AI Agents

Context

Built for banking within the Retail mortgage underwriting stack, this agent is used to verify that a borrower’s declared debts and liabilities are complete and accurate before affordability, DTI, and decisioning. It runs when the application, bureau file, and (where required) bank/loan statements are available, so underwriters rely on evidenced obligations—not self-reporting.

What it does

The agent reconciles the application’s declared liabilities with third-party evidence. It parses the credit bureau file to identify open trade lines, balances, scheduled payments, delinquency status, and public records (liens, judgments), and cross-checks those against the application entries. Where bank or lender statements are provided, it confirms recurring payments (cards, auto, student, buy-now-pay-later, personal loans), detects undeclared obligations (e.g., spousal/child support, tax payment plans), and distinguishes paid-off vs. active accounts. It then produces a clean liability register—what to include, exclude, or clarify—plus a short rationale for every add/remove/edit and links to the exact page/line or bureau segment used.

Core AI functions

Structured ingestion of bureau files and statements; OCR/field extraction for creditor, account type, limit/balance, payment amount and cadence, status and dates; entity normalization to unify creditor and account variants; recurrence detection in statements to surface obligations not visible on bureau; cross-document consistency checks against the application; policy-driven inclusion/exclusion rules (e.g., deferred student loans, authorized-user treatment, soon-to-be-paid debts with proof); and reason-code generation with record-level lineage.

Problem solved

Manual liability checks are slow and inconsistent. Undeclared or misclassified debts inflate affordability, DTI is miscomputed, and exceptions surface late—driving rework and delays to clear-to-close.

Business impact

Affordability becomes accurate and defensible: DTI reflects the true payment burden, ask-backs are precise, rework drops, and cycle time shortens. Files reach decision with fewer surprises and a clear audit trail from each included payment to its source.

Integration and adjacent use cases

Integration complexity: Medium. The agent ingests the application, bureau file, and (when available) bank/loan statements from the borrower/broker portal, credit bureau feed, email, or DMS, then writes the reconciled liability register, flags, and evidence links into the underwriting workbench; core systems are unchanged.

It is commonly combined with:

• Income & employment validator,

• Self-employed income validator to finalize affordability,

• Purchase agreement & contract validator,

• Land registry & title deed validator to align parties and collateral,

• Appraisal report consistency,

• EPC verifier for property evidence,

• Collateral documentation verifier for coverage and loss-payee wording, and

• Final completeness checker to ensure no pages or proofs are missing before sign-off.