Appendix: Financial Due Diligence (Quality of Earnings)

Adjusted (normalized) EBITDA is reported EBITDA plus or minus a schedule of adjustments, each intended to move the number closer to sustainable earnings. The adjustments, commonly called add-backs even when some reduce the figure, tend to fall into recognizable categories:

• Owner compensation above (or below) market. A founder who pays themselves well above a market salary for the role depresses reported earnings; normalizing to a market rate adds the excess back. The reverse happens when an owner takes little or no salary.
• One-time items. A litigation settlement, a failed system implementation, or non-recurring professional fees inflate expenses in a single period. If the item is genuinely non-recurring, it is added back so it does not drag the run-rate a buyer inherits.
• Run-rate adjustments. A price increase or a headcount change that took effect partway through the year is annualized so the figure reflects a full year at the new level, in either direction.
• Related-party normalization. Rent paid to a building owned by the seller, or a management fee charged by an affiliate, may sit above or below a market rate. Normalizing it to market removes a benefit or burden that will not travel with the business.

The discipline in each category is the same. An add-back is only as good as its support, and a buyer tends to challenge each one, because each dollar added to EBITDA is multiplied into the price.

Two deal terms ride alongside the earnings number. The first is the net-working-capital peg. A business needs a normal level of working capital (receivables plus inventory less payables, broadly) to run day to day, and the buyer expects to receive the target with roughly that level in place. The peg is a target working-capital level set in the purchase agreement, frequently a trailing average such as the last twelve months, because a single month-end can be unrepresentative. At close, actual working capital is compared to the peg, and a true-up adjusts the price: deliver less than the peg and the price is reduced, deliver more and it rises. The true-up mechanism is a standard term in private-company purchase agreements.

The second is debt-like items: obligations that behave like debt even when they are not labeled as such, for example accrued but unpaid bonuses, deferred revenue the buyer will have to service, unfunded liabilities, or deferred capital spending. These sit outside EBITDA but reduce the equity value a buyer will pay, so diligence flags them separately in the bridge from enterprise value to equity value. Missing one is a classic way a buyer overpays.

The finish line is fixed before the work begins. A buyer-defensible quality-of-earnings analysis has four properties. The bridge foots, so reported EBITDA plus the net adjustments equals adjusted EBITDA to the dollar. Each add-back is supported, so there is a documented basis a buyer could test rather than an assertion. Each adjustment is correctly labeled one-time or run-rate, so the reader can judge what persists. And the number is ownership-neutral, restating away benefits and burdens specific to the current owner rather than padding the figure to lift the price.

The task suits AI assistance because of how the work divides once the adjustment schedule exists. Much of the remaining work is mechanical and linguistic: totaling the bridge, averaging twelve months of working capital, labeling each item, and drafting a clear support note in a diligence voice. The judgment, whether an add-back is defensible, whether a cost is truly non-recurring, and what the numbers imply, is where the analyst adds value. That split, deterministic arithmetic on one side and judgment plus narrative on the other, is what the rest of this appendix is built around.

The most useful thing you can give a model for this task is an example of the destination. A prior deal's quality-of-earnings bridge and support note is exactly that: it shows the house format, the level of detail on each add-back, and the tone a diligence report carries. With that example in the folder, you are not asking the model to guess what a defensible bridge looks like; you are asking it to produce more of a known shape. The path is fixed: reported EBITDA and a candidate schedule at the near end, a supported adjusted-EBITDA bridge and a working-capital peg at the far end, with the worked example anchoring that far end.

Totaling a bridge is deterministic work: for a given schedule of adjustments, there is one correct net figure and one correct adjusted EBITDA. Averaging twelve months of working capital is the same, one correct number. Asking a language model to produce those figures in prose invites small, hard-to-catch errors in a number that gets multiplied into a price. So the workflow puts the math where it belongs. A template holds the bridge (reported EBITDA plus each labeled adjustment equals adjusted EBITDA) and the trailing 12-month working-capital average, and the model narrates and labels only. The model performs the non-deterministic work, the part with many acceptable versions: how to describe the basis for each add-back and how to sequence the findings.

This is the single most important design choice in the appendix. When the model recomputes the bridge in prose, it can drift from the template, and now two versions of the earnings number exist. When the template owns the arithmetic, there is one source of truth, and validation becomes a tie-out rather than a recalculation.

A single line makes the split concrete. The schedule shows the owner drawing $650,000 in compensation against a market rate of $400,000 for the role. The workflow carries that add-back from the schedule to one supported line, and each step has an owner.

• Template, deterministic. The add-back is $650,000 less $400,000, or $250,000, and the bridge adds it to reported EBITDA. There is one correct figure, so the template computes it and the model is told not to recompute it.
• Label, judgment. Above-market owner pay is a recurring feature of the current ownership that a new owner would not carry, so it is normalized as an ownership adjustment rather than a one-time item. That labeling is the analyst's call, not the model's.
• Support, judgment. The basis is a market-compensation reference for the role. The analyst confirms the $400,000 rate is defensible, because a buyer will test exactly this line.
• Narrate, non-deterministic. The model turns the figure and its basis into one measured line: "Owner compensation of $650,000 is normalized to a market rate of $400,000 for the role, a $250,000 add-back, reflecting an ownership-specific cost a new owner would not inherit."

The lab folder is deliberately small: the reported EBITDA and the candidate adjustment schedule, twelve months of net working capital, and the one worked example. Nothing else. A full data room would bury the accounts that matter and raise the chance the model latches onto something irrelevant. Minimum context is the fuel: give the model exactly what the task needs, and the destination example so it knows the shape of the answer. In diligence it is a security control as well as a quality one, because a target's financials are highly confidential.

The worked example does work here that written instructions tend to miss. It encodes the tacit house format, the level of support expected on each add-back, and the sequencing a diligence lead would struggle to spell out as a rule, and the model tends to match a shown example more reliably than it follows an abstract description of one. That is why the folder carries a finished bridge rather than a longer style guide.

The five moves are not a one-off. Because diligence runs the same shape on each new target, the prompt that carries these instructions is worth saving as a reusable skill or prompt template: the destination example, the deterministic-math rule, the minimal-folder list, and the one-time-versus-run-rate labeling guidance, packaged so the next target starts from a known shape rather than a blank page.

Pair it with a short support checklist the analyst runs before sign-off: confirm the bridge foots, trace each add-back to a basis in the schedule or flag it as unsupported, confirm each label fits the item, and confirm the peg's stated basis. The skill makes the draft repeatable; the checklist makes the review repeatable. Together they keep the human checkpoint from becoming a rubber stamp.

Download the folder and run the quality-of-earnings pattern in whatever AI you use. The folder holds the target's reported EBITDA and a candidate adjustment schedule, twelve months of net working capital for the peg, and a prior deal's bridge as your format example. Let the template hold the arithmetic, have the model total the bridge and label each add-back, then come back for support and validation. The target and every number are fictional.

The heart of validation is support and tie-out: take each adjustment in the bridge and follow it back to a basis in the schedule, and confirm the bridge foots, so reported EBITDA plus the net adjustments equals adjusted EBITDA to the dollar. Confirm, too, that each item is labeled one-time or run-rate in a way that fits, that the peg's stated basis (for example a trailing 12-month average) is the one actually computed, and that debt-like items are flagged rather than buried. A single unsupported add-back is enough to send the schedule back. Work the checklist below against your bridge before you call it done.

The exemplar below is what you would put in front of the deal team. It leads with the earnings number, labels each add-back, ties each figure to a basis, and separates one-time from run-rate.

Adjusted EBITDA. Reported EBITDA of $6,000,000 normalizes to $6,830,000 after $830,000 of supported adjustments. The largest is $250,000 of above-market owner compensation ($650,000 drawn against a $400,000 market rate for the role), an ownership-specific cost a new owner would not carry. A further $300,000 reflects one-time items, a litigation settlement and non-recurring professional fees, that will not recur. A $180,000 add-back normalizes above-market related-party rent to a market level. A net $100,000 run-rate adjustment annualizes a mid-year price increase ($220,000) less the full-year cost of two late-year customer-service hires ($120,000).

What persists and what does not. The owner compensation and related-party rent adjustments are recurring ownership normalizations, so the buyer should expect them to hold. The $300,000 of one-time items should not be read into the run-rate. The run-rate items are, by construction, the sustainable full-year view.

Open questions. The market-compensation reference and the market-rent basis are the two lines a buyer is most likely to test, and both should be supported with third-party references before the number is relied on.

Alongside the earnings number, the analysis sets the net-working-capital peg. Averaging the target's twelve month-end working-capital balances gives a trailing 12-month average of $9,126,083, and that is the peg carried into the purchase agreement as the target level at close.

The peg does real work in the deal. At close, actual working capital is compared to $9,126,083, and the true-up adjusts the price: if the target delivers less than the peg, the buyer funds the shortfall and the price is reduced; if it delivers more, the price rises. Stating the basis (a trailing twelve-month average, chosen so a single unrepresentative month-end does not set the target) is what makes the peg defensible, because the basis itself is negotiated.

The lab is clean by design: the schedule is curated, the folder is minimal, and the adjustments are supportable. Real diligence is messier, and the workflow tends to strain in three specific places.

• Over-normalized, seller-favorable add-backs. A seller-prepared schedule tends to add back everything it can, dressing recurring costs as one-time to lift the multiple. The buy-side job is to strike the add-backs that will not survive scrutiny, and a model reading the schedule at face value will carry them straight through if you do not test each basis.
• Working-capital manipulation near close. A trailing average can be gamed by pulling collections forward and stretching payables in the weeks before close, so delivered working capital looks healthy while the business is actually short. Seasonality does the same thing innocently. Either way, the peg and the true-up need a normalized, seasonally aware basis, not a raw month-end.
• Debt-like items hiding below EBITDA. Accrued but unpaid bonuses, deferred revenue the buyer will service, deferred capital spending, and unfunded liabilities each reduce equity value without touching EBITDA. A bridge that stops at adjusted EBITDA and ignores them can make a buyer overpay even when the earnings number is right.

On top of these, a model will sometimes produce a bridge that foots cleanly and is still wrong, because a footing bridge and a defensible bridge are not the same thing. That is exactly why the support-and-tie-out check is not optional. The workflow does not remove the analyst's judgment; it removes the mechanical drudgery so the judgment has more room.

Rate your own bridge against the rubric below. An honest score against the rubric shows which parts of the quality-of-earnings review are solid and which still need practice. Your scores roll up to the workflow maturity dashboard on the course hub.