketing Copy Under the MAP Rule and Reg Z 1026.24, and the Review Gate Before a Generated Line Ships
Generative AI now drafts mortgage emails, landing pages, and social copy at a scale no compliance team has reviewed a piece at a time. The moment a model writes a sentence about a rate or a payment, two regimes bite: Reg Z 1026.24 triggering terms and the MAP Rule, Regulation N at 12 CFR 1014, which bars material misrepresentation about a mortgage credit product and makes you keep every materially different version for 24 months. Here is the gate we put between the model and the send, the 'no closing costs' line it caught, and why AI turns the recordkeeping rule from a burden into a byproduct.
Underwriting the Bank-Statement Loan With AI: Deposit Analysis That Is Auditable, and the Ability-to-Repay Rule Non-QM Never Escapes
Bank-statement loans qualify a self-employed borrower on deposits instead of tax returns, which makes the income calculation the whole underwrite and the easiest number to get wrong. These are non-QM loans, and non-QM does not mean no ability-to-repay: Reg Z 1026.43(c) still requires a reasonable, good-faith determination on all eight factors. Here is how an AI agent classifies twelve to twenty-four months of deposits, the transfer that double-counted income until we caught it in shadow mode, and the reasonableness call the underwriter still signs.
The Documents That Are Not Done When the Loan Funds: Trailing-Document and Collateral-File Reconciliation With an AI Agent, Without Touching What the Custodian Certifies
A funded loan is not a delivered loan. The recorded security instrument comes back from the county weeks later, the final title policy trails, and until the collateral file is complete and certified the loan is a reps-and-warranties exposure. Fannie Mae requires a document custodian to take physical custody of the note and certify it, and for eNotes the authoritative copy is controlled on the MERS eRegistry. Here is where an AI agent reconciles the file and chases the trailing pieces, the match key that stops it from clearing the wrong exception, and the certifying act it is not allowed to perform.
When the Loan Is FHA or VA, the Waterfall Is the Investor's, Not the CFPB's: Government Default Servicing With AI After the 2025-2026 Overhaul
Most default-servicing AI is built to the CFPB Reg X rules and stops there. FHA and VA loans layer their own loss-mitigation waterfalls on top, and both changed hard in the last year: FHA sunset its COVID options and its HAMP on September 30, 2025 under Mortgagee Letter 2025-12, and VA replaced VASP with a new Partial Claim program that opened June 15, 2026. An agent that recommends a retired option on a government loan is not a small bug. Here is how we version the waterfall by investor and effective date, what the agent computes, and the determination a human still signs.
The Loss-Draft Desk: Where an AI Servicing Agent Moves a Hazard-Claim Check Toward a Repaired House Without Sitting on the Borrower's Money
A fire or a storm turns a performing loan into a claim check made out to the borrower and the servicer at once, and the servicer's job is to get the house repaired while protecting the lien. Where the security instrument gives the servicer authority over the proceeds, why the monitored-disbursement threshold is the control that matters, and the failure mode that turns a slow loss-draft desk into a UDAAP problem.
How Long the Agent's Evidence Has to Live: Retention Clocks for AI Mortgage Records Under Reg B, Reg Z, and Reg C, and the Default That Deletes Your Proof
An AI agent produces the record that proves a decision was compliant, and that record is worthless if the infrastructure deletes it before the exam or the lawsuit arrives. The 25-month ECOA clock, the three- and five-year TRID clocks, the HMDA retention period, and why retention has to be keyed to the loan event rather than a storage default that outlives nothing.
The Monthly Statement Is a Computed Document: Generating the Reg Z 1026.41 Periodic Statement With an AI Servicing Agent Without Misstating What the Borrower Owes
The mortgage billing statement looks like output and behaves like a calculation, and every field on it is a place a wrong number becomes a borrower-facing error at scale. What Regulation Z 1026.41 actually requires the statement to show, why the delinquency box has to trigger off the contractual due date and not the last payment, and the reconciliation an AI servicing agent runs before a statement is allowed to send.
Serving the Borrower Who Applied in Spanish: Limited-English-Proficiency Mortgage Origination With AI, the CFPB Line, and Where a Translated Disclosure Becomes a Liability
AI voice and chat agents make it cheap to talk to a borrower in their language, which is exactly why the risk moves from access to accuracy. Where ECOA and the UDAAP standard still draw the line after the CFPB pulled back its 2021 guidance, why we run the conversation in the borrower's language but keep the operative disclosures in English, and the translation-QA control that stops a servicing term from drifting in the second language.
The Approval That Still Owes the Borrower a Notice: Risk-Based Pricing, the Credit-Score-Disclosure Exception, and What an AI Pricing Agent Has to Trigger
The applicant was approved, just at a rate worse than the best-priced borrower gets, and that approval carries a notice obligation most lenders discharge without thinking about it. Where an AI pricing agent sits in the quote, why mortgage uses the credit-score-disclosure exception instead of the risk-based pricing notice, and the timing line a locked rate cannot cross without the notice out the door.
Loan Boarding QC: The Data-Mapping Defects an AI Agent Catches When Servicing Transfers
When servicing transfers, thousands of loans board onto a new platform through a data map, and the map is where borrower harm gets built in: a wrong escrow balance, a dropped trial payment plan, an ARM index that boarded wrong. Boarding QC on a sample misses the loan that fails. Where an AI agent reconciles the boarding tape against the source records at full coverage, how it ranks mismatches by borrower-harm risk, and why the correction stays a governed act.
The First 90 Days of a Mortgage AI Deployment: The Controls We Stand Up Before the Agent Touches a Live File
Most mortgage AI deployments fail at the start, when the agent goes live before the controls that make it safe exist. The first 90 days are a control build, not a rollout. What we run in shadow mode before the agent touches a file, why the go-live gate is set per control instead of globally, and how deploying controls-first is also the posture an examiner expects to see.
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