Know a company with years of records? Refer it and earn up to $100K
Mortgage lending

AI Mortgage Underwriting: How It Works and the Rules

How AI is used in mortgage lending and underwriting: Desktop Underwriter and LPA, document and income checks, adverse action rules, loan files and HMDA data.

Short answerChecked

AI mortgage underwriting is software that reads a loan file, checks income, assets and credit, and recommends a decision. The core tools are the automated underwriting systems of Fannie Mae (Desktop Underwriter) and Freddie Mac (Loan Product Advisor); newer tools sort documents and pull out figures. An underwriter still clears the conditions and decides, and a lender that denies credit must give specific, accurate reasons, whatever model it used.

Automated underwriting systems
2Desktop Underwriter (Fannie Mae) and Loan Product Advisor (Freddie Mac)
Reasons on a denial notice
Usually 4 or fewerRegulation B official interpretation
HMDA fields redacted for privacy
27FFIEC modified LAR
A signed licensing agreement with a fountain pen and archive boxes
Key points
  • Most "mortgage AI" sits in four places: document sorting, income and asset checks, automated underwriting and quality control.
  • Desktop Underwriter and Loan Product Advisor return a recommendation and a list of conditions. The lender makes the decision.
  • A denial needs specific reasons that match the factors the model actually used, even though the CFPB withdrew its 2022 and 2023 circulars on the point.
  • Public HMDA data shows who applied and what happened, with privacy fields removed. It holds none of the documents in the file.

How is AI used in mortgage lending?

In mortgage lending, "AI" covers several tools at different steps of one loan. Some are rules engines that have run for decades. Others are newer machine learning models that read scanned pages or spot patterns. The table shows where each sits and what a person still does.

Regulators treat these tools as models to be governed. The Federal Housing Finance Agency (FHFA), which oversees Fannie Mae and Freddie Mac, issued guidance to both "on managing risks associated with the use of artificial intelligence and machine learning (AI/ML)" (FHFA AB 2022-02). The bulletin is dated "AB 2022-02 (February 10, 2022; Revised May 2025)".

StepWhat the software doesWhat a person still does
Application intakeReads the application and flags missing answers or documentsThe loan officer checks that the story makes sense
Document sortingClassifies uploaded pages (pay stub, W-2, bank statement, tax return) and pulls out figuresThe processor fixes misread or missing pages
Income and asset checksCalculates qualifying income and matches it against payroll and bank dataThe underwriter reviews self-employed, variable or unusual income
Automated underwritingRuns Desktop Underwriter or Loan Product Advisor and returns a recommendation with conditionsThe underwriter clears conditions and makes the decision
Fraud and quality controlFlags figures that do not match across documentsQuality control staff review samples before and after closing
Denial noticesRecords the factors behind a negative resultCompliance staff check that the stated reasons are specific and accurate

What are Desktop Underwriter and Loan Product Advisor?

They are the automated underwriting systems (AUS) of the two government-sponsored enterprises. Desktop Underwriter (DU) "is Fannie Mae's automated mortgage loan underwriting system" (Fannie Mae). Lenders submit to it directly, and "Sponsored mortgage brokers have access to DU through Desktop Originator" (DO). DU returns a recommendation such as Approve/Eligible or Refer with Caution (Fannie Mae Selling Guide). Fannie Mae says "DU conducts its analysis uniformly, and without regard to race, gender, or other prohibited factors."

Loan Product Advisor (LPA) "is Freddie Mac's automated underwriting system (AUS) designed to provide greater certainty that loans meet Freddie Mac eligibility requirements" (Freddie Mac). Its results are risk classes: a loan that meets the requirements is an Accept, and other loans may receive a Caution (Freddie Mac LPA FAQ).

An AUS recommends; it does not decide. The lender must still "include documentation in the loan file that supports the lender's assessment of the borrower's credit history, employment and income, assets, and other financial information used for qualifying" (Fannie Mae Selling Guide).

Desktop Underwriter (DU)Loan Product Advisor (LPA)
Run byFannie MaeFreddie Mac
Main resultsApprove/Eligible, Approve/Ineligible, Refer with Caution, Out of ScopeAccept or Caution risk class, with feedback messages
Income and asset toolsDU validation service; Income CalculatorAsset and income modeler (AIM)

How does AI mortgage underwriting check documents, income and assets?

Document tools work in two passes. First they classify each page: this is a pay stub, this is a bank statement, this is page two of a tax return. Then they extract fields, such as employer, gross pay, statement balance or large deposits, and compare them with what the borrower typed into the application. A name or employer that does not match becomes a condition for the underwriter.

Income and asset checks increasingly use data rather than paper. Fannie Mae's Income Calculator is built to "calculate income from wage earnings, self-employment, business ownership, and rental properties using borrower income documentation and tax return data" (Fannie Mae). Its DU validation service lets lenders "Digitally validate assets, income, and employment" from third-party reports (Fannie Mae). Freddie Mac's AIM is a capability within LPA "that automates the assessment of borrower assets, income and employment for lenders" (Freddie Mac).

The hard cases stay with people: self-employed borrowers with falling income, gift funds, recent job changes, poor scans and handwritten letters of explanation. Those are the files where a misread page or a missing explanation causes trouble later, so the underwriter reads them in full.

What does a mortgage loan file contain?

It starts with the application. Fannie Mae and Freddie Mac "redesigned the Uniform Residential Loan Application (URLA) (Fannie Mae Form 1003)" so that application data is captured the same way everywhere (Fannie Mae). That application data is only a small part of the file. Washington's mortgage broker rule gives a practical list of the rest: "All forms of loan applications, written or electronic (the Fannie Mae 1003 is an example)", and third-party documents "including, but not limited to, credit, appraisal, title, verifications of employment and deposits, automated underwriting results" (Washington WAC 208-660-450).

Part of the fileTypical documentsWhat it shows
ApplicationForm 1003, initial disclosures, rate lockWhat the borrower stated and what was offered
CreditCredit report, explanations of inquiries and late paymentsPayment history at the time of application
Income and employmentPay stubs, W-2s, tax returns, verification of employmentAbility to repay
AssetsBank statements, gift letters, retirement statementsDown payment, closing costs and reserves
PropertyAppraisal, title report, insuranceCollateral value and ownership
UnderwritingAUS findings, conditions list, underwriter notes, exceptionsThe decision and what it depended on
Closing and afterClosing disclosure, settlement statement, quality control findingsFinal terms and any errors found later

A loan tape is the one-row-per-loan summary of many files; see the loan tape template.

What do fair lending and adverse action rules require when lenders use AI?

The Equal Credit Opportunity Act (ECOA) and its Regulation B require a creditor that denies credit, or takes other adverse action, to give the applicant specific reasons. The official interpretation says the reasons "must relate to and accurately describe the factors actually considered or scored by a creditor", and that "disclosure of more than four reasons is not likely to be helpful to the applicant" (Regulation B interpretation).

In 2022 the CFPB applied this to AI directly: "The adverse action notice requirements of ECOA and Regulation B, however, apply equally to all credit decisions, regardless of the technology used to make them." It added that the rules "do not permit creditors to use complex algorithms when doing so means they cannot provide the specific and accurate reasons for adverse actions" (CFPB Circular 2022-03). The CFPB withdrew that circular, and its 2023 follow-up on sample forms, on May 12, 2025 (CFPB withdrawn guidance). The withdrawal removed the guidance documents, not Regulation B's requirement of specific reasons.

FHFA makes the same point for Fannie Mae and Freddie Mac: "the use of AI/ML-based credit underwriting models in credit decision-making can present compliance risks due to a lack of explainability of the model" (FHFA). In practice, a lender should be able to trace each denial reason to an input the model really used. This page is general information, not legal advice.

What is HMDA data, and what does it leave out?

"The Home Mortgage Disclosure Act (HMDA) requires many financial institutions to collect, report, and publicly disclose information about mortgage loans" (CFPB). It is one row per application: loan type, purpose, amount, the action taken, and coded reasons for denial. "HMDA was originally enacted by Congress in 1975 and is implemented by Regulation C." The 2025 data, published on March 31, 2026, cover "approximately 4,768 HMDA filers" (CFPB).

"The publicly available data are modified to protect applicants' and borrowers' privacy" (CFPB). In the public file, "27 fields have been redacted for privacy reasons", including the loan identifier, application date, action taken date, street address, city, ZIP code and credit scores. Loan amounts are rounded to the midpoint of a ten-thousand-dollar band, and ages are grouped into bands (FFIEC).

So public HMDA data shows who applied and what happened to the application. It holds none of the documents, conditions or underwriter notes, and nothing about how the loan paid after closing. That is why it cannot stand in for a loan file.

Example: one application through an AI-assisted underwriting flow

An illustration written for this page, with a fictional borrower and fictional figures. A salaried nurse with weekend agency shifts applies for a 30-year fixed purchase loan of $380,000 on a $425,000 home.

StepWhat the tools didWhat the people did
Day 0: applicationForm 1003 submitted; 61 pages uploaded and classifiedLoan officer asked for the missing second bank statement
Day 1: incomeBase pay of $7,150 a month read from pay stubs and W-2s; agency shifts flagged as variable incomeProcessor ordered a two-year history of the agency income
Day 1: assetsA $12,000 deposit flagged as large and unsourcedBorrower supplied a gift letter from a parent
Day 2: AUSApprove/Eligible with 9 conditionsUnderwriter added one condition for the gift funds transfer
Day 9: conditionsTwo documents matched automaticallyUnderwriter cleared the rest and signed the approval
Day 30: closingClosing figures checked against the approvalQuality control picked the file for a post-closing review

Every condition, document and decision above stays in the file with a date. That record of what was asked, what was supplied and what was decided is what makes an underwriting file more than its application data.

Common mistakes when lenders add AI to underwriting

Most problems come from trusting the output more than the process behind it.

  • Using a model whose denial reasons cannot be traced to the inputs it used, then filling the notice with generic reasons.
  • Letting extracted figures overwrite what the underwriter verified, so the file no longer shows who decided what.
  • Training or tuning a model on approved loans only, so it never sees what a denial or a default looked like.
  • Skipping a fair lending review of new data sources, such as bank transaction data, before they affect decisions.
  • Keeping no record of the model version that scored each application.
  • Treating an AUS recommendation as the decision and skipping the underwriter's own review of conditions.

Are old mortgage loan files also a data asset?

They can be. AI data buyers have asked Origin for loan files from mortgage brokers and lenders with their outcomes, compiled across several years, with borrower identities removed first. The reason is the one in the example above: a file ties a credit decision to the documents, conditions and judgment behind it, and to what happened next. Public HMDA data and AUS findings alone cannot show that.

What comes out first: borrower names, Social Security and account numbers, addresses and phone numbers; credit reports, which the FCRA limits to permissible purposes; tax returns and bank statements as documents; and anything lender agreements restrict. Consumer borrower data also falls under Regulation P. The guide for mortgage brokers covers those rules and a sample-first process, and de-identification before licensing covers the method. Origin is an independent data partner. Your company decides what is shared and signs any license itself, and a buyer may pay Origin when a deal closes.

Does your company hold records like these? See what your data is worth.

Common questions

What is AI mortgage underwriting?

AI mortgage underwriting is the use of software to read a loan file, check income, assets and credit, and recommend a decision. It includes automated underwriting systems such as Fannie Mae's Desktop Underwriter and Freddie Mac's Loan Product Advisor, plus newer tools that sort documents and extract figures. An underwriter still clears conditions and makes the final decision.

Do mortgage lenders use AI to approve loans?

Lenders use automated systems to recommend approvals, but a person decides. Desktop Underwriter returns recommendations such as Approve/Eligible or Refer with Caution, and Loan Product Advisor returns an Accept or Caution risk class. The underwriter reviews the conditions those systems set, checks the documents and signs the approval. Fannie Mae's guide requires file documentation that supports the lender's own assessment.

What is the difference between Desktop Underwriter and Loan Product Advisor?

Desktop Underwriter is Fannie Mae's automated underwriting system and Loan Product Advisor is Freddie Mac's. Both assess credit risk and eligibility for sale to their owner. DU returns recommendations such as Approve/Eligible; LPA returns risk classes such as Accept or Caution. Brokers reach DU through Desktop Originator, sponsored by a lender.

Can a lender deny a mortgage based on an AI model without giving specific reasons?

No. Regulation B requires specific reasons for a denial, and its official interpretation says they must accurately describe the factors actually considered or scored. The CFPB withdrew its 2022 and 2023 circulars on complex algorithms in May 2025, but the regulation itself still applies to every credit decision, however it is made.

What documents are in a mortgage loan file?

A mortgage loan file holds the application (Form 1003) and disclosures, a credit report, pay stubs, W-2s and tax returns, bank statements and gift letters, the appraisal and title report, automated underwriting findings, the conditions list and underwriter notes, and closing documents such as the closing disclosure and settlement statement.

What is HMDA data used for?

According to the CFPB, HMDA data helps show whether lenders are serving the housing needs of their communities, gives public officials information for policy decisions, and sheds light on lending patterns that could be discriminatory. Many lenders must report each mortgage application each year under Regulation C. The public version is modified to protect privacy, so 27 fields are redacted.

Does public HMDA data include credit scores or addresses?

No. The public modified loan application register redacts 27 fields, including credit scores, street address, city, ZIP code, the loan identifier and the application and action dates. Loan amounts and property values are rounded and ages are grouped into bands. Lenders report the full data to regulators, but the public file leaves those fields out.

Sources and further reading

What is your company’s data worth?

AI companies pay $20K to $5M for company records. Six questions, no files.