How ADE parses and extracts structured data from mortgage packages: pay stubs, bank statements, tax returns, identity documents, and disclosure forms.
A mortgage application file can contain up to 500 pages and a dozen distinct document types, each arriving from a different source with a different layout. Lenders who process these packages with template-based systems build and maintain a separate extraction pipeline for each document type. ADE processes the entire package through a single parse step and extracts each document type using a schema, with no per-document-type configuration required.
What a Mortgage Package Contains
Standard mortgage origination packages include documents across five functional categories:
- Income verification. W-2 forms, pay stubs (typically 30 days to 3 months), IRS Form 1040 tax returns (typically 2 years), and for self-employed applicants, profit and loss statements and IRS Form 4506-C authorization.
- Asset and liability documentation. Bank statements (typically 2-3 months), retirement account statements, investment account statements, and current liability statements for existing debts.
- Identity documents. Government-issued photo ID, Social Security documentation, and for non-citizen applicants, relevant immigration or residency documentation.
- Credit and application forms. Uniform Residential Loan Application (URLA-1003), credit authorization forms, and credit report disclosures.
- Regulatory disclosures. Loan Estimate (LE), Closing Disclosure (CD), and lender-specific compliance disclosures.
Each of these arrives with format variation across lenders, employers, tax years, and jurisdictions. The pay stub from a large employer payroll system looks different from a pay stub generated by a small business accounting tool. W-2 forms are standardized by the IRS but vary in layout across issuers. Bank statements vary by institution and country, as covered in detail on the bank statement extraction page.
The ADE Pipeline for Mortgage Packages
The production workflow for a mortgage application package uses three ADE APIs in sequence.
Step 1: Parse the package. The entire mortgage package is submitted to the Parse API (for standard-sized files) or the Parse Jobs API (for large packages up to 1 GB or 6,000 pages). The parse step converts the entire package into layout-aware Markdown and hierarchical JSON with bounding-box coordinates on every block across all pages, preserving the structure of every document type in the package.
Step 2: Extract by document type. The Extract API is called once per document type with a schema tailored to that type's fields. The same parsed Markdown output is reused across all extract calls, so the package is parsed once and queried multiple times without re-processing. The W-2 schema extracts employer name, employee name, wages, and withholding totals. The bank statement schema extracts account numbers, balances, and transaction arrays. The pay stub schema extracts gross pay, net pay, year-to-date totals, and employer details.
Step 3: Route on confidence. Each parsed chunk carries a confidence score, and every extracted value includes a bounding-box citation linking it back to its page and location in the source document. Chunks below the confidence threshold route to a reviewer with the source citation pre-populated; chunks above the threshold proceed to extraction and pass to the underwriting system automatically.
Note on the Split API: ADE also provides a Split API that classifies and separates multi-document packages into individual sub-documents. Split is currently in Preview and not recommended for production use. The production approach described above uses the Parse and Extract APIs directly.
Per-Document-Type Schema Design
Each document type in a mortgage package has a distinct extraction schema. The schema specifies field names, types, and descriptions without encoding layout coordinates, so it handles format variation across issuers without modification.
Key schema patterns for mortgage document types:
| Document type | Extracted fields | Schema notes |
|---|---|---|
| W-2 | Employer name and EIN, employee name and SSN, wages (Box 1), federal withholding (Box 2), state wages and withholding, tax year | Enum for tax year; nullable for boxes not present on all W-2 variants |
| Pay stub | Employer name, employee name, pay period dates, gross pay, net pay, YTD gross, YTD deductions, pay frequency | Array for deduction line items; typed numeric for all monetary fields |
| IRS Form 1040 | Taxpayer name and SSN, filing status, adjusted gross income, total tax, taxable income, tax year | Multi-year packages use array schema; nullable for schedules not present |
| Bank statement | Account holder name, account number, institution name, statement period, opening balance, closing balance, transactions (date, description, amount, running balance) | Transaction array with typed fields; see bank statement extraction page |
| Identity document | Full name, date of birth, document number, expiry date, issuing country/state, document type | Uses attestation chunk type detection for signatures and seals |
| URLA-1003 | Borrower name, co-borrower name, property address, loan amount, loan purpose, employment information, declared assets and liabilities | Complex nested schema; arrays for employment history and liability items |
Schemas can be designed and validated interactively in the Schema Wizard Playground before production integration.
Compliance and Auditability in Lending Workflows
Mortgage underwriting operates under regulatory requirements (including RESPA, TILA, and HMDA in the US) that demand audit trails linking every underwriting decision to source documentation. ADE's bounding-box grounding satisfies this requirement at the field level: every extracted value carries its page number and coordinates in the source document, creating a verifiable chain from extracted field to source location that survives downstream processing and storage.
Zero Data Retention ensures that mortgage application documents containing PII and financial data are processed in memory without storage on LandingAI infrastructure. HIPAA support with BAA and SOC 2 Type II certification are documented at the Trust Center, covering the compliance posture relevant to financial services document workflows.
FAQ
Does ADE require a separate configuration for each document type in a mortgage package? No. ADE's visual-first parsing handles all document types in the package through a single parse step without per-document-type configuration. Each document type requires its own extraction schema, but the schema defines which fields to extract without encoding layout coordinates. A W-2 schema works across all W-2 variants from all employers because ADE's parsing layer identifies the document structure visually rather than by matching a stored template.
How does ADE handle a mortgage package that arrives as a single combined PDF? The entire package is submitted to the Parse API or Parse Jobs API as a single file. The parsed output preserves the structure of every document in the package across all pages. Multiple Extract calls with different schemas then run against the same parsed output to pull fields from each document type. The Split API can classify and separate the package into individual sub-documents, but it is currently in Preview and not recommended for production use.
What happens when a mortgage package contains a document type the schema does not cover? Fields defined in the schema return null for missing documents, using explicit null returns rather than omitted keys (with the extract-20251024 model). See extraction model versions for null handling behavior. Documents in the parsed output that are not targeted by any extraction schema are preserved in the parsed Markdown and can be extracted later by adding a schema for that document type, without re-parsing the package.
How does ADE handle scanned mortgage documents with variable scan quality? ADE's Document Pre-Trained Transformer architecture treats documents as visual systems and handles variable scan quality without requiring image pre-processing. The same model that parses a clean digital PDF parses a scanned pay stub or a photographed bank statement using the same visual reasoning process. Confidence scores on extracted fields reflect scan quality effects on extraction certainty, routing affected fields to human review.
Is ADE compliant with US mortgage lending regulatory requirements? ADE provides the technical infrastructure for compliance: audit-trail grounding linking every extracted value to its source location, Zero Data Retention for PII handling, SOC 2 Type II certification, and HIPAA BAA availability. Whether ADE's use in a specific lending workflow meets the requirements of RESPA, TILA, HMDA, or state-specific lending regulations depends on how it is integrated into the overall compliance architecture. Contact LandingAI through the financial services page to discuss compliance requirements for specific lending workflows.