Underwriting the Bank-Statement Loan With AI: Deposit Analysis That Is Auditable, and the Ability-to-Repay Rule Non-QM Never Escapes

The Income Number Is the Whole Loan

On a bank-statement loan the income calculation is not one input among many. It is the underwrite. A self-employed borrower who does not document income through tax returns is qualified instead on the deposits into their personal or business accounts, typically over twelve or twenty-four months, and every other number on the file, the debt-to-income ratio, the loan amount, the approval itself, moves with the qualifying income the analyst derives from those statements. Get the income right and the file is sound. Get it wrong by inflating it and you have originated a loan the borrower cannot carry, which is the exact harm the ability-to-repay rule exists to prevent, on the product where that rule is easiest to convince yourself you have satisfied.

That is what makes this a good place for an AI agent and a dangerous place to be careless with one. The work is arithmetic over a large document set, which is what an agent is built for. The consequence of the arithmetic being confidently wrong is a credit decision, which is why the calculation has to be auditable the way self-employed income calculation is, showing every deposit it counted and every one it excluded, not a single figure a model asserts.

Non-QM Does Not Mean No-ATR

The most expensive misconception about bank-statement lending is that because the loan is non-qualified, the ability-to-repay obligation is somewhere else. It is not. Regulation Z §1026.43(c) requires a creditor to make a reasonable and good-faith determination, at or before consummation, that the consumer has a reasonable ability to repay a covered closed-end residential mortgage, and that requirement applies to the loan whether or not it is a qualified mortgage. What a QM gives a lender is a safe harbor or a rebuttable presumption of compliance. What non-QM gives up is that presumption, not the underlying duty. So a bank-statement loan carries the full ATR obligation and less legal cover, which is the opposite of the informal sense that non-QM is the loosely documented product.

The rule names the eight factors the determination has to consider: current or reasonably expected income or assets, current employment status, the monthly payment on the covered transaction, the monthly payment on any simultaneous loan, the monthly payment for mortgage-related obligations, current debt obligations including alimony and child support, the monthly debt-to-income ratio or residual income, and credit history. An agent that computes a qualifying income and hands it off has touched the first and most error-prone of those factors, and the file still has to show that all eight were considered by someone who is allowed to make the call. The eight factors are the frame the ATR and QM analysis sits inside, and the bank-statement product does not get to skip any of them.

What the Agent Computes From the Deposits

The deposit analysis is where the agent does the volume of the work. It reads the full statement period, classifies each deposit, and separates the deposits that count as business income from the ones that do not, because the whole method depends on that separation being right. It excludes what the investor guideline excludes, applies the expense factor or reconciles the deposits against a CPA-prepared profit-and-loss statement where the program uses one, annualizes, and produces a qualifying monthly income with every included and excluded deposit itemized. The output is not a number. It is a worksheet a human underwriter and a later reviewer can both follow line by line.

An illustrative agreement figure we hold before this task goes live is in the high nineties against an independent re-calculation, measured on the qualifying income and set for this product specifically, because the bank-statement calculation fails differently than a tax-return calculation and does not deserve the same threshold as an easier task. The bar is per calculation type, not a single number stretched across every kind of income the agent touches, for the same reason a checklist task and an income task do not share a go-live gate.

The Deposit That Was Not Income

The failure mode that shaped this agent came out of shadow mode, and it is the one every deposit-based method is exposed to. A self-employed borrower moves money between their own accounts, from a business operating account to a personal account, from savings to checking, and each of those transfers lands as a deposit. Early on, the agent counted them, because on the face of a single statement a transfer in looks like any other credit. On a borrower with several accounts and regular internal movement, that inflated the qualifying income materially, which on a bank-statement loan flows straight into a larger loan the borrower cannot actually support. Shadow mode surfaced it because the agent's income was compared against a careful human calculation on live files, and the gap concentrated exactly on the borrowers with multiple accounts.

The decision from that is now part of the method. The agent detects likely inter-account transfers, deposits it can tie to a debit of the same amount from another of the borrower's accounts in the window, and excludes them pending sourcing rather than counting them, and it flags large or irregular deposits for the sourcing question instead of absorbing them into income. The point of measuring against human review was never to prove the agent was good on average. It was to find the cohort where it was wrong, and inflated income on multi-account borrowers was that cohort.

The Reasonableness Call the Underwriter Owns

The agent computes. It does not decide the loan. The reasonableness determination, whether the qualifying income the deposits support is consistent with the business, the occupation, and the rest of the file, is a judgment the underwriter owns, and so is the ability-to-repay sign-off that the eight factors were considered and met. The agent makes that judgment faster and better-supported by handing the underwriter a complete, itemized calculation with the exclusions already surfaced, rather than a stack of statements and a blank worksheet. It does not make the judgment for them, because a model's qualifying income is an input to an ATR determination, not the determination itself.

That places the calculation inside model governance, because an output that feeds a credit decision is a model under SR 11-7 and the discipline around it: documented intended use, validation before reliance, and monitoring on the outcomes that matter, which here is the rate at which underwriters adjust the agent's income and the defects a later review finds on loans it touched. At Sei we build the bank-statement agent to show its work, exclude what it cannot source, and stop at the reasonableness call, because on this product the number is the loan, and the number has to be one a person can defend, not just one a model produced.