Rewriting the Rules: How Microsoft AI is Revolutionizing Mortgage Underwriting

Freddie Mac Bulletin 2025-16 took effect March 3, 2026. It is not guidance. It is a requirement. Any seller or servicer delivering loans to Freddie Mac must now implement continuous, risk-based AI governance covering every use of AI or machine learning across the entire loan lifecycle: underwriting automation, document intelligence, chatbots, fraud detection, income calculation, and borrower outreach. Annual check-the-box reviews are no longer sufficient.

The adoption numbers explain why Freddie Mac acted. Fifty-five percent of mortgage lenders now use AI technologies in their operations. Gateless reports 70% to 75% auto-clearing rates on its Smart Underwrite platform, with projections of 85% by late 2026. GreenState Credit Union increased its general approval rate by 26% and its approval rate for protected classes by 32% using AI underwriting, with no added risk. Microsoft launched Azure AI Foundry Agent Service with multi-agent workflows that handle document verification, eligibility assessment, and loan calculations through GPT-4o powered agents. AI in underwriting is the production environment, not a pilot program.

This guide covers how Microsoft AI tools work at each stage of mortgage underwriting, what Freddie Mac's governance requirements mean for your implementation, and how to build an AI-assisted underwriting operation that passes compliance scrutiny. It also covers where ABT's MortgageExchange data layer, the Azure AI Foundry environments ABT hosts for mortgage companies, and the Microsoft 365 Copilot workspace under the M365 Guardian operating model fit into a real deployment.

Why AI Underwriting Is Spreading Across the Industry

Traditional underwriting has a math problem. Underwriters spend roughly 70% of their time on document review and data verification. That leaves 30% for actual risk assessment, the part that requires judgment, experience, and pattern recognition. AI flips that ratio.

Three forces accelerated adoption in the past 12 months:

Speed is now a competitive requirement. Lenders using AI-assisted underwriting approve loans in minutes, not days. When a borrower can get a decision from one lender in 20 minutes and from another in 5 days, the slower lender loses the deal. AI-assisted underwriting tools deliver average savings of $900 to $1,200 per loan in processing costs.

Document volumes keep growing. Fannie Mae's UAD 3.6 and UCD v2.0 updates added new data requirements. State regulations layer additional documentation demands. The average loan file now exceeds 500 pages. No human team can scale with that growth without AI-assisted document processing.

Fair lending scrutiny is increasing. The Government Accountability Office reported in May 2025 that financial agencies are training staff, forming AI working groups, and collaborating internationally to build AI oversight capabilities. The GAO has since called on the Federal Housing Finance Agency to provide guidance to Fannie Mae and Freddie Mac regarding fair lending requirements for property technology deployed in the homebuying process. Companies that deploy AI without governance frameworks face regulatory problems. Freddie Mac's Bulletin 2025-16 made that explicit.

The Microsoft AI Stack for Mortgage Underwriting

Microsoft offers five tools that form the AI underwriting stack. Each handles a different part of the process.

Azure AI Document Intelligence (formerly Form Recognizer) extracts data from mortgage documents using optical character recognition and natural language processing. It reads pay stubs, W-2s, tax returns, bank statements, and appraisal reports. It handles PDFs, scanned images, and photographed documents. Extraction accuracy exceeds 95% on structured financial documents.

Azure Machine Learning builds the predictive models that assess borrower risk. Train models on your historical loan performance data to predict repayment probability, identify risk factors, and score applications based on hundreds of variables beyond traditional credit metrics.

Azure AI Foundry Agent Service orchestrates multi-agent workflows. Microsoft published a detailed architecture in 2025 showing GPT-4o powered agents handling document verification, eligibility assessment, loan calculations, packet assembly, and digital signing integration. Each agent specializes in one task. The orchestration layer coordinates the sequence.

Microsoft Purview handles AI governance and compliance. It provides audit trails for every AI-assisted decision, tracks model performance, and enforces data access policies. This is the tool that helps you meet Freddie Mac's continuous governance requirements.

Power BI creates dashboards that give underwriting managers real-time visibility into AI-assisted pipeline metrics: auto-clearing rates, exception rates, turn times, and model confidence scores.

Stage 1: AI-Powered Document Processing

Every underwriting decision starts with documents. The faster and more accurately you process them, the faster you reach a decision.

Intake and classification. Azure AI Document Intelligence receives the complete document package and classifies each file by type: income documentation, asset documentation, property records, credit reports, and disclosures. Classification happens in seconds, not minutes. Misfiled documents get flagged and re-routed automatically.

Data extraction. The AI extracts structured data from each document. From a W-2: employer name, wages, federal tax withheld, state tax withheld, Social Security wages. From a bank statement: account holder, account number, ending balance, average balance, large deposits. From an appraisal: property address, estimated value, comparable sales, condition rating.

Cross-referencing. Extracted data flows into validation rules. Does the income on the pay stub match the income on the tax return? Does the deposit pattern on the bank statement support the stated income? Does the property address on the appraisal match the purchase contract? The AI flags discrepancies for human review instead of requiring underwriters to find them manually.

Completeness checking. The system compares received documents against the required checklist for the specific loan type (conventional, FHA, VA, jumbo). Missing documents generate automated requests to the borrower or processor before the file reaches the underwriter's queue.

Stage 2: Income and Employment Verification

Income verification is where traditional underwriting consumes the most time and where AI delivers the largest savings.

Multi-source income analysis. The AI analyzes income across document types simultaneously. Pay stubs provide current earnings. Tax returns provide annual income trends. Bank deposits provide actual cash flow. The system reconciles these sources and flags significant discrepancies. A borrower claiming $8,000 per month on the application but showing $5,500 per month in bank deposits gets flagged, not approved.

Self-employed income calculation. Self-employed borrowers produce the most complex income documentation. Two years of tax returns, profit and loss statements, Schedule C or K-1 forms, and business bank statements. Azure Machine Learning models trained on self-employed loan data calculate qualifying income from these documents in seconds, applying the same depreciation add-back and business expense analysis that experienced underwriters perform manually.

Employment stability assessment. The AI evaluates employment history for gaps, job changes, and industry risk factors. A borrower with 10 years at the same employer in a stable industry scores differently than one with three job changes in two years in a volatile sector.

Alternative data analysis. For borrowers with non-traditional income, Azure ML models analyze rental payment history, utility payment patterns, and banking transaction trends. This expanded data view supports credit-worthy borrowers who may not qualify under rigid traditional analysis, supporting both fair lending goals and business growth.

Stage 3: Credit and Risk Assessment

AI does not replace the underwriter's risk judgment. It gives the underwriter better data and more time to apply that judgment.

Predictive risk modeling. Azure Machine Learning models analyze hundreds of variables across your historical loan portfolio. Beyond credit score and DTI ratio, the models evaluate spending pattern changes, credit utilization trends, payment timing patterns, and geographic risk factors. The output is a risk probability score that supplements traditional AUS findings.

Exception detection. The AI identifies loans that fall outside standard guidelines but may still be approvable with compensating factors. A borrower with a 38% DTI but substantial reserves and excellent payment history gets flagged for exception review rather than automatic decline.

Fraud pattern recognition. Machine learning models trained on known fraud cases identify suspicious patterns: straw buyer indicators, income inflation red flags, property flipping schemes, and identity inconsistencies. These patterns are difficult for human reviewers to spot across individual files but visible to AI systems analyzing data at scale.

Collateral risk assessment. AI models analyze property data, comparable sales trends, and market conditions to assess collateral risk. A property in a declining market with limited comparable sales receives a different risk treatment than one in a stable market with strong recent sales activity.

The Data Layer: Where MortgageExchange and Azure AI Foundry Fit

An AI underwriting engine is only as good as the data it sees. The lenders that get production value from Azure AI Document Intelligence, Azure Machine Learning, and Azure AI Foundry are not the ones with the smartest models. They are the ones that have already solved the data pipeline problem. Borrower applications, employment verification, asset documentation, credit bureau pulls, and product configurations have to leave the loan origination system in a structured form, land in a tenant-isolated data environment, and reach the AI models tagged with the right loan type, product code, and program. Without that pipeline, the AI is reading a pile of unstructured PDFs and best-guessing what it sees. With it, the AI is reading clean, normalized data that matches the institution's actual lending guidelines.

ABT operates that data layer for hundreds of mortgage companies, banks, and credit unions under the product name MortgageExchange. It is the managed integration that connects an institution's loan origination system, such as Encompass, Calyx, or Empower, to the core banking platform, such as Fiserv DNA, Symitar Episys, or Jack Henry, and from there into the data environment where the AI underwriting models run. Without that interface, an underwriter is rekeying borrower data from one system into another. With it, the application arrives at the AI pre-populated, validated against the core, and tagged with the right product and program codes. For more on the integration layer, see our overview of mortgage data integration for financial institutions.

The next layer is where the AI underwriting actually runs. ABT hosts the Microsoft Azure AI Foundry environment that operates on the MortgageExchange data lake for mortgage clients that want production AI underwriting without standing up an Azure tenant from scratch. The Foundry environment runs the document intelligence models, the Azure Machine Learning risk models, and the multi-agent workflows that handle document verification, eligibility assessment, and loan calculations against the institution's actual lending guidelines. The institution continues to own the loans, the customers, and the regulatory relationships. ABT operates the Azure subscription that runs the models, configures the data boundaries so the AI is grounded to the institution's own loan data rather than to public training corpora, and produces the Microsoft Purview audit log that Freddie Mac Bulletin 2025-16 requires for every AI-assisted decision.

The Underwriter Workspace: Microsoft 365 Copilot Inside M365 Guardian

An AI underwriting decision is not the end of the workflow. The underwriter still has to read the AI's findings, evaluate the exceptions, document the overrides, and communicate with the processor, the loan officer, and sometimes the borrower. That work happens in Outlook, Teams, Word, and Excel. The lenders that get the largest productivity gain from Azure-side AI close the loop with Microsoft 365 Copilot on the underwriter's desktop, where it can summarize the AI's exception findings inside Outlook, draft the override memo inside Word, surface the income-discrepancy details inside Teams, and reconcile cashflow numbers inside Excel. Copilot reads the institution's own loan files, emails, and conversations, grounds its answers in that tenant-isolated data, and cites the source documents back to the underwriter. It is the agentic surface that turns the Azure underwriting output into action without the underwriter switching between five different apps. ABT manages the Microsoft 365 tenant that Copilot runs in and configures the data boundaries so Copilot is grounded to the institution's own loan data, not the public internet. For pricing context on the SMB-first Copilot tiers, see our Microsoft 365 Copilot pricing buyer's guide.

The governance envelope around that whole pipeline, the Azure Foundry environment that runs the underwriting models, the MortgageExchange data lake that feeds them, the Copilot workspace that the underwriter uses, and the Outlook, Teams, SharePoint, and OneDrive surfaces that handle the resulting communications, runs under an ABT operating model branded M365 Guardian. ABT manages the Microsoft 365 tenant. ABT hosts the Azure subscriptions. Guardian layers Microsoft Entra ID Conditional Access for identity, Microsoft Defender for threat detection, Microsoft Purview for retention and data loss prevention, Microsoft Intune for device compliance, and Microsoft Sentinel for SIEM aggregation over the whole footprint. For an AI underwriting deployment, Guardian is what gives the chief compliance officer a single configuration-evidence story to hand to a Freddie Mac post-purchase reviewer or a state regulator examining for fair lending compliance.

Meeting Freddie Mac's New AI Governance Requirements

Freddie Mac Bulletin 2025-16 changed the compliance landscape for every lender using AI. The bulletin covers AI-enabled vendors and tools too, not just models you build in-house. Here is what your organization needs.

Cross-functional governance body. Freddie Mac requires a governance committee that includes leadership from IT, cybersecurity, risk management, legal, compliance, and business operations. This body oversees AI policy, reviews model performance, and makes decisions about AI deployment across the loan lifecycle.

Senior management accountability. The framework now requires CIO, CTO, or Chief Risk Officer sign-off on the lender's AI policies. This is not a committee recommendation. It is executive accountability for AI governance decisions.

Continuous model monitoring. Annual reviews are no longer sufficient. You need ongoing monitoring of model performance, accuracy, and bias detection. Microsoft Purview provides the audit infrastructure. Azure Machine Learning tracks model drift, prediction accuracy, and fairness metrics in real time.

Documented risk management. Every AI model needs a risk assessment covering intended use, data inputs, output interpretation, and failure modes. Freddie Mac wants to see that you understand what your AI does, how it makes decisions, and what happens when it gets something wrong.

AI-specific security guardrails. The framework defines guardrails to prevent AI-specific threats including model inversion, data poisoning, and prompt injection attacks. Lenders must separate responsibilities to prevent conflicts of interest in AI development and deployment.

Demonstrable human oversight. Human underwriters must review AI outputs, not rubber-stamp them. Your process needs to document how underwriters evaluate AI recommendations, when they override them, and how overrides are tracked.

Fair lending testing. Run regular adverse impact analyses on AI-assisted lending decisions. Compare approval rates, pricing, and terms across demographic groups. Azure Machine Learning's responsible AI toolkit includes fairness assessment features designed for this purpose.

Vendor oversight. If you use third-party AI tools (Gateless, ICE Mortgage Technology, or any vendor with AI in the workflow), you must demonstrate governance over those tools too. Freddie Mac's requirements cover the entire AI supply chain, not just your internal models.

For a complete compliance checklist and implementation timeline, see our companion article: Freddie Mac AI Mandate Compliance Checklist.

The 2026 AI Underwriting Adoption Landscape

The regulatory mandate arrived alongside an acceleration in AI adoption that makes governance more urgent, not less.

Budget commitments are growing. A Celent study commissioned by Zest AI found that 83% of lenders plan to increase their generative AI budgets in 2026. Two-thirds of lenders have already completed or will implement GenAI strategies by 2026, a faster adoption rate than previous lending technology waves.

Origination volumes respond to AI investment. Lenders implementing AI-powered platforms report a 50% increase in mortgage origination volume. Those without AI face an increasingly steep competitive disadvantage as borrowers expect faster decisions and streamlined processes.

Regulatory encouragement, not just enforcement. Over a third (37%) of lenders say the current regulatory environment encouraged them to accelerate AI adoption in underwriting. Freddie Mac's governance framework is not designed to block AI. It is designed to ensure AI is deployed responsibly.

For a broader view of how AI vendor risk affects mortgage companies, see our analysis of FHFA's AI vendor risk guidance.

The Human-Plus-AI Underwriting Model

The most effective AI underwriting implementations follow a clear division of labor.

AI handles: document classification, data extraction, cross-referencing, completeness checking, income calculation, initial risk scoring, fraud pattern detection, and compliance rule checking. These are high-volume, pattern-based tasks where speed and consistency matter more than judgment.

Humans handle: exception decisions, compensating factor evaluation, borrower conversations, complex income analysis for unusual situations, final risk determination, and override documentation. These are judgment-intensive tasks where experience and context matter more than speed.

The result is faster underwriting because quality improves. When the AI handles document review, the underwriter reads a structured summary instead of 500 pages. When the AI flags discrepancies, the underwriter investigates specific issues instead of searching for them. Gateless reports that this model produces 70% to 75% auto-clearing rates today, with the remaining files receiving focused human review.

Justin Kirsch

CEO, Access Business Technologies

Justin co-founded ABT in 1999 and leads the company's work with mortgage lenders navigating AI adoption and compliance. ABT operates the MortgageExchange data layer, the Microsoft Azure AI Foundry environment, and the Microsoft 365 Copilot workspace under the M365 Guardian operating model for hundreds of mortgage companies, banks, and credit unions building AI underwriting that passes Freddie Mac Bulletin 2025-16 governance review.