AI Income Calculation for Self-Employed Borrowers: The Part of Underwriting Where the Math Has to Be Auditable

The One Number That Decides the Loan

For a wage earner, qualifying income is close to a lookup. Pull the pay stub, confirm the W-2, annualize, done. For a self-employed borrower, qualifying income is a construction. It is built from two years of personal and business tax returns, adjusted for non-cash expenses, tested for trend, and haircut for anything that will not continue. Two competent underwriters can land on different numbers from the same returns, and both can defend their answer.

That is the number an AI income engine has to produce, and it is the number an examiner, a model validator, and an investor QC reviewer will all pull apart. This is not a document-extraction problem with a calculation bolted on. It is a calculation problem where the documents are the inputs, and the difference matters for how you build it.

What the Calculation Actually Is

Fannie Mae's Selling Guide B3-3.2 governs self-employment income, and Fannie publishes the arithmetic as Form 1084, the Cash Flow Analysis worksheet. Freddie Mac's parallel guidance sits in Single-Family Seller/Servicer Guide Section 5304. These are not vibes. They are line-referenced worksheets that pull specific figures from the 1040, Schedule C, Schedule E, the 1120-S, the 1065, the K-1, and the depreciation schedules, then add back the non-cash deductions and subtract the income that does not recur.

The 1084 for a sole proprietor starts from Schedule C net profit, then adds back depletion and depreciation, subtracts the meals exclusion and any non-recurring other income, and adjusts for business use of home and mileage depreciation. For an S-corporation borrower it runs through the 1120-S, the K-1, and a business liquidity test before any income flows to the borrower. Each line has a source, and each source is a specific box on a specific form for a specific tax year.

An AI engine that produces a qualifying income figure without producing the 1084 line by line is producing a guess with a confident tone. The line-by-line worksheet is the deliverable, not a byproduct.

Why We Build the Worksheet, Not the Answer

The failure mode we see in income automation is a system that reads the returns, emits a monthly income number, and stops. The underwriter cannot see how it got there, so the underwriter either trusts it blindly or recalculates by hand, and the second one erases the time savings. Neither is acceptable in a regulated file.

We build the engine to output the full Form 1084 with every line populated, every figure tagged to the source document and the specific line on that form, and every add-back and subtraction shown as a step. The underwriter sees the same worksheet they would have built, filled in, with the source for each number one click away. Their job changes from data entry and arithmetic to review and judgment: is the trend supportable, does the declining-income rule apply, is this K-1 distribution actually available to the borrower.

The engine does the mechanical work. It does not decide whether the income qualifies. A declining self-employment income trend is a judgment under B3-3.2, and the underwriter owns it. The engine surfaces the trend and the rule; the human makes the call.

The ATR/QM Reason This Has to Be Exact

Income is not just an approval input. It is a legal requirement. The Ability-to-Repay rule, 12 CFR 1026.43(c), requires the lender to make a reasonable, good-faith determination that the borrower can repay, using verified income and following, at 1026.43(c)(4), third-party records that provide reasonably reliable evidence of income. For self-employed income, those records are the tax returns and, where required, the tax transcripts.

The connection to the AI engine is direct. The engine's income figure is part of the ATR determination, so the engine's inputs have to be the verified documents, and the calculation has to be reconstructable. When a defaulted loan gets litigated, the ATR defense is the file, and the file has to show verified income calculated by a documented method. A worksheet tied to source documents is that defense. A black-box number is not.

Treating the Engine as a Model Under SR 11-7

A qualifying income calculation is a model. It takes inputs, applies a method, and produces an output that drives a credit decision. That puts it inside the model risk management expectations of SR 11-7, and inside the NIST AI Risk Management Framework if you are mapping controls to it.

Concretely, we manage the income engine the way the guidance says to manage a model. It has a documented methodology that maps every line to the Selling Guide reference. It has a development and validation record. It gets backtested: we run the engine against a set of files an experienced underwriter has already calculated by hand and measure agreement line by line, not just on the final income figure, because two errors can cancel to the right total and still be wrong on the worksheet. It has ongoing monitoring, so when the engine's numbers start drifting from the underwriters' overrides, that shows up on a dashboard and triggers a look.

The backtest is the part teams skip and the part validators ask for first. Our bar before an income engine goes into production is agreement with the human-calculated worksheet on the source lines, within a defined tolerance, across a population that includes the hard cases: Schedule C with home office and mileage, S-corp with a liquidity question, partnership K-1 with guaranteed payments, rental income on Schedule E with the fair-rental-days test. If the engine only works on clean W-2-plus-Schedule-C files, it is not ready, because those were never the files that needed help.

The Documents Lie in Predictable Ways

Self-employed files are where extraction quality gets tested, because the documents are messy in specific ways. A borrower provides a tax return that is not the one they filed. A K-1 is missing a page. A Schedule C shows a year of losses the borrower forgot to mention. An amended return sits in the file next to the original and the engine has to know which one governs.

The engine has to be built for adversarial input, not clean input. Our controls: match the return to the IRS tax transcript where the lender pulls one, so a self-prepared return that does not match what was filed gets flagged rather than trusted. Reconcile figures across related documents, so a Schedule C net profit that does not tie to the 1040 raises an exception. Detect the amended-return and multiple-version case and force a human to pick the governing document. The engine's confidence in a number is a function of how many independent sources agree on it, and a number with one unverified source is surfaced as low-confidence, not passed through as fact.

Fair Lending Lives in the Method

There is a fair-lending dimension to income calculation that gets missed because the calculation feels purely mechanical. It is not. Discretion in how income is calculated is discretion that can produce disparate outcomes, and self-employed borrowers skew toward certain demographics. If the engine applies add-backs more generously for some file patterns than others, or if underwriter overrides of the engine cluster by geography in a way that correlates with protected class, that is a fair-lending exposure sitting inside an operational tool.

We monitor it. The engine applies the same documented method to every file, which removes one source of unexplained variation. And we track overrides: when underwriters change the engine's income figure, we look at the distribution of those overrides for patterns that correlate with anything they should not. A consistent method plus override monitoring is a stronger fair-lending posture than a floor of underwriters each applying the Selling Guide from memory.

The Audit File

When a QC reviewer, an investor, or an examiner asks how the income was calculated, the file produces:

• The completed Form 1084 (or the Freddie 5304 equivalent), every line populated
• Each figure tagged to its source document and the specific form line
• The tax transcript match result, where pulled
• The reconciliation exceptions and how they were resolved
• The engine's confidence on each derived figure
• The underwriter's review, any override, and the documented reason for it
• The methodology version the engine ran, under change control

This is the file that turns a self-employed income question from a three-hour recalculation into a ten-minute review. The reviewer is not re-deriving the number. They are checking a worksheet that shows its work.

What We Tell Underwriting Leaders

Self-employed income is the right first place to put an AI income engine, precisely because it is the hardest. It is where underwriters spend the most time, where the most inconsistency lives, and where the audit exposure is highest. An engine that produces an auditable Form 1084 turns the worst part of the underwriting queue into the fastest.

The way we build it at Sei is to make the worksheet the product and the number a consequence, and to run the whole thing as a governed model rather than a feature. That is more work than shipping a monthly-income field, and it is the reason the output survives a QC pull, a model validation, and an ATR challenge instead of becoming the finding.