How Should M&A Attorneys Verify an AI-Assisted IOI Comparison Table?

Define the task and the accountable reviewer

Write a narrow instruction before generating the table: extract stated terms from the approved source set, attach document locators, flag uncertainty, and make no recommendation. The task is to check whether the table faithfully represents the documents. Determining which offer is preferable, whether a provision is enforceable, or what response serves the client belongs to separate human analysis.

Assign a preparer to manage extraction and corrections, a responsible M&A attorney to resolve wording and legal-context questions, and a finance reviewer when a proposed normalization requires financial interpretation. One person may fill more than one role, but responsibility should be explicit. The attorney should define permitted recipients and the person authorized to release the reviewed table.

ABA Formal Opinion 512, issued in July 2024, interprets the ABA Model Rules and says review should fit the tool and task; it does not prescribe this exact checklist. Check applicable local professional rules and current guidance rather than treating the opinion as law everywhere. Professional judgment remains with the lawyer. [1]

Clear document permissions before using live IOIs

Before inputting transaction documents, the responsible attorney should review firm policy, client instructions, confidentiality undertakings, and the proposed tool's contractual terms. Identify who can access inputs and outputs, whether information can be used for training, applicable retention settings, and whether shared workspaces expose the material to unauthorized colleagues. Obtain technical assistance where the terms or settings cannot be understood reliably.

Under the Model Rules analysis in Opinion 512, confidentiality review is fact-specific, and some uses require informed client consent. Determine disclosure and consent requirements under the rules applicable to the matter and the actual tool configuration. A generic engagement clause or a product label should not replace that assessment. [1]

If permissions remain unresolved, do not upload live IOIs. Use synthetic documents to test the workflow or perform manual extraction in an approved environment. Redaction can support data minimization, but removing names alone does not establish confidentiality or privilege. Treat the comparison table and correction record as sensitive material too; intended circulation needs its own human check.

Freeze the source set and define the comparison fields

Create a document register before extraction. Record the submitting party's approved identifier, filename, document date, version, receipt record, and included attachments. Identify any later clarification separately and have the attorney decide whether it supplements or supersedes an earlier statement. Do not silently let the newest filename replace the source without checking its contents.

Choose rows that fit the actual documents. Potential fields include stated valuation basis, transaction perimeter, consideration components, retained ownership, financing conditions, diligence assumptions, timing, and any exclusivity or confidentiality language. These are review prompts, not assertions that every IOI contains those terms. Add an exceptions field for material wording that does not fit the initial schema.

For each buyer-and-term cell, retain the source wording, a concise extracted statement, any normalized display value, the document locator, a validation status, and reviewer comments. Keep extraction separate from interpretation. If the original is scanned, retain the page image as the reference and identify optical character recognition, or OCR, text as a derivative rather than the source of truth.

Validate populated cells and then search for omissions

Work row by row across the table. Open the cited original for each material populated cell, confirm the correct party and version, and read enough surrounding text to capture conditions and exceptions. Look for misplaced columns, copied terms from another offer, dropped negative wording, and summaries that turn a conditional statement into an unconditional one. Record the result rather than relying on a general impression that the table looks plausible.

Then reverse the direction: read each complete IOI and map its material statements back to table rows or the exceptions field. This coverage pass catches information that never reached the table and therefore has no cell to inspect. Review attachments and footnotes within the approved source set, not just the main narrative.

The two passes answer different questions: whether included content is faithful, and whether the table is materially complete for its defined purpose. NIST describes generative AI confabulation as confidently presented erroneous content, including output that diverges from its input. Polished language is not evidence that either check has passed. [2]

Preserve ranges, units, and financial bases

Check each number character by character against the original, including currency symbols, decimal separators, scale labels, percentages, and signs. Preserve a stated range as a range. Do not replace it with a midpoint, treat an upper bound as a fixed offer, or assume an omitted currency. If a value depends on assumptions elsewhere in the IOI, carry those assumptions into the display or a clearly visible note.

Keep differently defined quantities separate. A statement labeled enterprise value should not be relabeled as equity proceeds. A proposed cash component should not absorb contingent consideration or retained ownership merely to create one headline number. Review debt, cash, working-capital language, and the transaction perimeter before deciding whether values are comparable. Where the basis is unstated, label it unknown rather than supplying a customary assumption.

If a reviewer authorizes a derived figure, show it separately from the quoted term and preserve the formula, inputs, units, and source locations. Recalculate it with a deterministic spreadsheet or calculation tool. An AI explanation of arithmetic is not independent verification. Do not calculate a multiple without a verified numerator and denominator using an explicitly approved period and definition.

Keep transaction structure and legal qualifiers intact

Use separate fields where similar words describe different proposed structures. Check whether the document addresses assets, shares, selected operations, excluded assets, retained liabilities, or continuing ownership. Do not force those alternatives into an apparently uniform comparison that conceals a different transaction perimeter. Any common comparison basis should be a separately labeled human-reviewed analysis, not a rewritten source term.

An illustrative wording check: if an IOI states that a proposal is subject to financing approval, a cell reading 'financing secured' is not a faithful extraction. Likewise, a proposed diligence schedule should not become a guaranteed closing date. Preserve qualifying phrases such as 'anticipated,' 'subject to,' and 'to be agreed' when they affect the statement's meaning.

Capture any language addressing binding effect, conditions, expiration, or exclusivity without deciding its legal consequence automatically. The responsible attorney should interpret the actual wording under the relevant law and facts. Neither the document's IOI label nor the table's shorthand establishes enforceability. Route those questions to counsel rather than asking the model to resolve them or draft a negotiating response.

Verify references and give unknowns distinct meanings

Open every material locator and confirm that it leads to the correct document, page, and passage. Distinguish PDF page numbers from printed page numbers when they differ. A valid-looking reference that points to an unrelated paragraph is an error, even if the cell happens to be correct elsewhere. NIST specifically recommends reviewing and verifying sources and citations in generative AI output during testing and ongoing monitoring. [2]

Use explicit statuses. 'Not stated' means the approved source set was checked and did not state the term. 'Ambiguous' means relevant wording exists but its meaning is unresolved. 'Unreadable' identifies a source-quality problem. 'Conflicted' preserves inconsistent statements. 'Not reviewed' means no reviewer conclusion yet. Do not collapse these states into a blank, a zero, or 'not applicable.'

For each unresolved cell, record the question, its impact on the table, the responsible person, and the permitted next step. An unreadable annex may require a better copy; conflicting versions may require attorney review. Drafting a clarification question is different from sending it. No outreach should occur without human review and specific authorization.

Pilot with known answers and explicit stopping conditions

Begin with synthetic or appropriately authorized test documents containing known terms. Include a range, an unstated field, a condition, an attachment, and a structure mismatch. Have the reviewer prepare the expected extraction independently before comparing it with the AI output. Measure fidelity, omissions, reference accuracy, and treatment of unknowns, not which buyer the system favors.

Adopt clear pilot gates for this use case: no unresolved material transcription errors, no invented terms, working references for material cells, and reviewer confirmation of coverage. Stop AI-assisted use if the system repeatedly invents values, mixes parties or versions, cannot preserve conditions, follows instructions embedded in source documents, or exposes information beyond the approved scope. Use manual extraction while the cause is investigated.

These are proposed workflow criteria, not universal professional standards. Record the tested configuration and repeat relevant tests after changes to the model, extraction method, or document layout. NIST cautions against extrapolating capability from narrow, anecdotal assessments. A successful pilot supports a bounded use decision; it does not eliminate review of a live table. [2]

Record corrections and authorize the reviewed version

Keep a correction record showing the original cell, revised wording, source locator, reason, reviewer, and affected table version. Recheck dependent cells after a correction. If regeneration overwrites a reviewed term, treat the changed output as unreviewed. Before circulation, the responsible attorney should confirm the source register, unresolved flags, correction history, and intended recipients.

Describe the release accurately: verified extraction for the stated source set, with unresolved items visible. Do not label it a valuation opinion, legal approval of the proposals, or buyer recommendation. Keep the original documents available to authorized readers and retain materials under applicable firm and matter policies. A later IOI or clarification triggers review of affected fields, not automatic replacement or circulation.

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; it does not represent clients or provide legal advice. The validation process described here is a recommended framework, not a claim that the platform implements every control. Attorneys retain professional review, and authorized humans retain disclosure, negotiation, and transaction decisions.