AI Agents

Context

Built for banking within the Retail mortgage underwriting stack, this agent is used once the appraisal is received to confirm that the report, valuation basis, and appraiser credentials are consistent with policy and the rest of the file (application, title/registry, EPC). It runs before clear-to-close so value, scope, and identity mismatches don’t surface late.

What it does

The agent reads the full appraisal (narrative, comps grid, photos, certification, license details, scope of work) and reconciles it with the application and property evidence. It validates the subject property match (address, parcel/legal description), confirms the appraiser’s licence/credential status and effective dates, and checks that the valuation approach/method used (sales comparison/cost/income) is appropriate per product and property type. It verifies effective date and report freshness, reviews the comparable selection and adjustments for completeness/consistency (distance, recency, condition, GLA/lot adjustments), and flags internal contradictions (e.g., condition noted as average while photos/descriptions imply material issues). Discrepancies—mismatched parcel/address, expired or unverified licence, missing exhibits/pages, unrealistic adjustments, or scope that conflicts with product rules—are returned with concise rationales and links to the exact page, table, or credential lookup used.

Core AI functions

Document parsing and section detection for appraisals and addenda; OCR and field extraction for subject identifiers, dates, licence IDs, method and scope declarations, comps and adjustments; cross-document consistency checks against title/registry and application; rule evaluation for product/jurisdiction appraisal requirements (method appropriateness, freshness windows, credential validity); quality checks on comps grid coverage and adjustment coherence; and reason-code generation with page-level lineage to the evidence.

Problem solved

Appraisal inconsistencies are often found late, forcing rework or revised conditions. Manual checks across narrative, comps, credentials, and other file artefacts are slow and error-prone, and product rules on method/freshness are applied inconsistently.

Business impact

Files move to decision faster and with fewer exceptions. Value is evidenced against a verified report, credentials and methods align to policy, ask-backs are precise, and collateral risk is reduced. Rework drops, cycle time shortens, and audits improve through clause-level, source-linked findings.

Integration and adjacent use cases

Integration complexity: Medium. The agent ingests appraisal PDFs and credential lookups from portal/email/DMS and writes validation outcomes, exceptions, and evidence links into the underwriting workbench; cores are unchanged.

It is commonly combined with:

• Land registry & title deed validator (to reconcile parcel/legal description),

• EPC verifier (to check energy-certificate requirements),

• Purchase agreement & contract validator (to align parties, price, and dates),

• Collateral insurance & pledge verifier (to confirm coverage and loss-payee wording),

• Income & employment verifier / Self-employed income analyzer (borrower evidence),

• Debt & liability cross-checker (declared vs. bureau), and

• Final completeness checker (to catch missing pages or expired lookups before sign-off).