Updated November 7, 2026
AI mortgage underwriting uses software to read borrower documents, extract data, check income and eligibility against GSE guidelines, and flag exceptions for a human underwriter. Done well, it shortens cycle times and lowers per-loan cost. It also needs defined criteria, human review of exceptions, and compliance guardrails.
This guide covers the workflow in four stages, then weighs the benefits against the barriers lenders cite and shows where an Encompass® team can start.
How is AI used in mortgage underwriting?
AI underwriting sits between document intake and the credit decision. It reads files, pulls data, and checks that data against lending rules before an underwriter opens the loan.
It is no longer a fringe experiment. A STRATMOR Group survey found 38% of lenders were using artificial intelligence or machine learning in their operations as of 2024. A separate survey of more than 150 mortgage professionals found 57% believe AI-driven underwriting will be the most significant technological change in the industry this year.
The objective is not to automate the credit decision itself. It is to automate the work around that decision.
Lender Toolkit is a SaaS platform of automation tools, AI-powered solutions, and professional services built exclusively for mortgage lenders running ICE Mortgage Technology's Encompass LOS, including its AI Underwriter™ offerings.
Step 1: Pick one underwriting workflow to automate
Start narrow. Choose one underwriting segment with clearly defined service-level agreements, such as income documentation review or condition tracking. A defined SLA gives you a baseline to measure against. [S1]
Set your efficiency target with context. Freddie Mac reports that lenders who fully use its own automated underwriting systems can save $1,500 per loan and cut production cycles by five days. That figure applies to Freddie Mac's systems, so treat it as a benchmark for what is possible, not a promise for any other tool.
Step 2: Automate document processing with OCR and NLP
Document processing is where most manual touchpoints live. Optical character recognition (OCR) turns scanned pages into text. Natural language processing (NLP) classifies each document and extracts the fields that matter from pay stubs, bank statements, and tax forms.
Before you turn it on, write down the rules:
- Define which document types and formats the system accepts.
- Set exception handling: what happens when a page is unreadable, missing, or mismatched.
- Require an audit trail showing what was extracted and from where. [S2]
This keeps files clean, complete, and consistent before they reach the underwriter.
Step 3: Assess risk and eligibility against GSE guidelines
Once data is extracted, the system checks it against GSE guidelines for income, assets, and eligibility. This is where AI loan underwriting earns its keep, and where it can fail.
Rule-based logic suits this step because GSE guidelines are explicit. A fixed rule produces the same answer every time and can be traced. A general-purpose language model alone is a weaker fit. In a study using the MortarBench open-source benchmark for mortgage origination tasks, Gemini 3.1 Pro reached a correct answer rate of 77.1% on the strictest evaluation measure. That is impressive for a general model, but one wrong answer in roughly four is not acceptable for a credit file. Use rules for the checks that have a right answer, and send the rest to a person.
Step 4: Keep a human in the loop with compliance guardrails
AI does not replace your underwriters. It reduces the clerical work so they spend time on exceptions and judgment calls. Keep a human in control of every flagged file.
Train the underwriting team on the tool and build a feedback loop so reviewers can correct rules that misfire. [S3] This also answers the most common complaint about AI underwriting: errors and bias. Columbia University research found that general-purpose AI models wrongly flagged bank deposits as potentially of foreign origin 13.3% of the time when they were associated with English names. Human review, documented overrides, and regular testing of automated rules are your guardrails. For more on this, see Responsible Mortgage AI.
What are the benefits and barriers of AI underwriting?
The benefits are measurable: Freddie Mac reports savings of $1,500 per loan and five fewer days in production for lenders that fully use its systems. Adoption is already widespread, with 38% of lenders using AI or machine learning in operations as of 2024.
The barriers are just as real. A joint survey by the Mortgage Bankers Association, AARMR, and BCG found nearly 60% of lenders view regulatory and compliance uncertainty as the primary barrier to AI implementation. Accuracy, explainability, and accountability when a model gets a file wrong are fair concerns. Treat them as design requirements, not objections.
How can Encompass lenders get started with Lender Toolkit?
If manual touchpoints are slowing your underwriters, begin with automation inside Encompass rather than a separate system. Prism handles end-to-end loan condition and underwriting automation, including AI-driven income calculation. See the details in Inside Prism: The All-in-One Encompass Automation Engine.
The PowerTools suite has published pricing: Free at $0/month, Expert at $795/month, and Optimized at $1,495/month. See PowerTools for the plug-ins. Prism pricing is by consultation, so Schedule a Demo. Related reading: Unleash the Power of Encompass and How AI Underwriting Streamlines Mortgage Approvals and Minimizes Loan Losses. You can also see Lender Toolkit at ICE Experience 2025, watch walkthroughs on youtube, or browse the site with Search powered by Jetpack. More at Lender Toolkit.
FAQs
How is AI used in mortgage lending beyond underwriting?
Lenders also apply it to disclosures, post-closing tasks, condition tracking, and data entry. These areas share the same pattern: repetitive, rules-driven work that software can handle before a person reviews exceptions.
What is the biggest barrier to AI underwriting?
Regulatory and compliance uncertainty. Nearly 60% of lenders in the MBA, AARMR, and BCG survey named it the primary barrier. Pilot narrowly, measure results, and keep people in control of exceptions.
Can general-purpose AI models be trusted to underwrite a loan on their own?
Not on their own. A benchmark showed a top model at 77.1% on the strictest measure, which leaves too many wrong answers for a credit file. Use them under rules and human review.
Does AI underwriting need an audit trail?
Yes. Record what the system extracted, from which document, and every human override. That record lets your compliance team see how the rules behave and review any file that was flagged.
How do rule-based underwriting automation and an LLM differ in mortgage work?
GSE guidelines are explicit, so fixed rules suit guideline checks: they give the same answer every time and can be traced. A general-purpose language model is a weaker fit for those checks. Use rules for guideline checks and keep people on judgment calls.
Conclusion
AI mortgage underwriting works best when you automate one defined workflow, set clear document and exception rules, check data against GSE guidelines with traceable logic, and keep an underwriter in charge of exceptions.
Start with one segment and one SLA. Measure cycle time before and after, and document every override so your compliance team can see how the rules behave.


