M&A attorneys should treat an AI-flagged financial conflict as an unverified issue, not a corrected number. Preserve each source and its date, period, entity scope, units, and metric label. Have a human check the cited passages, then assign the accounting question to the seller's responsible financial professional and the legal implications to deal counsel. Record the explanation and supporting evidence before changing a confidential information memorandum, disclosure schedule, or buyer response. Do not let AI select preferred earnings or approve release. A controlled log should connect the issue to every affected document, with seller authorization and separate financial and legal review. Lawyers remain responsible for appropriately reviewing AI-assisted work. [1]
Hold the disputed statement while checking the flag
Start by identifying what the conflict could affect. A difference in revenue, adjusted earnings, debt, or working capital may change a buyer's understanding of the business. Ask the person controlling document distribution to hold further use of the affected statement while the team checks it. This is a targeted workflow recommendation, not a conclusion that every discrepancy requires suspending the entire transaction.
Open the original files rather than relying on the AI summary. Check the page, table, footnote, and spreadsheet cell the tool cites. Confirm whether the displayed amount matches the extraction, including signs, decimal points, column headings, and scanned text. A false flag should be documented as an extraction or comparison error, not silently discarded. NIST's generative AI profile recommends verifying output sources and citations before deployment and during ongoing monitoring. [3]
If the flagged document was already distributed, log that fact immediately. Counsel should assess whether interim communication or other action is needed under the transaction's circumstances. Do not wait for a convenient replacement number before raising a potentially consequential issue internally.
Check whether the figures describe the same thing
Require the comparison to identify the legal entity or business perimeter, reporting period, currency, scale, accounting basis, and actual-versus-forecast status. Record document dates separately from the periods covered. A recently prepared presentation can still contain an older reporting period. A business-unit figure should not be matched automatically to a consolidated figure.
For an illustrative example, a teaser might state annual revenue while a confidential information memorandum, or CIM, states revenue for the latest twelve months. The amounts can differ without either being wrong. The financial reviewer should identify the exact start and end dates and underlying records. If both documents simply say revenue, the issue may require clearer labels rather than substituting one amount everywhere.
Treat reported earnings, adjusted earnings, and forecasts as separate candidates until their definitions are established. SEC staff guidance explains that non-GAAP labels can be misleading and that similarly labeled measures may not be comparable. That guidance concerns the U.S. securities framework, not a universal rule for every private-company CIM. Counsel must assess applicability and any relevant transaction-specific exceptions before drawing a regulatory conclusion. [2]
Create a discrepancy log that preserves both candidates
Give each discrepancy a stable identifier. Record both amounts exactly as displayed, their labels, filenames and versions, publication or preparation dates, and page or cell references. Add the entity scope, period, currency, units, and accounting basis. Preserve a controlled copy of the source or an authorized reference to it. The log should distinguish what the document says from what AI inferred.
Each entry also needs an accountable owner, the question requiring an answer, a review deadline, distribution history, and affected materials. Record priority as an internal triage decision, not an AI determination of legal materiality. Possible states include awaiting source, financial review, legal review, held unresolved, and closed with evidence. A closed item must identify the supporting explanation and the humans who approved the relevant correction.
Keep the financial finding separate from legal analysis. The first might explain a period mismatch or an omitted adjustment. The second might address a representation, contractual notification requirement, or disclosure question. Access and privilege treatment require matter-specific legal assessment; merely placing either note in a discrepancy log does not establish privilege.
Assign accounting answers and legal decisions to different owners
Deal counsel should coordinate the escalation without becoming the default accounting decision-maker. Direct the issue to the seller's authorized finance lead, such as its controller or chief financial officer. Where the question exceeds that person's remit, involve the appropriately engaged accountant or financial diligence professional. Confirm the engagement scope before asking anyone to provide assurance or an opinion.
Send a neutral, human-reviewed request: identify the conflicting passages, specify the comparison, and ask which records explain the difference. Request the calculation bridge, version history, and supporting schedules where relevant. Do not ask only which number the seller prefers. The seller's response is evidence to evaluate, not automatic verification that overrides the underlying records.
Financial professionals should explain accounting treatment and validate the supporting calculation within their roles. Counsel should evaluate document language and legal consequences. The seller's authorized decision-maker approves the business information for release. ABA Formal Opinion 512 emphasizes that AI cannot replace the professional judgment required of lawyers; its model-rule analysis must be applied with the governing jurisdiction's rules in mind. [1]
Do not use AI to normalize earnings or settle disagreements
Limit AI to identifying possible conflicts and organizing supported evidence. Do not instruct it to choose the most credible number, remove recurring expenses, harmonize accounting policies, or infer missing adjustments. A plausible explanation can conceal an unsupported change. A disclaimer attached to the resulting number does not supply the missing financial evidence.
For adjusted earnings, request the starting measure, each adjustment, its description, and its source. Have the financial reviewer check the bridge using a spreadsheet or calculator, including duplicated items, signs, period allocation, and changes between versions. Preserve disagreements about an adjustment as disagreements. Counsel may need to revise the description, but should not resolve the accounting judgment through a wording change.
SEC staff distinguishes EBITDA based on GAAP net income from differently calculated measures such as Adjusted EBITDA, and identifies potential problems with inconsistent adjustments between periods. These interpretations illustrate why labels and bridges matter. They do not establish that any particular private-sale adjustment is permissible or impermissible; the applicable rules and facts require qualified review. [2]
Set confidentiality permissions before submitting deal records
The supervising attorney and the firm's technology or information-security reviewer should determine whether the proposed tool and configuration are permitted for the matter. Examine provider terms, retention, training use, access arrangements, and deletion provisions rather than assuming an account described as private is sufficient. Check client instructions and contractual restrictions on the seller's and third parties' records.
ABA Formal Opinion 512 calls for a fact-specific assessment before inputting representation-related information into generative AI. It discusses informed consent for self-learning tools that risk later disclosure and explains that general engagement-letter boilerplate is insufficient when informed consent is required. Whether consent or additional safeguards are necessary depends on the tool, matter, and governing professional rules. [1]
Use synthetic material for initial testing where possible. For authorized live use, submit only the necessary records or excerpts and restrict access to approved reviewers. Redaction must preserve the periods and labels needed for comparison, but removing names alone does not necessarily remove confidentiality concerns. If permissions or provider treatment of the information cannot be established, stop the upload and use manual review.
Run a bounded trial with explicit stopping conditions
Before using AI across a deal file set, the supervising attorney should approve a limited trial. Give the tool a narrow task: extract labeled figures from specified documents, preserve their locators, and flag possible mismatches without selecting a corrected figure. Have a financial reviewer and a legal reviewer manually establish the expected findings for the trial documents.
Include controlled examples of genuine conflicts, different reporting periods, differently defined metrics, and difficult tables. Compare flags against the original pages and check missed conflicts as well as false positives. Require the tool to report missing evidence rather than invent a locator or explanation. These are recommended trial controls, consistent with NIST's emphasis on documented testing and limits on extrapolating performance from narrow assessments. [3]
Stop the trial if the tool invents sources, repeatedly loses units or period labels, misses a consequential seeded conflict, or selects unsupported earnings. Stop immediately for unauthorized access or disclosure. The responsible humans should investigate and authorize any restart. Changes to the tool, instructions, or document format warrant renewed testing; a successful trial is not permission to remove review of material outputs.
Track verified corrections through every affected disclosure
Once the financial explanation is supported, record whether the issue was an extraction error, a labeling problem, a version mismatch, or a financial correction. Preserve the original candidates and attach the reviewer-approved explanation. Do not overwrite the history. If the explanation remains unsupported, keep the item open and withhold the affected conclusion rather than converting uncertainty into an approved amount.
Build an impact list covering the teaser, CIM, financial schedules, diligence responses, valuation inputs, draft contractual definitions, and disclosure schedules where applicable. An authorized document owner should make proposed changes and prepare a comparison against the prior version. Financial review checks the figures and labels; legal review checks the surrounding statement and transaction implications. Seller authorization follows the agreed release process.
Record who received earlier versions and how an approved correction will be communicated. Do not assume replacing a data-room file reaches prior recipients. Counsel must decide any notification or remedial steps under the applicable law and agreements. AI may help draft an internal comparison, but must not contact buyers, approve disclosures, or negotiate the consequences.
Close the issue only with evidence and a human release decision
Before release, require confirmation of the financial basis, legal review, seller authorization, and completion of the affected-document checks. If a deadline arrives first, escalate the unresolved status instead of selecting the newer, larger, or more convenient figure. Under ABA Model Rule 4.1, lawyers may not knowingly make false statements of material fact to third persons. Local adoption and the facts govern; an AI flag alone neither proves falsity nor creates a blanket duty to disclose every inconsistency. [4]
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.
Acquivo is user-operated software with no success fee. The discrepancy workflow described here is a recommended framework, not a claim that particular AI, logging, or approval features are available. Software does not represent the client or supply accounting or legal approval. Use the log to leave a clear record of what was questioned, what evidence resolved it, which documents changed, and who authorized release.
Key Points
- Treat an AI flag as unverified until a human checks the original passage, label, period, units, and scope.
- Preserve both candidate figures and their source versions in a log with an owner, status, deadline, and affected-document list.
- Route accounting questions to appropriately engaged financial professionals, legal consequences to counsel, and release authorization to the seller's authorized decision-maker.
- Do not let AI choose earnings, infer adjustments, approve disclosures, contact buyers, or resolve accounting disagreements.
- Establish tool-specific confidentiality permissions and a bounded trial with stopping conditions before live use.
- Close a discrepancy only with supporting evidence and completed human review of downstream corrections.
Sources and references
Source retrieval dates record research access, not article publication. Read our editorial and date policy.
- [1] ABA Formal Opinion 512: Generative Artificial Intelligence Tools, July 29, 2024 · Accessed
- [2] SEC Division of Corporation Finance: Non-GAAP Financial Measures, last updated December 13, 2022 · Accessed
- [3] NIST AI 600-1: Artificial Intelligence Risk Management Framework, Generative Artificial Intelligence Profile, July 2024 · Accessed
- [4] ABA Model Rule 4.1: Truthfulness in Statements to Others · Accessed
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.