Bankers should use AI to find and organize possible adjustments, not to decide that an item qualifies as an EBITDA addback. Put each suggestion in a source-linked register that identifies the underlying ledger entry, owner explanation, accounting support, recurrence pattern, and any replacement or continuing cost. A qualified reviewer should then test whether the adjustment is relevant to the agreed normalization purpose, consistently treated, and supported by records. Mark each item as accepted, rejected, pending, or escalated, with a reason and next evidence request. Keep unresolved exceptions visible. This approach makes the review auditable without implying that software certifies earnings or predicts buyer acceptance.
Start with the decision the register must support
An addback review is not a hunt for expenses that look unusual. It is a controlled explanation of how a reported profit figure is being recast for a defined purpose. The purpose might be a management presentation, a lender discussion, a valuation analysis, or another transaction workstream. Those uses can apply different conventions and review thresholds, so the banker should state the intended basis before asking an AI tool to extract candidates.
The register should therefore begin with the reported period, source financial statements, chart-of-accounts mapping, normalization objective, and accountable reviewer. It should distinguish reported EBITDA from the proposed adjusted figure. It should also preserve the original description and amount rather than replacing them with an AI-generated label. A suggestion is only a work item until the evidence and the reviewer support a disposition.
Use AI for extraction, classification, and questions
A narrow AI task can be useful: scan an approved set of general-ledger exports, expense detail, invoices, payroll reports, contracts, and management explanations; identify language that may warrant review; link each candidate to a source location; and draft questions. The output should include confidence or uncertainty notes, but those notes are triage signals, not an eligibility decision.
Give the tool a constrained instruction. It can be asked to find unusual descriptions, one-time-looking charges, owner-related items, transaction costs, discontinued activities, or expenses with a possible replacement requirement. It should not be asked to maximize adjusted EBITDA, select the most favorable result, approve a seller claim, contact a buyer, or make a disclosure. FINRA describes document summarization and querying as possible AI uses while emphasizing accuracy, privacy, model risk, and human supervision for member firms. [3]
Build a source-linked evidence register
Each row should be traceable without reopening a chat transcript. Useful fields include a stable item ID, period, account and journal reference, original amount and currency, proposed adjustment, source-file name and page or row locator, management or owner explanation, and the person who supplied that explanation. Add a link to accounting support such as an invoice, payroll record, contract, board authorization, or bank statement where the file set permits.
Add separate fields for recurrence, operating necessity, replacement cost, tax or accounting treatment where relevant, and the reviewer's conclusion. Record whether the item is supported, partially supported, contradicted, or missing evidence. A short rationale should explain the disposition in plain language. This structure prevents a polished AI summary from becoming the de facto workpaper and lets another reviewer reproduce the path from suggestion to conclusion.
Test the owner claim against accounting support
The owner or management explanation is important context, but it is not the same as independent support. Compare the narrative with the ledger, invoice, contract, payroll detail, payment record, and period timing. Check that the proposed amount agrees to the underlying records and that a single event has not been split across accounts or periods. If the amount is estimated, record the method, inputs, and owner of the estimate instead of presenting a false exact figure.
Ask what would change if the source document were absent. An item with only a verbal explanation should remain pending or be escalated according to the firm's review policy. Conflicting descriptions should remain visible. Do not allow the AI system to resolve the conflict by choosing the explanation that produces the larger adjustment. The reviewer should request the missing record, narrow the adjustment, or reject it with a documented reason.
Separate recurrence from unusual appearance
A charge may look unusual because its amount is large, its description is vague, or it appears in an unfamiliar account. None of those facts proves that the cost is nonrecurring. Review prior periods, later periods when available, budgets, contracts, maintenance schedules, and the operating context. Look for similar costs under different descriptions and for related gains that could make a one-sided adjustment misleading.
The SEC's non-GAAP interpretations warn that excluding normal, recurring, cash operating expenses necessary to operate a business can produce a misleading measure. The guidance also treats an expense that recurs occasionally or at irregular intervals as recurring, and addresses inconsistent treatment across periods. [1] For a private-company transaction analysis, this is not a universal formula for buyer acceptance, but it is a strong reason to make recurrence a documented human review question rather than an AI label.
Ask whether the cost must be replaced
An owner-related salary, family expense, discretionary benefit, or unusual service may be presented as removable. The practical question is whether the business can operate after the transaction without a comparable cost. Identify the work performed, who performs it, the hours or responsibilities involved, and whether another employee, contractor, or manager would need to take over. Record a replacement-cost estimate only when its basis is stated and reviewable.
Do not treat the absence of a current payroll line as proof that the work has no value. Conversely, do not add a replacement cost automatically without understanding the operating model. The register can show a proposed gross adjustment, an estimated continuing cost, and the net effect as separate fields, with the replacement estimate marked as an assumption until reviewed. If the evidence cannot support a reasonable conclusion, preserve the item as unresolved rather than forcing a number.
Apply permissions and confidentiality controls before the pilot
Before uploading financial records, define which files the tool may access, whether the provider retains prompts or documents, who can see outputs, and how records will be deleted or exported. Use a redacted or synthetic test set when the workflow can be evaluated without live deal information. Access should follow the engagement's permissions and the firm's information-handling policy. A convenience feature does not expand a user's authority to disclose client or counterparty material.
NIST describes its AI Risk Management Framework as voluntary guidance for managing AI risks and incorporating trustworthiness into the design, development, use, and evaluation of AI systems. [2] For this use case, that means documenting the permitted data set, the task boundary, known failure modes, reviewer role, and evidence retention decision. It does not mean that a tool is secure, confidential, privileged, accurate, or compliant merely because it is described as AI-enabled.
Set human dispositions and stopping conditions
Use a small status vocabulary such as accepted for review basis, rejected, pending evidence, and escalated. Require the reviewer to record why the status was chosen, which source was checked, and what would resolve a pending item. A second qualified reviewer or appropriate accounting, tax, or legal professional may be needed where the issue turns on a specialized judgment or the transaction's governing documents. This article does not decide those questions for a particular engagement.
Stop the pilot if the tool repeatedly invents citations, merges separate transactions, cannot preserve source locators, exposes restricted data, or produces suggestions that reviewers cannot efficiently verify. Stop using an output for a decision if the source is missing, the period is wrong, the amount does not reconcile, or the proposed treatment depends on an unsupported assumption. FINRA says firms should evaluate AI tools before deployment and maintain controls for privacy, integrity, reliability, accuracy, and supervision where its rules apply. [3]
Deliver a review-ready register, not a promise
The final work product should show reported EBITDA, each proposed adjustment, evidence links, recurrence analysis, replacement-cost analysis, reviewer disposition, unresolved questions, and the date of the latest review. Keep rejected items in the record with reasons so they are not reintroduced by a later prompt. Keep assumptions separate from sourced facts, and label any estimate with its basis and sensitivity rather than presenting it as accounting truth.
Acquivo is a software platform that delivers the entire sell-side M&A process, valuation, marketing materials, buyer matching, and deal management, on one platform for business owners, M&A attorneys, and boutique investment banks. In a user-operated workflow, the platform can serve as a place to organize the proposed-addback register and review trail. The banker still owns the evidence standard, permissions, judgment, and final disposition. The implementation outcome is a proposed-addback register with evidence, unresolved questions, and human disposition for each adjustment.
Key Points
- Use AI to extract candidates, source locations, and review questions, never to decide what qualifies as an addback.
- Keep owner claims, accounting support, recurrence analysis, replacement costs, assumptions, and reviewer disposition in separate fields.
- Test unusual-looking expenses across periods and related gains before calling them nonrecurring.
- Treat replacement cost as a separate, supportable estimate rather than silently accepting the gross proposed adjustment.
- Pilot only with approved permissions, a narrow task, human review, and explicit stopping conditions for missing or unreliable evidence.
Sources and references
Source retrieval dates record research access, not article publication. Read our editorial and date policy.
Choose your next step
Acquivo is user-operated software. Owners and their professionals remain responsible for decisions, communications, and professional review.
Educational information, not individual legal, tax, or investment advice. Consult qualified professionals for your circumstances. Acquivo provides software, not representation or negotiation on your behalf.