Calculating Self-Employed Income With an AI Agent: The 1084 Cash-Flow Analysis, the 4506-C Transcript, and the Reasonableness Call the Underwriter Owns

The Hardest Number in the File

Qualifying income for a W-2 wage earner is close to mechanical: the pay stub, the W-2, the verbal verification of employment, and a formula. Qualifying income for a self-employed borrower is the hardest number in the file. It requires reading two years of personal and business tax returns, identifying which figures are recurring income and which are one-time or paper losses, adding back the non-cash deductions, subtracting the income the borrower does not actually receive, and forming a view about whether the resulting number will continue. Two competent underwriters can read the same returns and land on different qualifying income, and that variance is the problem.

The variance is slow, because self-employed files sit in underwriting queues waiting for the analyst with the time and the expertise to work them. The variance is inconsistent, because the same borrower can qualify at one lender and not at another on the same returns. And the variance is a fair-lending exposure, because self-employment correlates with characteristics that ECOA and Regulation B protect, and an income-calculation process that produces inconsistent results across similar borrowers is a process that cannot easily prove the inconsistency is not discrimination.

An AI agent addresses the first two problems directly and the third structurally. It runs the cash-flow analysis the same way on every file, and it documents every step to the line on the return it came from. What it does not do is make the determination Fannie reserves for the underwriter. We build that agent, and the design question that matters is where its work stops.

The 1084 Is a Method, Not a Calculator

Fannie Mae's Cash Flow Analysis (Form 1084) is the standard worksheet for developing self-employed income from tax returns, and the underwriting factors and documentation requirements for a self-employed borrower live at Selling Guide B3-3.2. The 1084 walks through the borrower's returns by business structure, sole proprietorship on Schedule C, partnership on the 1065 and K-1, S corporation on the 1120-S and K-1, C corporation on the 1120, and pulls the income and the adjustments each structure requires.

The adjustments are the substance. Depreciation and depletion are non-cash deductions that reduce taxable income but not actual cash flow, so they are added back. Amortization and casualty losses are treated similarly. A one-time capital gain is recurring income only if the borrower has a documented history of it. Meals and entertainment are a real cash expense the tax code only partly deducts, so the analysis subtracts the non-deductible portion. Business use of a home, mileage depreciation inside a vehicle deduction, each has a specific treatment that determines whether it helps or hurts qualifying income.

The agent runs this method deterministically. It reads the returns, identifies the business structure, applies the 1084 line by line, adds back what the method adds back, subtracts what it subtracts, and produces the monthly qualifying income with every adjustment traced to the specific line on the specific return and the specific schedule it came from. The output is a worked calculation, traceable from the tax return to the qualifying income, that an underwriter or an examiner can follow without redoing it. That traceability is the difference between an income figure someone has to trust and one someone can verify.

The Transcript Is What Makes the Return True

A tax return in the file is a document the borrower provided. A tax return the agent has confirmed against the IRS is a document the file can rely on. The ability-to-repay rule at 12 CFR 1026.43(c)(4) requires verification of income using reliable third-party records, and it accepts several forms of them: the filed tax returns themselves, IRS transcripts, and other government, employer, or financial-institution records. The rule does not single out an IRS transcript as the required record. In practice most lenders and investors reconcile the returns in the file against an IRS transcript pulled through the borrower-authorized Form 4506-C process, because the transcript is the cheapest way to confirm the returns match what was actually filed, but that reconciliation is a lender and investor verification standard built on top of the rule, not a documentation type the ATR rule mandates.

The agent drives the 4506-C flow and reconciles the transcript to the returns in the file. A return whose figures match the transcript is a verified return. A return whose figures do not match the transcript is a finding, and the agent surfaces the specific line where the return and the transcript diverge. It does not decide what the divergence means, because an amended return, a transcription error, or a fabricated return all produce a mismatch and the difference among them is a judgment. The agent's job is that the mismatch is found on every self-employed file and put in front of the underwriter with the two figures side by side, so the verification the file relies on is actually performed rather than assumed.

The Line the Agent Does Not Cross

Fannie's B3-3.2 does not treat self-employed income as a formula. It requires the underwriter to determine that the income is stable, that the business has the capacity to continue generating it, and that the income used to qualify is reasonable in light of the business's trend. A borrower whose income rose from year one to year two presents a different question than a borrower whose income fell, and a declining trend is not disqualifying but it changes what income the underwriter can reasonably use. This is a judgment about the future, and Fannie assigns it to the underwriter.

The agent computes the two-year figures and the trend. It does not conclude the reasonableness determination. When the income is declining, the agent presents the decline, the magnitude, and the two years' worked calculations, and routes the file to the underwriter with the trend flagged. When the income depends on a single large item, the agent shows the dependency. When the business structure or the returns raise a question the 1084 method does not resolve, the agent surfaces the question rather than resolving it silently. The determination that the income is stable, continuing, and reasonable to use is the underwriter's, made on the computation the agent produced.

Building the agent to stop at that line is the whole design problem, because an agent that computes qualifying income is one small step from an agent that decides qualifying income, and the step is the one that matters. We keep the boundary explicit in the agent's output: the agent produces a calculation and a set of flags, and the qualifying-income figure the loan uses is the underwriter's decision recorded against the agent's calculation, not the agent's number adopted by default. The same boundary governs the ability-to-repay analysis the income figure feeds into. The agent's contribution is the accuracy and the consistency of the computation the underwriter is deciding on.

Why Consistency Is a Fair-Lending Control

The fair-lending argument for running self-employed income through an agent is not that the agent is unbiased. A model can encode bias, and an income-calculation agent that applied the 1084 differently to similar borrowers would be a fair-lending problem, not a solution. The argument is narrower and stronger: an agent that applies the same documented method to every file produces a record that lets the lender test for disparate treatment, because the calculation is explicit and comparable across borrowers.

When income calculation lives in individual underwriters' spreadsheets and judgment, two similar self-employed borrowers can receive different qualifying income and the lender has no efficient way to know, let alone to explain. When the calculation runs through a documented method the same way every time, the lender can compare the qualifying income the method produced across similar profiles, look for patterns that correlate with a protected characteristic, and either confirm the differences are explained by the returns or find a problem before an examiner does. The consistency is what makes the disparate-impact testing possible. An inconsistent process cannot be tested; a consistent one can. That does not remove the obligation to test, and the agent itself has to be inside the model-risk-management program and validated for exactly the bias it might introduce. But a documented, uniform calculation is the precondition for the monitoring, and a pile of individual spreadsheets is the absence of it.

The Failure Mode We Engineer Against

The pattern that produces the worst self-employed-income outcomes is the lender that treats income calculation as tribal knowledge held by a few senior underwriters. The files queue behind those underwriters, the qualifying income depends on which one worked the file, the add-backs are applied from memory rather than to a documented standard, and when a regulator asks how the lender calculates self-employed income the answer is a shrug toward experience. That process is slow, it is inconsistent, and it cannot defend itself.

The architecture we run makes the calculation a documented method the agent executes and the underwriter judges. Every self-employed file gets the same 1084 analysis, traced line by line to the returns, reconciled to the IRS transcript, with the trend and the dependencies flagged for the underwriter's reasonableness call. The senior underwriters stop spending their scarce time on the mechanical cash-flow arithmetic and spend it on the judgment the agent surfaced, the declining trend, the questionable continuation, the transcript mismatch that might be more than a typo. The files move faster because the computation is done when they arrive, the results are consistent because the method is uniform, and the process can be tested for fair lending because the calculation is explicit.

The Honest Read

Self-employed income is the hardest number in a mortgage file, and it is hard in three ways at once: slow, inconsistent, and exposed to fair-lending risk. An AI agent that runs the Form 1084 cash-flow analysis the same way on every file, traces every add-back to its line on the return, and reconciles the returns to the IRS transcript addresses the slowness and the inconsistency directly, and it makes the fair-lending testing possible by making the calculation explicit and comparable.

It does not make the reasonableness determination. Fannie assigns to the underwriter the judgment that the income is stable, continuing, and reasonable to use, and we build the agent to stop exactly there, presenting the computation and the flags and leaving the determination to the human who is accountable for it. At Sei, that is the boundary we design around: the agent owns the calculation, the underwriter owns the call, and the record shows which is which. The number gets faster and more consistent, and the judgment stays where the rule puts it.