What problem does it solve?
A mortgage application arrives as a stack of documents built for a person to read: pay stubs, W2s, tax returns, bank statements, profit and loss statements for the self employed, and letters explaining anything unusual. An underwriting team has to relabel and index every page, work out which income sources count under the applicable agency or investor guide (Fannie Mae, Freddie Mac, FHA, VA and USDA each have their own rules for calculating qualifying income), add them up by hand and run the whole calculation again every time a new document arrives to clear a condition.
The Mortgage Bankers Association put per loan production costs at USD 11,102 in the fourth quarter of 2025, across independent mortgage banks and mortgage subsidiaries of chartered banks. Income verification and document handling add to that cost and to how long a file takes: a loan sits waiting on a person to open a PDF, find the right page and type numbers into a spreadsheet, and it sits there again every time a condition brings in one more document.
The task is well suited to AI because the rules are explicit and public (published agency income guides) and the documents are structured enough to extract reliably, but it stays a credit decision: the qualifying income figure that comes out of this process can decide whether a borrower gets the loan they applied for, at the rate and term they were quoted.
- Per loan production costs decreased to USD 11,102 per loan in the fourth quarter of 2025, down from USD 11,109 per loan in the third quarter, according to the Mortgage Bankers Association's Quarterly Mortgage Bankers Performance Report.IMBs Report Production Profits in Fourth Quarter of 2025 (2026)
How does it work?
- Classify and index. Documents arriving by upload portal, email or fax are split into pages, classified by type (pay stub, W2, bank statement, tax return) and indexed against the loan file automatically.
- Extract and calculate. Wage, self employed, rental and other income types are extracted from the source documents and calculated against the applicable agency or investor income guide, with a confidence score on every figure.
- Check document authenticity. Bank statement deposits and document metadata are checked for internal consistency and compared against known alteration patterns, flagging suspicious files for a fraud specialist before they reach underwriting.
- Route by confidence. High confidence calculations flow into the loan origination system with the source page attached for every figure; anything below the threshold goes to an underwriter with the extracted data prefilled, not a blank form.
- Clear conditions without retyping. When a new document arrives to satisfy a stipulation, the same pipeline reruns, updates the calculation and shows the underwriter exactly what changed.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Mainstream
- Channels
- Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Handling time reduction | 78.5% vendor reported | 67% to 90% Not pooled: up to 85%, up to 67% | 4plus 2 up to | 4 vendor |
| Hours saved | Not pooled | 8500 hours to 150,000 hours | 2 | 1 organization, 1 vendor |
| Accuracy | Too few to pool | 90% | 1 | 1 vendor |
| Cycle time reduction | Too few to pool | 23% | 1 | 1 vendor |
Value drivers: Lower cost to serve, Speed and cycle time, Employee productivity, Risk and loss reduction.
Indicative value
A community bank originating 5,000 mortgage and home equity loans a year
USD 87,500 to USD 330,000
Underwriting and processing hours cost avoided per year
How this is calculated
Formula: loansPerYear * hoursSavedPerLoan * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Loans processed per year loansPerYear, loans per year | 5,000 | 5,000 | The reference lender. |
| Underwriting and income verification hours saved per loan hoursSavedPerLoan, hours per loan | 0.5 | 1.2 | Editorial assumption, replace with your own. Haventree Bank's up to three hours to under one hour reduction is an upper bound for its largest files and covers bank statement review only, not the full income verification workflow. HomeTrust Bank's reported saving (8,500 hours and USD 90,000 a year across its loan processing teams) cannot be converted into a per loan figure because its loan volume is not disclosed, so it does not set this range. |
| Fully loaded cost of underwriting and processing staff time costPerHour, USD per hour | 35 | 55 | Editorial assumption for US retail mortgage operations staff. Replace with your own fully loaded cost. |
What it leaves out: Gross labor time avoided only. It leaves out software licensing, integration work, and any revenue effect of a faster, more predictable closing timeline. The hours saved per loan input is an editorial assumption for this reference lender, not derived from the evidence on this page: HomeTrust Bank's own reported saving is USD 90,000 a year, at a loan volume Ocrolus does not disclose.
Who already uses it?
9 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
American Federal Mortgage Corporation
United States · Banking · 2026
American Federal Mortgage Corporation, a US mortgage lender, had underwriters spending close to an hour manually reviewing bank statements for complex, self employed borrower income. It adopted Ocrolus to classify documents at intake, calculate income and surface early insights to processors through Ocrolus Inspect. Ocrolus reports complex self employed income reviews fell 90 percent, from about an hour to five to ten minutes, and that underwriters now clear files in about nine hours instead of twelve.
- Handling time reduction: 90%, per complex self employed income review
"Complex self-employed income reviews cut 90% — from ~1 hour to 5–10 minutes"
Claimed by: vendor - Handling time reduction: up to 50%, initial processor review, per typical file
"Initial processor review on a typical file reduced by up to 50% with Inspect"
Claimed by: vendor
Haventree Bank
Canada · Banking · 2026
Haventree Bank, a Canadian alternative residential mortgage lender, had its underwriting team manually transferring data from bank statements into an Excel income calculator, taking up to three hours per file. It adopted Ocrolus to automate bank statement income analysis and document authenticity checks. Ocrolus reports the bank cut bank statement review time by 67% and moved from an inconsistent 9 to 12 months of statements reviewed per file to a consistent full 12 months, without adding headcount as loan volumes grew.
- Handling time reduction: 67%
"Since adopting Ocrolus, Haventree Bank has cut bank statement review time by 67%."
Claimed by: vendor
Key Mortgage Services, Inc.
United States · Banking · 2026
Key Mortgage Services, a US mortgage lender, had its setup team manually classify and index every document in a loan file and manually cross check verification of employment data against the 1003 application and the automated underwriting system, taking 20 to 30 minutes per file. It adopted Ocrolus Classify to automate document classification, indexing and those cross checks. Ocrolus reports classification time fell by up to 67 percent, letting the setup team handle more loan volume without adding headcount.
- Handling time reduction: up to 67%, per file, document classification and indexing
"Up to 67% reduction in classification time: Indexing dropped from 20-30 minutes to 5-10 minutes per file"
Claimed by: vendor
MCM Holdings, Inc.
United States · Banking · 2026
MCM Holdings, a Miami mortgage lender operating since 1998 and licensed in more than 30 states, saw manual underwriting workflows limit speed as it expanded its non-QM and bank statement lending programs. It integrated Ocrolus into its Encompass loan origination system to automate bank statement analysis, income verification and document classification. Ocrolus reports the lender cut bank statement review time by over 90%, from four or five hours to about 10 minutes, while reducing average time to close by 23% and supporting 25% year over year loan volume growth.
- Handling time reduction: at least 90%, per bank statement review
"Over 90% reduction in bank statement review time, from five hours to about 10 minutes"
Claimed by: vendor - Cycle time reduction: 23%, average time to close
"Underwriting tasks that once took hours now take minutes, helping reduce average time to close by 23% and supporting 25% year-over-year loan volume growth."
Claimed by: vendor
Nations Direct Mortgage
United States · Banking · 2026
Nations Direct Mortgage, a US wholesale mortgage lender, indexed and reviewed loan files by hand, including income analysis that took underwriters 30 to 45 minutes per file. It adopted Ocrolus to automate document classification and income calculation inside its Encompass loan origination system. Ocrolus reports income review time fell 67 percent to under ten minutes, with underwriters accepting the Ocrolus recommended income calculation as is on 9 in 10 files, and file setup time fell 83 percent.
- Handling time reduction: 67%, per file income analysis
"67% faster income review: Underwriters reduced time on income analysis from 30–45 minutes to under 10 minutes"
Claimed by: vendor - Accuracy: 90%, files reviewed
"90% income acceptance rate: Underwriters accepted Ocrolus-recommended income as-is on 9 in 10 files"
Claimed by: vendor
Volo Loans
United States · Banking · 2026
Volo Loans, a mortgage lender that finances US citizens and Canadians buying homes internationally, evaluates W-2 income, bank statements and complex business tax returns across multiple entities, and ran a highly manual underwriting process before automation. It adopted Ocrolus to classify documents and automate income analysis while keeping underwriter review and validation in place. Ocrolus reports the change cut income and document review time by up to 85%, saving an average of five to six hours per loan file.
- Handling time reduction: up to 85%, per loan file
"Income and document review time dropped by up to 85%, saving an average of five to six hours per loan file and enabling the Volo Loans team to stay lean while improving the borrower experience."
Claimed by: vendor
Westpac
Australia · Banking · 2026
Westpac built five AI agents on Amazon Bedrock AgentCore that work alongside bankers in its home lending and credit card application processes. Two agents extract data from a customer's payslip and verify it against lending policy and the required calculations, while a third, an "account scrutiny agent", classifies transactions and spending from manual bank statements. The agents hand their output to a banker to review before the application continues, and the bank says they are running in production, not as prototypes.
- Hours saved: at least 150,000 hours
"McMullan said that the assistance of the AI agents in lending processes produced "more than 150,000 hours" of time being saved, which could be "reinvested" by teams to serve customers better."
Claimed by: organization
Deephaven Mortgage
United States · Banking · 2024
Deephaven Mortgage, a lender working through independent mortgage brokers, serves borrowers locked out of the traditional GSE market by non-W-2 income or short credit histories, and uses bank statement analysis to assess their actual cash flow. It integrated Ocrolus into its loan origination workflow to capture, verify and analyze bank statement data automatically instead of calculating real income by hand. Ocrolus reports that Deephaven's own internal studies on underwriter productivity found more than two hours saved per mortgage application.
- Time saved per task: at least 120 minutes, per mortgage application
"According to Deephaven's internal studies on underwriter productivity, Ocrolus helped increase efficiencies and helped save more than 2 hours per mortgage application for their underwriters."
Claimed by: vendor
HomeTrust Bank
United States · Banking · 2024
HomeTrust Bank, a North Carolina community bank, used to spend over four hours a week per team member manually relabeling mortgage documents and a similar amount of time verifying income by hand. It adopted Ocrolus, integrated with its Encompass loan origination system, to classify documents and calculate wage, self employed, rental and other income automatically across loan origination, processing and underwriting. Ocrolus reports the bank estimates annual savings of 8,500 hours and USD 90,000 through the resulting efficiencies.
- Hours saved: about 8500 hours, per year
"HomeTrust Bank estimates annual savings of 8,500 hours across loan processing teams and $90,000 through efficiencies in document processing."
Claimed by: vendor - Cost savings: about USD 90,000, per year
"HomeTrust Bank estimates annual savings of 8,500 hours across loan processing teams and $90,000 through efficiencies in document processing."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Current agency and investor income calculation rules (Fannie Mae, Freddie Mac, FHA, VA, USDA)
- A labeled set of historical loan files to test classification and income calculation accuracy
- A document authenticity or fraud watchlist feed for altered statement detection
Systems to integrate
- Loan origination system (for example Encompass by ICE Mortgage Technology)
- Document upload portal and email or fax intake
- Income and employment verification services
- Underwriter task queue for flagged and low confidence files
Complexity: Medium
Document classification and extraction are largely off the shelf; the real work is mapping each investor's and agency's own income calculation rules, integrating with the loan origination system, and building an escalation path that gives the underwriter prefilled data rather than a blank recalculation.
- 1
Start with one income type and one loan program
Prove accuracy on wage earner W2 income for conventional loans before expanding to self employed, rental and government loan income, where the rules are more complex.
- 2
Set a confidence threshold and route below it to a person
Every calculation carries a confidence score. Anything below the threshold goes to an underwriter with the extracted figures and source pages attached, not a blank recalculation.
- 3
Keep every figure traceable to its source page
Underwriters must be able to click from a calculated income figure back to the exact page and line of the document it came from, so a challenge takes seconds, not a document search.
- 4
Route flagged files into the underwriter's existing queue
Low confidence and suspected fraud cases land in the loan origination system's own task list, so underwriters do not have to work a separate tool for exceptions.
- 5
Reverify accuracy against agency guide updates
Agency and investor income guides change; schedule a recurring check of the calculation rules against the current published guides, not only against last year's test set.
Guardrails
- A person, not the model, makes and signs the final income and eligibility determination
- Every calculated figure links to the exact source document page it came from
- Confidence thresholds route uncertain extractions to a person instead of a best guess
- Document authenticity checks stay on and are monitored even under application volume pressure
KPIs to instrument
- Income calculation accuracy against a manually reviewed sample, by income type
- Time from document receipt to a cleared condition
- Share of files requiring manual income recalculation
- Document authenticity flags raised versus confirmed fraud
Human in the loop
An underwriter reviews and signs off on every loan file; the AI reduces retyping and first pass assembly, it does not remove the underwriter's judgment. Low confidence calculations, unusual income types and any authenticity flag go to a person before the file moves forward, and a sample of high confidence calculations is checked periodically against a manual recalculation.
Common failure modes
- Income rules go stale
- Agency and investor income guidelines change periodically. An unmaintained rules engine keeps calculating against an old version and understates or overstates qualifying income. Review the rules against the agencies' published guides on a fixed schedule.
- Confidence miscalibrated on unusual income
- A model that is confidently wrong on seasonal, gig or multiple job income routes bad numbers straight through. Measure accuracy per income type, not only in aggregate.
- Fraud checks disabled under volume pressure
- Document authenticity checks get treated as an optional step when application volume spikes, letting altered bank statements and pay stubs through. Keep the check mandatory and monitor its trigger rate for unexplained drops.
What are the risks and rules?
EU AI Act
Depends on design
Classification and indexing of documents for a person to review may fall under the Article 6(3) derogation for narrow procedural tasks, if the provider documents that assessment. That derogation does not apply once the system profiles a natural person: calculating a named borrower's qualifying income from their pay stubs and bank statements evaluates that person's economic situation, which is profiling under GDPR Article 4(4).
A system intended to calculate qualifying income for the credit decision is high risk under Annex III point 5(b), evaluating the creditworthiness of natural persons, whether or not a person reviews its output. Human oversight of that output is a separate obligation under Article 14, not a way to take the system out of the high risk category.
Rules that apply
Guidance
- Guidelines on loan origination and monitoring (EBA/GL/2020/06) (European Banking Authority, Europe). Sets governance and creditworthiness assessment standards for loan origination, including automated elements of the process.
Controls to put in place
- Human underwriter sign off on every loan file, with the AI's role limited to assembly and calculation
- Full audit trail from every calculated figure back to its source document page
- Periodic sampling of automated calculations against an independent manual recalculation
- Mandatory document authenticity checks with monitored trigger rates
Frequently asked questions
- Does AI income verification replace the underwriter?
No. It removes the manual relabeling, indexing and retyping, and gives the underwriter a calculation with every figure linked to its source page. The underwriter still reviews and signs the file, and any low confidence or unusual case is routed to a person before the file moves forward.
- How much time can this save on a mortgage file?
Ocrolus reports that Haventree Bank cut bank statement review time by 67%, taking a review that took up to three hours down to under one hour, and that HomeTrust Bank estimated annual savings of 8,500 hours across its loan processing teams after automating income calculation and document handling.
- Which income types are hardest to automate?
Self employed, rental and other variable income sources are harder than a single wage earner pay stub, because they depend on multiple documents and more agency specific rules. Start with the simplest income type and expand once accuracy is proven.
- Is this a high risk system under the EU AI Act?
It depends on what the system is intended to do. Pure document classification and extraction for a person to review can fall under the Article 6(3) derogation for narrow procedural tasks. A system intended to calculate a borrower's qualifying income for the credit decision profiles a natural person's economic situation and is high risk under Annex III point 5(b), evaluating creditworthiness, whether or not a person reviews the output; Article 14 then requires human oversight of that high risk system.
How to cite this page
Blits.ai AI Use Case Library, "AI income and document verification for mortgage underwriting", last verified 29 September 2026, https://www.blits.ai/ai-use-cases/mortgage-income-and-document-verification. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 1 October 2026: Added deployments: Westpac.
- 29 September 2026: First published