Skip to content
Sei

Overtime, Bonus and Commission: The Trending Math an AI Income Agent Has to Get Right, and Where Fannie Mae Moved the Rules in 2026

6 min readRamkumar Venkataraman
Share

Who This Is For

Heads of underwriting, income analysts and QC managers at lenders that sell to Fannie Mae, and model risk managers asked to approve an AI agent that calculates qualifying income. If your underwriters already know the variable income rules cold, the parts on earning-type separation, bonus timing and the 2026 renumbering are the ones worth your time.

Variable Income Fails Quietly

Self-employed income gets the attention because it is hard. Variable wage income gets less, because each paystub looks simple. The defects come from arithmetic choices nobody wrote down: which months were averaged, whether a bonus was counted once or annualized twice, whether a decline was noticed. Each choice moves the qualifying figure by a few hundred dollars a month, which is the margin between a loan that meets the debt-to-income limit and one that does not.

Those errors also survive review. A reviewer re-checking the file sees a monthly income figure and a worksheet, and unless the worksheet shows which earnings went into which average, the reviewer re-checks the math, not the method. An agent that only produces a number has the same problem at a larger scale.

Where the Rules Live Now

On March 4, 2026, Fannie Mae reorganized the income assessment chapter of the Selling Guide in Announcement SEL-2026-02. Bonus and overtime policy, which had lived in B3-3.1-03 alongside base pay, moved to B3-3.3-02, Bonus, Commission, Overtime, and Tip Income, and base income moved to B3-3.3-01, now split into fixed and variable base pay. Fannie described much of the move as a restatement for clarity that did not change the original requirements, while it also updated parts of the bonus, commission and tip policy.

For an underwriting team, a renumbering is a nuisance. For an agent whose every decision cites a rule, it is a change event. The Rulebook keys each rule by an internal identifier and keeps the citation as a versioned attribute with an effective date, so a loan decided before March 4 still cites B3-3.1-03 and a loan decided after cites B3-3.3-02, and both point to the same underlying rule where Fannie did not change its substance. Where the substance did change, it is a new rule version with its own effective date and its own regression run against archived files. Our post on rulebook change management covers that process in detail. The short version is that an auditor reading a 2025 file should see the citation that was in force in 2025.

The Substance: History, Trend, and Never Averaging a Decline

Three requirements carry most of the weight. Fannie's general income policy recommends a two-year history of variable income and accepts a shorter history, not less than 12 months, when positive factors offset it. The current B3-3.3-02 repeats the 12-month floor. Fannie also requires a trending analysis, and its language on declines is specific: if a declining trend has since stabilized and the borrower is likely to remain employed at that level, the current, lower amount must be used, and if the trend is still declining the income may not be stable, further analysis is required, and "in no instance may it be averaged over the period when the declination occurred."

Regulation Z's ability-to-repay rule sits underneath all of it. 12 CFR 1026.43(c)(2)(ii) and (c)(4) require the creditor to consider the income it relies on and verify it with reasonably reliable third-party records. For a qualified mortgage under the general QM definition, the creditor has to consider income and verify it consistent with that rule, and agency guides are the usual way lenders show they did. So the trending method is a compliance artifact as well as a credit one.

Four Places the Math Goes Wrong

The first is lumping earning types. A written verification of employment often reports a single "overtime" or "other" figure per year, and a paystub's year-to-date column may mix overtime, double time, holiday premium and shift differential. Each of those has its own history and its own trend. Overtime can be declining while shift differential is rising, and an average of the combined figure hides the decline that Fannie says must not be averaged. Our Income Desk extracts each earning code from the paystub as its own series, and when a written VOE combines types, it raises the question of whether the breakdown can be obtained rather than averaging the lump.

The second is annualizing a bonus. Fannie's own example in B3-3.3-02 is an annual bonus paid each March 31, which is annualized by dividing by 12. The trap is the year-to-date figure on a September paystub. If the borrower received a $9,000 bonus in March, the September year-to-date bonus column shows $9,000, and a calculation that divides year-to-date earnings by nine months elapsed produces $1,000 a month in bonus income when the real figure is $750. The agent uses the bonus payment dates it can see on paystubs and the written VOE to decide the bonus frequency first, then the divisor.

The third is the wrong end date. Year-to-date earnings run through the pay period end date, not the pay date. A paystub dated October 3 for the period ending September 27 covers a little under nine months. Divide by the months to the pay date and the monthly figure comes out low. Divide by nine when the period ended early in a month and it comes out high. Either way, the result depends on a date someone picked without writing it down. The agent computes months elapsed from the period end date it extracted, and the worksheet shows that date so the reviewer can see the choice.

The fourth is the decline nobody looked for. A simple 24-month average treats the last two years as equal. When the third year is lower, the average overstates current income, which is the case the "never average over the decline" sentence exists for.

A Worked Example

The numbers here are illustrative, not from a real file. An hourly borrower's written VOE breaks out overtime as $14,200 in 2024 and $11,900 in 2025. The borrower's paystub for the period ending September 27, 2026, shows $6,300 in year-to-date overtime.

A two-year-plus-YTD average across the 33 months comes to $32,400 divided by 33, or about $982 a month. That figure averages straight through a decline, from $1,183 a month in 2024 to $992 in 2025 to $700 in 2026 year to date. The trend is still falling, so under Fannie's language the overtime may not be stable at all, and if the underwriter concludes some of it is stable, the figure cannot come from an average across the decline. The defensible numbers are the current $700 a month or zero, and the choice between them is underwriting judgment about the employer and the borrower's role.

The gap between $982 and $700 is $282 a month. On a borrower at 44% DTI on the averaged figure, using the current figure can push the ratio past a 45% or 50% limit depending on the loan's other numbers, which is why this decision matters and why the agent does not make it alone.

What the Agent Decides, and What It Hands Back

For each earning type, the Income Desk extracts the history from the written VOE and paystubs, computes the monthly figure for each period with the end dates and divisors it used, and classifies the trend as increasing, stable, declined then stabilized, or still declining. Each classification cites the rule and the documents, page by page, in the Loan Graph.

Increasing and stable trends with at least the required history produce a qualifying figure, which then goes through Triple Check before it is written to the LOS: an independent checker recomputes the figure from the extracted values without seeing the maker's answer, and a disagreement stops the work. A trend that declined and stabilized produces the current lower figure, with the stabilization evidence cited. A trend that is still declining, or a history between 12 and 24 months, is never resolved by the agent. It goes to the underwriter as a named exception, with the series, the trend, the candidate figures and the rule, because both decisions turn on "positive factors" and the likelihood of continued employment, and those are judgments Fannie assigns to the lender's underwriter.

Re-runs matter too. A new paystub that arrives two weeks before closing extends the series, and the trend can change. The agent recomputes and, if the qualifying figure moves, flags it, so the loan does not close on a number the latest document no longer supports. Our post on verification of employment covers how the VOE itself is obtained and reconciled.

What This Looks Like at QC

When a QC reviewer opens a file, the question they should be able to answer in a minute is how the variable income was calculated and why that method. In the pre-close QC deployment at Better Mortgage, human review per loan file fell from 4 hours 50 minutes to 47 minutes, and 78% of items cleared at high confidence with a citation trail. Income is one of the areas where that trail does the most work, because the reviewer is checking a method rather than redoing a calculation. A worksheet that shows each earning type, each period, each divisor and the trend call, each tied to a page and a Selling Guide section, is one a reviewer can approve or challenge quickly.

Variable income will never be fully automatic, and it should not be. The agent's job is to make every arithmetic choice explicit and cited, do the cases the guide settles, and put the cases the guide leaves to judgment in front of the person whose judgment it is.

Ramkumar Venkataraman

Ramkumar Venkataraman

CTO & Co-Founder

Related Articles

Mortgage

The Clean File: Why the Strongest Mortgage Operations Build Quality In at Intake, Not at QC

Most mortgage defects are not created in underwriting or closing. They enter the file in the first days, as a missing page, a stale paystub or a figure that disagrees with its source, and they get more expensive at every stage they survive. An account of what a clean file actually means, why quality keeps getting pushed downstream to the most expensive people in the operation, and how our agents move the checks to the moment a document arrives.

7 min read
Read more
Mortgage

Reading a Leave and Earnings Statement Like an Underwriter: Military Income, the ETS Date and VA Residual Income for an AI Income Agent

A servicemember's Leave and Earnings Statement carries taxable base pay, nontaxable allowances that depend on duty station and dependents, allotments that can reveal debts the credit report does not show, and an expiration of term of service date that decides whether the income continues. A technical account of what our income agent reads from an LES, how it handles a PCS move and an ETS inside twelve months, and how VA's residual income test in 38 CFR 36.4340 is computed and cited.

6 min read
Read more
Mortgage

When Rates Drop, Your Servicing Portfolio Gets Called by Someone: Refinance Recapture With an AI Voice Agent, and the Tests That Decide Who It Calls

A rate move puts thousands of loans in the money in the same week, and the servicer has the data to reach those borrowers first. Before an AI agent dials any of them, it has to know which loans can actually close: VA's seasoning and recoupment tests under 38 U.S.C. 3709, the early payoff window on loans you sold, which consent covers a marketing call, and what the agent may and may not say about savings. What we learned running a voice agent on a lender's funnel, applied to the recapture campaign.

6 min read
Read more

You Ain't Seen Nothin' Yet

Book a Demo

Pack up some of your complex historical files — any loan type, any investor. We run them through intake, income and condition clearing, and in 30 minutes you see every condition we created and cleared efficiently for your own team, and why.

  • Any loan type, any agency guideline or custom investor overlays.
  • Every finding cited to the guideline or document it came from

Case study highlight

Loan review fell from 4 hours 50 minutes to 47 minutes per file, and 78% of items now clear at high confidence — so QC touches only the exceptions.

Pre-Close QC Agent · Better Mortgage— read the case study

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. One of our experts will get back to you.

Which desks would you like to discuss?*

Choose the desks you’d like us to cover. Pick “Not sure yet” if you’d rather we worked it out on the call.

Roughly, so we bring the right person to the call.

Complete the verification to submit the form.