AI Disclosure Policy

Margie — Pre-Underwriting Analysis Platform

Effective Date: March 23, 2026

Last Updated: March 23, 2026

Document Classification: Compliance — AI Governance

1. Purpose

This AI Disclosure Policy describes how Margie (“Platform”), operated by Margie (“Company,” “we,” “us”), uses artificial intelligence services in its pre-underwriting analysis of mortgage borrower documents. This policy is intended to provide transparency to mortgage brokerages, their compliance officers, and regulators regarding the role, scope, and limitations of AI within the Platform.

Margie is a pre-underwriting analysis tool designed for licensed Illinois mortgage brokers. It does not originate loans, make credit decisions, or serve as a substitute for professional underwriting judgment.

2. Overview of AI Usage

Margie processes borrower-submitted documents through a six-stage forensic analysis engine. AI is used in two narrowly scoped capacities within this pipeline:

1. Optical Character Recognition (OCR) and Structured Field Extraction — AI reads uploaded documents and extracts discrete data fields (e.g., employer name, gross pay, account balances, tax filing amounts).
2. Narrative Analysis Generation — AI produces human-readable commentary summarizing the findings of each analysis stage for broker review.

AI is not used to make decisions, score borrowers, predict outcomes, or determine loan eligibility. The distinction between AI-assisted processing and deterministic analysis is fundamental to Margie's architecture and is described in detail in Section 5.

3. AI Service Providers

Margie integrates three third-party AI services, each serving a distinct function. No single provider has access to a borrower's complete file.

3.1 Microsoft Azure Document Intelligence

• Function: OCR and structured field extraction from borrower documents.
• Models Used: Prebuilt extraction models including prebuilt-tax.us.w2, prebuilt-payStub.us, prebuilt-bankStatement.us, prebuilt-idDocument, prebuilt-mortgage.us.1003, and additional prebuilt models as applicable.
• Data Transmitted: Uploaded document images or PDFs, transmitted via TLS 1.2+ encrypted connections.
• Data Retention: Azure Document Intelligence does not retain customer document data after processing is complete.
• Model Training: Microsoft's prebuilt models are trained on Microsoft-curated datasets. Customer documents are not used to train, retrain, or improve Microsoft's models.

3.2 Anthropic Claude (claude-sonnet-4-5)

• Function: Generation of narrative analysis commentary for each of Margie's six analysis stages.
• Data Transmitted: Structured field data extracted during earlier processing stages. Raw borrower documents are not sent to Anthropic.
• Data Retention: Under Anthropic's Commercial Terms of Service, API inputs and outputs may be retained for up to 7 days solely for service delivery purposes. Data is automatically deleted after this period.
• Model Training: Anthropic's Commercial Terms of Service contractually prohibit the use of customer API data for model training.

3.3 Google Gemini (gemini-2.5-flash)

• Function: Automated field population for single-borrower document extraction via semantic parsing.
• Data Transmitted: Extracted text from borrower documents, transmitted via TLS-encrypted connections.
• Data Retention: Under Google's Gemini API terms, data may be retained for up to 55 days for automated abuse monitoring.
• Model Training: Under Google Cloud's Service Specific Terms for paid API usage, customer data is not used to train or improve Google's foundation models.

4. How AI Is Not Used

The following limitations are architectural constraints of the Platform, not merely policy commitments:

• AI does not make credit decisions. The Platform does not approve, deny, or recommend loan applications.
• AI does not access credit reports or credit scores. No credit bureau data is transmitted to or processed by any AI service.
• AI does not determine loan eligibility or pricing.
• AI does not perform automated decision-making as defined under applicable consumer protection law.
• AI does not generate risk scores, creditworthiness assessments, or predictive analytics of any kind.
• No AI model is trained on borrower data submitted through the Platform.

5. AI-Assisted vs. Deterministic Processing

Margie's analysis pipeline maintains a strict separation between AI-assisted processing and deterministic rule-based analysis.

5.1 AI-Assisted Processing

AI is used at two points in the pipeline:

AI-assisted outputs serve as inputs to the deterministic engine or as presentation layers for broker review. They do not independently determine findings.

5.2 Deterministic Analysis Engine

Margie's core underwriting analysis is performed by a deterministic rules-based engine spanning six evaluation stages. This engine executes 66 discrete compliance and consistency checks, including but not limited to:

• Income calculation verification and year-over-year trending
• Employment tenure and stability assessment
• Asset sufficiency and seasoning validation
• Debt-to-income ratio computation
• Document cross-referencing and consistency checks
• Regulatory compliance flag identification

Every check in this engine follows predefined rules and thresholds. No machine learning, neural networks, or probabilistic AI models are used in the evaluation logic. Results are deterministic: the same inputs will always produce the same outputs.

5.3 Why This Distinction Matters

• AI extraction errors can be caught by deterministic cross-checks
• Analysis findings are reproducible and auditable
• No “black box” model influences compliance-relevant determinations
• Brokers can trace any finding back to specific rules and source data

6. Human Oversight Requirements

Margie is designed as a broker-in-the-loop system. The following human oversight requirements are built into the Platform:

1. All AI-generated narrative content is clearly labeled as AI-generated within the Platform interface and in exported reports.
2. Brokers review all findings before taking any action based on Platform output. The Platform does not transmit results to borrowers, lenders, or third parties.
3. The Platform includes explicit disclaimers stating that its output is not a substitute for professional underwriting judgment.
4. Regulatory disclaimers referencing the Equal Credit Opportunity Act (ECOA), Real Estate Settlement Procedures Act (RESPA), and Gramm-Leach-Bliley Act (GLBA) are displayed on all outputs.
5. Brokers retain full discretion over how (and whether) to act on any finding.

7. Data Handling and Security

7.1 Document Processing

• Borrower documents are processed in memory only during analysis. Documents are not written to persistent storage on Margie's servers.
• Extracted field data and analysis results are encrypted at rest using AES-256-GCM encryption.
• All data in transit is protected by TLS 1.2+ encryption.

7.2 Data Minimization

• Each AI provider receives only the minimum data necessary for its specific function.
• Azure receives document images for OCR. Anthropic receives extracted fields (not documents) for narrative generation. Google receives extracted text for field population.
• No single AI provider receives a borrower's complete file or the full output of Margie's analysis.

7.3 Infrastructure

• Application: Next.js deployed on Vercel
• Database: Google Firebase Firestore (encrypted at rest)
• Billing: Stripe (PCI DSS Level 1 compliant)
• Domain: getmargie.com / app.getmargie.com

For complete information on data handling, refer to Margie's Privacy Policy.

8. Model Training Prohibitions — Summary

Margie does not fine-tune, customize, or contribute to the training of any third-party AI model using borrower data.

9. Accuracy, Limitations, and Disclaimers

9.1 OCR and Extraction Accuracy

• Handwritten text, poor-quality scans, unusual document formats, or non-standard layouts may reduce extraction accuracy.
• Extracted field values should be treated as preliminary until confirmed by the deterministic analysis engine's cross-referencing checks or by broker review.
• The Platform's deterministic checks are specifically designed to catch extraction inconsistencies.

9.2 Narrative Generation Accuracy

• Narratives may occasionally contain imprecise characterizations of findings.
• Narratives are summaries for broker convenience, not authoritative compliance determinations.
• All narrative output is clearly labeled as AI-generated.
• Brokers should verify narrative claims against the underlying extracted data and deterministic findings.

9.3 General Disclaimers

• Margie is a pre-underwriting analysis tool, not a licensed underwriter, lender, or credit decision system.
• Platform output does not constitute a loan approval, denial, pre-qualification, or pre-approval.
• Platform output does not constitute legal, financial, or compliance advice.
• All findings are advisory and require review by a licensed mortgage professional before any action is taken.
• The Platform complies with applicable provisions of the GLBA, ECOA, and RESPA. Compliance disclaimers are affixed to all outputs.

10. Contact Information

Questions regarding this AI Disclosure Policy or Margie's use of AI should be directed to:

Margie
Email: contact@getmargie.com
Website: getmargie.com

11. Policy Updates

This policy may be updated to reflect changes in AI service providers, processing methods, or applicable regulatory guidance. Material changes will be communicated to active subscribers. The “Last Updated” date at the top of this document reflects the most recent revision.

This document is provided for compliance transparency purposes and is intended to be reviewed alongside Margie's Privacy Policy, Terms of Service, and GLBA Compliance Statement. A Data Processing Agreement is available upon request for brokerage partners.