Occupation manual 002 / Mortgage origination
A conservative workflow for approved-channel document tracking, missing-item follow-up, borrower-question preparation, disclosure milestones, communication review, and escalation—without automating underwriting, pricing, protected-class treatment, or adverse action.
Short answer: use AI like a file-preparation clerk. It may format an approved checklist, organize status-only fields, draft neutral follow-ups, prepare questions for a qualified responder, and flag missing source links. It must not decide or imply eligibility, approval, denial, underwriting, pricing, rates, fees, terms, conditions, protected-class treatment, or adverse-action reasons.
The federal origination framework distinguishes operational work across originator activity, disclosures, appraisals, underwriting, and record controls. A bounded tool can help prepare a queue or draft, provided the institution has approved the system, inputs, and review route.
Regulation B applies to mortgage credit and covers evaluation, discouragement, discrimination, notification, and adverse action. CFPB guidance says a creditor cannot use a complex or AI-supported model as an excuse for inaccurate or nonspecific adverse-action reasons.
Confirm authorization or licensing, employer procedures, loan-program path, system of record, approved document channel, approved AI tools, prohibited information, and accountable human owners before using AI.
Use an approved checklist and status fields such as requested, received, and human-verified. Keep document contents and borrower identifiers out of unapproved tools.
AI may format open items and neutral follow-up drafts from the approved checklist. A human verifies that each request is required, accurate, appropriately timed, and sent through an approved channel.
Turn questions into a preparation sheet with the authoritative policy source, facts to verify, and qualified responder. Do not let AI improvise credit, pricing, disclosure, or compliance answers.
Show system-of-record dates for disclosures and handoffs, their source, delivery evidence, and human signoff. The AI reminder surface never calculates the controlling deadline on its own.
Escalate anything touching approval, denial, counteroffer, incompleteness, underwriting, pricing, rates, terms, protected information, or adverse action to the authorized credit and compliance process.
| Job | Approved input | AI output | Human gate |
|---|---|---|---|
| Intake clerk | Blank approved checklist and nonidentifying status fields | Organized missing-item queue | Authorized staff verifies every requirement and status |
| Meeting-prep clerk | Approved, minimal agenda fields | Questions, open items, and source links | Loan officer validates facts and controls the conversation |
| Follow-up drafter | Human-approved item list and template | Neutral draft with no outcome promise | Authorized human reviews recipient, facts, attachments, and sends |
| Milestone clerk | System-of-record dates and current policy links | Reminder view and missing-evidence flags | Compliance/operations confirms triggers, applicability, and completion |
| Policy-update clerk | Current public regulator and institution sources | Dated summary with direct links | Qualified reviewer decides applicability and changes practice |
The CFPB tells borrowers to follow precise lender instructions for document submission, use the specified route, keep originals, and confirm receipt. Separately, the FTC says covered mortgage lenders and brokers need a written information-security program with safeguards for customer information. That makes tool approval and data minimization part of the workflow—not an optional footer.
Loan Estimate and Closing Disclosure duties depend on formal triggers, applicability, and delivery rules. Use current Regulation Z, approved compliance procedures, and the institution's system of record. The worksheet captures the official source, system due date, delivery evidence, and human signoff; it does not turn a chatbot into a deadline calculator.
If the workflow encounters a denial, counteroffer, credit-model output, or possible adverse-action reason, stop the AI preparation route. CFPB guidance says creditors using complex algorithms still must provide the applicant with accurate and specific principal reasons. A general chatbot must not choose, rewrite, hide, or invent those reasons.
Act as a mortgage file-preparation clerk, not a loan officer, underwriter, creditor, or compliance decision-maker. From the approved checklist and status-only fields, create: (1) a missing-item queue, (2) neutral follow-up drafts, (3) borrower questions that require a qualified responder, and (4) milestone reminders that cite their source. Do not make or imply eligibility, approval, underwriting, pricing, rate, fee, term, condition, protected-class, or adverse-action decisions. Do not invent requirements or reasons. If any input appears to contain borrower or customer information, stop and label it for approved-system review.
Download the desk, test it only with fictional or approved nonconfidential examples, and have mortgage compliance, information security, and authorized credit personnel review it against the actual institution and jurisdiction.