The Agentic Bank

Mortgage Document Agent

⬡ Stack Ingests, classifies and extracts data from the mortgage document pile.
◆ Autonomous Orchestrator

Reads the mortgage 'stack' … pay stubs, W-2s, bank statements, tax returns, appraisals … classifies every page and reconciles the extracted fields against the application. Detects the missing document, chases it over borrower comms, and checkpoints the file so a stalled chase resumes.

Memory

Working The loan file: documents in, checklist outstanding, discrepancies found.
Episodic Earlier touches on this application and the borrower's prior loans.
Semantic Document-type schemas, GSE (Fannie/Freddie) data standards, verification rules.
Procedural Extraction playbooks per document type and lender format.
Store File-based memory tool + document store

Orchestration

orchestrator-worker MCP

Harness · Managed Agents … long-running case session; a file-based NOTES record tracks the document checklist and discrepancies outside context; compaction on long files.

Tools

{ } Document intake + OCR API Loan origination system (LOS) Computer use { } Income / asset verification API ›_ GSE data-standard validator Code exec { } Borrower comms (document chase) API

Evals & guardrails

  • Field-level extraction confidence; low-confidence fields flagged for a judge agent to re-extract.
  • Reconciliation check: extracted income/assets must tie to the application.
  • Completeness gate against the GSE document checklist before the file advances.
  • Sampled agent-as-judge review of extraction accuracy vs. source documents.

Offline reflection

Learns which lender / document-format combinations cause extraction failures and updates collection playbooks to request the right artifact up front next time.

Frontier edge

  • Multimodal document reasoning: reads pay stubs, scanned tax returns, appraisal photos and the legacy LOS green screen natively in a single pass.
  • Long-horizon autonomy: shepherds a file across days of borrower back-and-forth, checkpointing the document checklist so a stalled chase resumes itself.
  • Continual learning (eval-gated): folds each lender / format extraction failure into its collection playbooks offline, so it requests the right artifact up front next time.

A sample run

Trigger Borrower uploads a 60-page document package for a refinance.
  1. 1Classify every page; extract income, assets and liabilities.
  2. 2Reconcile against the application; flag a self-employment income gap.
  3. 3Auto-request the missing two years of business tax returns.
  4. 4Update the file checklist and the discrepancy notes.
Output A reconciled, indexed file in ~25 minutes with one outstanding document chased; income discrepancies flagged for the underwriting agent.

In numbers

6,200
Files processed / day
31 min
Median doc-to-decision time
94%
Auto-extracted fields

Handoffs

More on the Home Lending / Mortgage desk