AI loan file review: agents check the mortgage file, people handle the exceptions

AI loan file review uses software agents to check every document in a mortgage file. They compare each document with the loan application, with the other documents and with your rules. Files that pass every check go on to underwriting. Your team sees only the checks that failed, with the figures and source pages attached.

The agents get the file ready for a decision. The underwriter still makes it.

What the agents do with a loan file

An AI agent, in this sense, is software that runs a series of steps without a person starting each one. In a loan file, those are the steps processors do by hand today.

1. CollectThe file comes in by email, upload or API. If a document is missing, an agent emails the sender to ask for it.
2. Classify and splitA long upload is split into the pay stubs, W-2s, tax returns and statements inside it.
3. ExtractEach field is read and given a confidence score, which shows how sure the software is. Transactions that run across pages are joined into one table.
4. VerifyEach document is checked against the application, the other documents and your rules.
5. RouteFields the software isn't sure of go to a person. Failed checks go to whoever owns them, and clean data goes to your loan origination system (LOS).

Most of the time saved comes from the Route step. If the underwriter still reads every page of every file, the agents have only changed who reads the file first.

What an exception looks like

An exception is one failed check, shown with the evidence behind it. It tells the underwriter which two figures disagree, where each one came from and which rule failed. A flag that only says "check income" sends someone back through the whole file.

The figure shows how Docsumo reads one couple's tax forms. One borrower draws a salary from an S corporation and owns 25% of it. That puts a W-2, a K-1 and the company's 1120-S in the file. The other borrower is self-employed and was paid by two businesses on 1099-NEC forms. Each form lists what is read from it and what it's checked against. Almost every check passes. The two 1099-NEC forms add up to $46,800, but Schedule C reports $41,300 in gross receipts. That check fails by $5,500.

The flag doesn't say which figure is right. It shows the underwriter where the gap is and how big it is. The underwriter decides whether it matters and what to ask for. In Docsumo, checks between documents like this one are on the Enterprise plan.

Exceptions in a mortgage file come in four kinds, and each kind has a clear owner.

The four-month limit comes from Fannie Mae's Selling Guide. Credit documents, including income and asset documents, can be no more than four months old on the note date. Federal tax returns follow a separate rule in the same section. The deposit rule is Fannie Mae's large-deposit test, which bank statement verification for mortgages covers in full. Most other documents are checked against the application itself, the Form 1003.

Page-by-page review vs exception-based review

In a page-by-page review, a person reads every page of every file. In an exception-based review, software reads every page, and people read what failed. Here is how the work changes for processors and underwriters.

Rules to settle before a pilot

Agents follow the rules you give them. A vague rule raises vague flags. Agree on these with your underwriting and QC leads before the first file goes through.

• A tolerance for each checkHow far apart two figures can be before the gap counts as an exception. A tight tolerance sends more files to people.
• An age limit for each documentCount it to the note date. Fannie Mae's limit for income and asset documents is four months, and tax returns follow their own rule.
• Name matchingDecide whether "Front Range Dental Group PC" and "Front Range Dental" count as the same employer.
• Hold or noteDecide which failures stop the file and which only add a note for the underwriter.
• An ownerName who clears each kind of exception, whether it's a processor, an underwriter or QC.
• The evidenceSay what every flag must show, such as both values, their pages and the rule.
• The recordLog who cleared each flag and when, so QC can trace it later.

Our take. An agent is worth having when the underwriter can act on its flag without reopening the file. So every exception should arrive with the figures that disagree, the pages they came from and the rule that failed. A flag without that evidence just sends someone back to read the file.

Prefunding QC with the same checks

Fannie Mae requires lenders to review a sample of loans before closing. It calls this prefunding quality control. The lender must have a documented monthly process for picking loans, with picks from each production channel. The reviews must happen early enough for mistakes to be fixed before the loan closes. A full-file review checks how income was calculated and the documents behind it. It also covers employment, assets, the appraisal and occupancy, among other items. Occupancy means how the borrower will use the home.

Two Fannie Mae rules limit how agents can help QC. The reviewers must have no part in originating, processing or underwriting the loan they review. The QC process must also run apart from production, where that's practical. The agents' checks are part of processing, so they can't count as the independent QC review. What they can do is show the QC reviewer where to look first. The reviewer sees which checks passed and which failed, with the pages behind each one. The reviewer still assesses every item Fannie Mae lists.

Defect data shows where files go wrong. ACES Quality Management sells QC software. It publishes defect data from the post-closing reviews that lenders run in that software. In the first quarter of 2026, the critical defect rate in its data was 1.71%, up from 1.38% in the quarter before. Legal, regulatory and compliance findings made up 26.02% of critical defects. Income and employment made up 20.07%, the second-largest share.

QC reviews only the loans it samples. Agents can check the pay stubs, W-2s and tax returns behind the income on every file.

What stays with the underwriter

The agents check documents. The underwriter decides the loan.

Income shows this most clearly. For a self-employed borrower, Fannie Mae asks the lender for a written cash flow analysis. Its purpose is to find the "stable and continuous" income the borrower will have. The lender can use Form 1084, another cash flow analysis, or an automated tool such as Fannie Mae's Income Calculator. Docsumo extracts the tax return data that goes into that analysis. It doesn't produce Form 1084 or Freddie Mac Form 91 calculations.

Whether a bonus will continue and whether a job change matters are judgments for the underwriter too. So is the credit decision. An agent can put the evidence in front of the underwriter, but it shouldn't make the decision.

Where Docsumo fits

Docsumo is document AI for teams that decide on files, such as loan files. Its agents collect, classify, extract and cross-check every document in a loan file, then post the result to your system. Your team only sees the exceptions. In the case review, the "Show only failed" switch hides every check that passed. A reviewer can then ask Copilot what failed and why.

Classification and splitting are on the Business plan. AI Split takes a plain-English definition of each document type. A new document type works from the first document, with no ramp-up period. One upload can hold up to 500 pages. Checks between documents run as cross-document validation, and each loan's documents stay together in case management. Both are on the Enterprise plan, as listed on the pricing page.

Each field gets a confidence score, and you set the threshold per field. Fields below it go to a person, as in any human-in-the-loop setup. A reviewer who clicks a field sees its source line on the page, and each correction improves the model. You can add rules of your own, such as your large-deposit test, as workflow steps. A step can run code or call an LLM (a large language model). Docsumo also checks the running balances on bank statements and reads pay stubs, W-2s and tax returns.

Checked data goes to your LOS through the API and webhooks. Docsumo supplies the software, and your own staff review the exceptions. It doesn't make the credit decision, and it runs in the cloud only. It's SOC 2 Type 2 compliant and ISO 27001 certified, with HIPAA and GDPR compliance, single sign-on and role-based access (security). Audit logging is on the Business plan. For what a log should record, see audit trails.

• 95%+95%+of documents processed straight through, with no manual review
• 99%99%field-level accuracy on 250+ document types
• <5 min<5 minper document, down from 2+ hours

For other tools that read mortgage files, see the best mortgage document automation software. For the whole loan, from application to closing, see mortgage process automation.

Start from the defects your QC team already finds

Your QC reports already list the findings that come up most. Take the five most common from last quarter and write each one as a rule, with a tolerance and an owner. Then run a few closed loans through the agents and compare their flags with what QC found. If the agents missed a finding, fix that rule before you add more files.

Book a demo and bring two closed loan files with the borrower details masked, or sample files, plus their QC findings. You can also start a free trial.

Frequently asked questions