sei
Mortgage

The Large Deposit Question: Verifying Assets With an AI Agent Under Fannie Mae B3-4.2-02, and Where Asset Sourcing Meets the Bank Secrecy Act

7 min read
Ramkumar Venkataraman
Share

Adding Up the Balances Is the Easy Part

Asset verification looks like the simplest thing an underwriter does. Read the bank statements, confirm the borrower has enough for the down payment, closing costs, and reserves, and move on. The balances are the easy part. The work that matters is the sourcing: where did the money come from, is it the borrower's own funds, and does anything in the account movement suggest funds that were borrowed, gifted without documentation, or deposited from a source the file cannot explain. A borrower with enough money in the account can still fail asset verification, because having the funds and being able to source them are different questions, and the second one is where loans get conditioned and where a regulated lender has an obligation that goes past eligibility.

We build the agent that runs across mortgage origination on lender platforms, and asset verification is a clean example of a task that reads like arithmetic and behaves like an investigation. The agent reads the statements, identifies the deposits that require sourcing, and applies the guideline. It also sits at the point where an underwriting question and a Bank Secrecy Act question can be the same deposit seen from two angles, and knowing which one you are looking at is part of doing the job.

The 50 Percent Rule Defines the Work

The rule that generates most asset conditions is the large-deposit standard. Under Fannie Mae's Selling Guide B3-4.2-02 on depository accounts, when bank statements are used to verify assets, the lender must evaluate large deposits, and a large deposit is defined as a single deposit that exceeds 50 percent of the total monthly qualifying income for the loan. Crossing that threshold means the deposit gets evaluated, not that it automatically has to be documented, and the guideline is narrower than a blanket sourcing rule in two ways worth building to.

The evaluation only applies where the funds are needed. On a refinance, B3-4.2-02 does not require the borrower to source large deposits at all, because the transaction is not consuming the borrower's own funds to close in the way a purchase does. On a purchase, if a large deposit is unsourced, the lender does not have to obtain documentation. It can instead reduce the funds it counts by the amount of the unsourced deposit and proceed if the remaining verified funds are enough to cover the down payment, closing costs, and reserves. Sourcing is required only when the file actually needs the deposited funds to close and the deposit is not otherwise explained on the statement.

The guideline draws a useful line on what needs explaining. If the source of the deposit is printed on the statement and is a source that plainly is not a loan, a direct deposit from the employer, a Social Security payment, a tax refund, or a transfer from another verified account of the borrower, the lender does not need further documentation. If the source is not identified, or if it is identified but still leaves open whether the funds were borrowed, the deposit has to be documented. The reason the rule exists is that an undocumented large deposit could be a loan the borrower took to cover the down payment, which would change the debt picture the underwrite is built on, so the funds have to be shown to be the borrower's own.

The agent applies this by computing the large-deposit threshold from the qualifying income on the specific loan, not from a fixed dollar figure, because 50 percent of qualifying income is a different number on every file. It reads each statement, identifies every deposit over that loan's threshold, and checks whether the statement itself sources it. The deposits the statement explains, it clears. For a deposit the statement does not explain on a purchase, the agent first tests whether the funds are even needed, recomputing the borrower's available funds with the unsourced deposit removed, because if the remaining verified funds still cover the down payment, closing costs, and reserves, the deposit can be excluded and no condition is warranted. It flags for documentation only the deposits the file actually relies on, with the deposit, the date, the amount, and the reason it needs sourcing laid out, so the condition the underwriter sends the borrower is specific and necessary rather than a blanket request for an explanation of the whole account. On a refinance it does not raise large-deposit conditions at all, which keeps the agent from imposing documentation the guideline does not require.

Reading the Statement Is a Document-Intelligence Problem

Identifying large deposits sounds mechanical and is not, because bank statements are not a standard format. Every institution lays them out differently, deposits and transfers and cleared checks are described in inconsistent ways, and the same underlying event, a payroll deposit, can appear as a dozen different strings across a dozen banks. An agent that reads statements has to parse formats it was not explicitly programmed for, distinguish a deposit from a returned item or a reversal, recognize a transfer from another of the borrower's accounts as a transfer rather than a new deposit, and do it accurately enough that it is not flagging routine payroll as a large deposit or missing a real one.

This is the document-intelligence core of the task, and it is where the income calculation work and the asset work share the same engine: extract structured facts from an unstructured financial document, then apply a guideline to the facts. The agent extracts every transaction, classifies each one, and reconciles transfers between the borrower's own accounts so that money moving from savings to checking is not double-counted as both a withdrawal and a new deposit. Getting the classification right is what keeps the flags meaningful, because an agent that flags every payroll deposit as needing sourcing is an agent the underwriter learns to ignore, and an agent the underwriter ignores is worse than no agent at all.

Gifts, Reserves, and Funds That Have to Be Seasoned

Beyond large deposits, asset verification carries a set of sub-rules the agent has to hold. Gift funds are permitted on many loan types but have to be documented with a gift letter and, depending on the program, evidence of the transfer and the donor's ability to give, and a gift that shows up as an undocumented deposit is a large deposit until it is papered correctly. Reserves, the funds the borrower has left after closing, have to be verified as actually available and not, for example, retirement funds counted at full value when the program only allows a percentage. Borrowed funds secured against an asset are treated differently from the borrower's own funds. Each of these is a rule with a documentary requirement, and each is a place a file gets conditioned when the deposit and the documentation do not line up.

The agent carries these rules and applies the one that fits what it sees. A deposit that matches a gift letter in the file, it ties to the gift and clears. A deposit that looks like a gift but has no letter, it flags for the letter. Retirement assets used for reserves, it applies the program's allowable percentage rather than the full balance. The value here is not that any one of these is hard. It is that a file has many accounts and many deposits and several applicable rules, and applying all of them correctly across the whole file, every time, is exactly the consistency a manual review loses on the twentieth file of the day.

Where the Underwriting Question Becomes an AML Question

Here is the part that makes asset sourcing more than an eligibility exercise. A mortgage lender that is subject to the Bank Secrecy Act, and FinCEN's 2012 final rule subjects non-bank residential mortgage lenders and originators to AML program and suspicious activity reporting requirements under 31 CFR Part 1029, has an obligation that does not stop at whether the funds are eligible. An unexplained large deposit is an underwriting condition. A deposit with characteristics that suggest the funds are the proceeds of illegal activity, or that fit a structuring pattern, or that come from a source the borrower cannot or will not explain, is a different thing, and the lender's obligation there is to its AML program, not just to the investor's eligibility rule.

Most large deposits are not this. They are a bonus, a car sale, a transfer the borrower forgot to mention, and they resolve with a document. But the same deposit that raises an eligibility question can, in a small number of cases, raise a suspicious-activity question, and an underwriting process that only ever asks the eligibility question is a process that will not notice when it is looking at the other one. The agent's role here is narrow and important: it surfaces the characteristics that distinguish a routine sourcing question from one that should reach the lender's BSA function, and it routes those cases to the people who own that determination. The agent does not file a SAR, does not decide that activity is suspicious, and does not make the AML judgment, because that is a determination the lender's compliance function owns, with its own process and its own narrative and reporting discipline. What the agent does is make sure the deposit that should have been seen by that function is not buried in an underwriting queue where the only question anyone asked was whether the borrower had enough money.

The Honest Read

Asset verification is a sourcing investigation dressed as arithmetic. The 50 percent large-deposit rule in Fannie Mae's Selling Guide defines most of the work, and applying it means computing the threshold from each loan's qualifying income, reading bank statements that come in every format there is, classifying transactions accurately enough that the flags mean something, and holding the gift, reserve, and borrowed-funds rules alongside the large-deposit one. That is a document-intelligence problem with a guideline on top, and it is the kind of consistent, every-file-every-time work an agent does better than a queue.

The part that makes it more than eligibility is the point where an unexplained deposit stops being a condition and becomes a source-of-funds question a regulated lender cannot leave inside an underwriting queue. At Sei, we build the asset agent to source deposits against the guideline and to recognize the small set of cases that belong with the lender's BSA function, and to route rather than decide, because the SAR determination is compliance's to make. The balances are the easy part. The agent earns its place on the sourcing, and on knowing which question a given deposit is really asking.

Ramkumar Venkataraman

Ramkumar Venkataraman

CTO & Co-Founder

Related Articles

Mortgage

The Three Tolerance Buckets: Balancing the Closing Disclosure Against the Loan Estimate With an AI Agent

Between the Loan Estimate and the Closing Disclosure sits the fee comparison that decides whether a lender owes the borrower a refund. Zero tolerance, ten percent aggregate, and no tolerance are three different rules on three different sets of fees, and a changed-circumstance re-disclosure can move a fee from one bucket to another. Where an AI agent tracks every fee from LE to CD, catches the tolerance breach before consummation, and computes the cure the rule requires.

Aug 26, 20266 min read
Read more
Mortgage

Mortgage Fraud Detection at Origination: What an AI Agent Can Flag, What It Cannot Decide, and Where the SAR Obligation Starts

Most origination fraud is not a forged document, it is a set of facts that are each plausible and collectively wrong: an owner-occupancy claim that does not fit the file, an employer that only exists on paper, a gift that is really a loan. Where an AI agent reads the whole file for the pattern instead of each document in isolation, what FinCEN's mortgage AML rule and the Red Flags Rule actually require, and the line between a fraud flag and a fraud determination.

Aug 25, 20266 min read
Read more
Mortgage

Verification of Employment With an AI Agent: The Written VOE, the Day-of-Closing Verbal, and Why The Work Number Is a Credit Report

VOE reads like a phone call and is actually three separate controls with three separate failure modes. The written verification, the verbal within ten business days of the note, and the reverification when something changes. Where an AI agent runs each one, what Fannie Mae B3-3.1-07 actually requires, and why treating a database VOE as free of FCRA is the mistake that shows up in a dispute.

Aug 24, 20267 min read
Read more

BOOK A DEMO

Embed Sei AI in your workflows
Tell us about your operations. We'll show you how Sei handles borrower calls, processes loan documents, and monitors compliance for mortgage lenders and banks.
  • Deploy in weeks, not months
  • Trained on FDCPA, TCPA, TILA, UDAAP, and RESPA
  • SOC 2 Type II and PCI DSS L1 certified
  • Integrates with your LOS, CRM, and telephony

Please provide your full name so we know how to address you.

Tell us which company you represent so we can personalise our response.

Use your work email so we can connect you with the right specialist.

Choose the topics you’d like us to cover during the demo.

Complete the verification to submit the form.

sei

AI operations platform for mortgage lenders, servicers, and banks. Handle borrower calls, process loan documents, and monitor compliance.

© 2026 Sei Software Technologies Inc. All rights reserved.