Seller Discretionary Earnings With AI: Recast and Find Add-Backs

A broker can use AI to recast a seller's financials by importing a clean P&L, asking it to flag possible add-backs, then checking each candidate against ledger entries and evidence. Calculate seller discretionary earnings only from accepted adjustments. AI is useful as a finder and worksheet assistant; it cannot certify that a cost disappears for the buyer.

Start with a defined period and a reconciled profit figure. If the annual P&L, tax filing, general ledger and bank records disagree, an AI-generated add-back list will make the disagreement harder to see. The broker's deliverable is a bridge from reported earnings to supported SDE, plus an evidence request list and a column for rejected items.

Follow me on Instagram@sagnikteaches

Set one earnings period and freeze the source P&L

Seller discretionary earnings, or SDE, is a recast measure commonly used for a business with one working owner. The International Business Brokers Association glossary describes discretionary earnings before items such as income taxes, interest, depreciation and amortisation, one owner's entire compensation, non-operating or non-recurring items and personal expenses paid by the business. That description is a starting framework, not permission to add every unusual-looking debit. The broker needs to show what the business really earned before the defined adjustments.

Connect on LinkedInSagnik Bhattacharya

Export the annual P&L and general ledger for the same 12 months. Keep the original exports read-only and label the period precisely. Compare the P&L net profit to the bookkeeping trial balance and filing, noting timing or classification differences. Ask the seller to confirm whether the business is cash or accrual based; switching bases halfway through can double count revenue or expenses. If the year is incomplete, label the trailing period and show its months. Do not silently annualise nine months of trading.

Subscribe on YouTube@sagnikteaches

Before AI sees the files, remove customer names, payroll identifiers, bank account details and anything irrelevant to the recast. If the AI tool is outside the broker's controlled environment, check authority to share the seller's confidential data and the product's data terms. A broker may instead provide a redacted category-level export with transaction IDs that can be resolved internally. The client-data anonymisation tutorial gives the mechanics; the recast still needs an exact line back to the ledger.

Build a bridge that can be recalculated by hand

Use one row per proposed adjustment. The minimum columns are ledger account, transaction reference, period, amount, seller's explanation, SDE category, evidence link, buyer-replacement question, reviewer decision and accepted amount. Preserve zero and rejected rows. A reviewer should be able to filter to accepted items and reproduce the total with a calculator. Never put an AI-generated subtotal in the only copy of the worksheet.

Here is an illustrative one-year P&L for a small owner-operated distribution business. Revenue is $900,000. Cost of sales is $430,000, leaving $470,000 gross profit. Operating expenses are $390,000, including the owner's salary, payroll, rent, vehicles and marketing. Interest is $8,000 and depreciation is $12,000. Reported pre-tax profit is therefore $60,000. The example assumes these figures reconcile to the source statements; a real transaction needs that proof.

The $191,000 is an illustrative, supported bridge, not a valuation. The seller might initially present $205,000 by adding back marketing, but a buyer is likely to need marketing to sustain the revenue. The broker should keep both numbers visible in an internal working sheet, with the rejected $14,000 clearly outside the buyer-facing supported total. If a buyer later has a credible reason to treat part of that spend as non-recurring, revise the evidence and the bridge deliberately rather than hiding the change.

Make AI find candidates, not approve them

AI can scan ledger descriptions for owner-related, unusual and one-off items faster than a person browsing hundreds of rows. Its output should be a candidate list with transaction references, not a new financial statement. First provide the category definitions and tell the assistant to abstain when the entry is ambiguous. Then ask it for questions needed to confirm each flag. Give it a small, redacted batch for calibration before processing a full year.

Task: review these redacted expense lines for possible SDE add-backs.
Return columns: transaction ID, amount, candidate category,
reason to investigate, evidence needed, replacement-cost question.
Do not calculate accepted SDE. Do not classify an item as accepted.
Flag recurring operating costs as likely retained costs.
Source period: 1 January to 31 December, illustrative only.
Ledger lines: [paste approved redacted rows].

Suppose the illustrative output flags transaction 418, “director vehicle”, $12,000, as a full personal add-back. The broker checks mileage records and learns that 25% of the vehicle use was for deliveries. The accepted candidate becomes $9,000, assuming the records support that split. The $3,000 business-use portion stays as an operating cost. That correction is the difference between a persuasive worksheet and a sales claim the buyer can overturn in minutes.

Another illustrative output might label “equipment relocation” as one-off. Ask for the contractor invoice, date, site closure details and any further scheduled moves. If the business relocates equipment every year as part of normal delivery operations, the label is misleading even though this invoice occurred once. The broker's question is whether a comparable cost is likely to recur in operating the acquired business, not whether the exact invoice number appears twice.

Separate owner compensation from replacement labour

A common recast error is to add back all compensation for two owner-managers. SDE usually assumes one owner's compensation is added back; the buyer may have to replace the other person's work. If the second owner manages purchasing for 30 hours a week, removing that salary without an offset overstates the earnings available to a buyer. Record duties, hours, market replacement role and payroll cost separately. A broker should consult an accounting or transaction specialist before finalising a contested normalisation.

For example, imagine the owner receives a $70,000 salary and a spouse receives $45,000 for running the office. Payroll records prove both amounts. The AI tool flags $115,000. Interviews show that the spouse handles invoicing, debtor follow-up and supplier payments. A buyer still needs those tasks performed. The broker accepts the defined owner's $70,000 in the SDE bridge but does not automatically add the $45,000. A separate buyer scenario could show the cost of hiring an office manager, clearly labelled and supported rather than tucked into a higher SDE.

Benefits require the same care. A health benefit attached to the owner may be discretionary in the SDE definition, but the payroll and benefits records must show the amount. A vehicle lease may serve both business deliveries and personal journeys. A meal expense may be customer entertainment rather than a purely personal purchase. The seller's recollection is a lead for the evidence request; it is not evidence itself. Ask AI to draft the questions in plain language, then send them after a human checks tone and scope.

Put non-recurring claims on a calendar

“One-off” is a claim about the future economics of the business. A legal bill for a completed ownership dispute may genuinely be unusual; ordinary contract review is not. A repair caused by a single storm may differ from recurring maintenance on ageing equipment. Place each candidate on a three-year timeline, with prior and subsequent occurrences, so a buyer can see whether the same kind of cost repeats under different ledger descriptions.

Consider an illustrative $6,500 website redesign. The seller calls it one-off. The ledger shows $4,000 of design work in the previous year and the current year has a monthly hosting and content charge. The exact redesign invoice may be singular, but ongoing customer acquisition needs a working website. The broker can request the scope and distinguish an exceptional migration cost from routine marketing. If no defensible split exists, leave it out of supported SDE and disclose the question in diligence.

AI is useful here for matching descriptions that vary: “site move”, “warehouse transition” and “relocation contractor” may point to one event. It can also mistakenly group different expenses by shared words. Verify every flagged transaction against the general ledger and invoice. The messy-data cleaning tutorial helps prepare descriptions, but keep original wording and transaction IDs beside any cleaned label so the reviewer can trace each edit.

Turn the candidate list into an evidence request

Send a concise request grouped by claim, not a dump of every ledger line. For each proposed add-back, ask for the source transaction, proof of payment or accrual, the reason it is personal or exceptional, and whether a buyer would incur a comparable cost. Use an “evidence received” date and a reviewer initials column. If a seller supplies a screenshot cropped above the amount, request the full relevant record. If a cost is split, record the calculation that produced the accepted percentage.

A filled illustrative request might read: “For ledger transactions 418–425, vehicle expenses total $12,000. Please provide the lease schedule and the mileage or use log supporting the proposed 75% personal allocation. For transaction 602, please provide the $7,000 relocation invoice and confirm whether another move is planned. For marketing account 740, explain why $14,000 would cease after sale.” This asks for evidence without pre-accepting the seller's account.

AI can draft that email from the candidate rows, but it must not invent a missing document or imply the buyer has accepted a claim. Compare the draft to the register. If AI writes “the vehicle is entirely personal”, correct it to “the seller proposes a personal portion, pending records”. If it collapses several transactions into one total, ask for line references. Avoid uploading employee names, customer accounts or other unrelated records just to make the request sound personal.

Reconcile the bridge against the financial statements

Once each accepted row has evidence, recalculate from the frozen pre-tax profit. Check that a depreciation amount has not already been excluded from the starting figure. Check that owner payroll is in operating expenses and not included in a separate management-fee line. Check that a refunded legal expense is not added back without also treating the refund. A simple two-person review catches many duplicated rows that a polished AI summary obscures.

One illustrative mistake: the P&L lists $10,000 in “professional fees” and the ledger shows a $4,000 transaction for a sale adviser. The seller's candidate sheet also lists “sale costs, $4,000” as a separate line. If the broker adds both the full $10,000 account and the $4,000 transaction, SDE is overstated by $4,000 and perhaps more. The correct approach is to identify only the supported portion of the account, leave ordinary professional fees in expenses and show one adjustment row per accepted amount.

Another illustrative mistake is period mismatch. The seller presents the current 12-month P&L but provides an invoice for an exceptional cost booked in the previous year. The cost may be real, yet it cannot be added back to this period's starting profit. Put it on the comparative-year bridge instead. AI can check dates if they are structured, but humans must resolve late postings, accrual reversals and filing adjustments. A business-sale due-diligence checklist can carry unresolved questions into the wider deal process.

Show a range without turning this into a valuation

Where evidence is incomplete, present a supported SDE and a clearly labelled list of disputed candidates. For the main example, the supported figure is $191,000. The seller's additional $14,000 marketing claim would make $205,000 if accepted, but it remains excluded. A buyer can then challenge the operating assumption directly. Calling $205,000 “the SDE” before diligence would hide the single disputed issue in the headline.

Do not multiply either figure by a generic “market multiple” in this worksheet. Price depends on transferability, working capital, assets, liabilities, customer concentration, owner dependence, deal structure and buyer risk. AI is particularly prone to finding a neat multiple from an unrelated industry or period. Keep the deliverable to the recast and its evidence. The buyer qualification and NDA tutorial covers controlled access to fuller information after the initial anonymous stage.

A broker can finish with a compact audit packet: original P&L, tie-out to the ledger, accepted add-back register, rejected and pending claims, evidence index, assumptions note, and dated reviewer sign-off. Ask a second person to recalculate the bridge and open a sample of the evidence links. If that person cannot explain why each adjustment is accepted, the worksheet is not ready. AI has done its job when it shortens the search while leaving the human calculation clearer.

Test the recast against two buyer objections

The first objection is concentration. Suppose the same illustrative business earns $180,000 of its $900,000 revenue from one customer. That is 20% of sales. It does not change the arithmetic of the $191,000 SDE bridge, but it changes how much confidence a buyer may place in maintaining that earnings level. Put a customer-concentration schedule alongside the recast without uploading customer identities to a general AI tool. Ask the seller for contract term, renewal history and whether the relationship depends on the departing owner. Do not disguise a buyer risk by adding a larger earnings adjustment.

The second objection is working capital. A seller may say that a $30,000 stock purchase was “exceptional” because inventory was high at year-end. If the stock is needed to fulfil normal orders, the purchase is not automatically an add-back to earnings. Examine whether the item was expensed or recorded on the balance sheet, check the stock count and ask the accountant how the accounting treatment affects the P&L. AI can flag a large stock-related transaction, but it cannot determine the correct adjustment from the transaction description alone.

For a cash-basis illustration, a customer pays a $24,000 annual service contract on the final day of the year. A broker may be tempted to include the entire receipt in this year's earnings while also describing next year's work as “already sold”. That risks presenting one receipt as if it benefits two periods. Preserve the accounting basis of the source statements, document the timing and ask a qualified accountant whether a normalising adjustment is appropriate. The worksheet should never mix cash receipts and accrual expenses merely because the combined result looks smooth.

Use a review meeting to close the open rows

After the first AI pass, schedule a short review with the seller and their accountant. Send the disputed rows beforehand: transaction ID, proposed amount and the exact missing evidence. During the meeting, record the decision beside each row: accepted, rejected, pending or moved to another period. Avoid editing the original export while the seller explains it. If the explanation changes, retain the old note and date the new one so the buyer can see why the bridge moved.

For example, a $3,200 “owner travel” candidate may be supported by a hotel invoice, yet the trip could have been a supplier visit that a buyer would repeat. The reviewer asks for the diary entry and meeting purpose. If it was a supplier negotiation, the expense stays in operating costs. If it was a personal holiday booked through the business, it may be a candidate add-back, subject to proof and accounting review. The same account name can hide opposite decisions, which is why an automated blanket rule fails.

Record the revision after the meeting as a new worksheet version. If the owner vehicle allocation falls from $9,000 to $6,000 when a fuller mileage log arrives, supported SDE changes from $191,000 to $188,000. Update the table, calculation and any draft buyer material together. A version date prevents the broker's memorandum, data room and discussion notes from carrying three incompatible SDE figures. The buyer should never have to discover the changed assumption by comparing attachments unaided.

SDE questions a buyer may raise

Can the broker simply add back every owner payment?

No. The worksheet must distinguish one owner's compensation, benefits and personal expenses from costs the buyer will still incur. A second working owner or an essential replacement role can require a separate adjustment. Each proposed item needs a ledger source, reason, supporting evidence and a reviewer decision.

Is an AI-generated SDE figure a business valuation?

No. SDE is an earnings measure used in a recast. It does not establish a sale price, multiple, financing terms or buyer-specific return. A qualified professional should review the financial statements and transaction context before the figure is presented as verified or used in a valuation.

What if the seller cannot document an add-back?

Keep it in a separate unverified column and exclude it from the supported SDE total. Ask for a dated invoice, payroll record, contract, general ledger entry or other evidence that matches the period and amount. A plausible explanation without a matching source is a question for diligence, not an accepted adjustment.

Further reads

• Confidential Information Memorandum AI: A Business Broker's Workflow — Present supported figures in a buyer-facing memorandum.
• Business Broker Marketing: AI Buyer Outreach and Anonymous Listings — Prepare anonymous outreach after the numbers are checked.
• Business Broker CRM With AI: What to Compare Before You Buy — Choose a system for controlled deal records.
• Selling an Accounting Practice: Prepare the Due Diligence Pack With AI — See how a specialist practice assembles supporting records.
• Business Teaser Template With AI: Attract Buyers, Hide the Seller — An anonymous teaser template with five anonymisation rules, a before-and-after rewrite and a leak check for any AI-drafted listing.
• Do AI Business Valuation Tools Work? What They Miss for Brokers — Test valuation software against three closed deals and identify the facts a quick model cannot see.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: International Business Brokers Association glossary entries for add-backs, discretionary earnings and normalised financial statements.