How Can M&A Attorneys Use AI to Identify Conflicting Answers Across Diligence Materials?

Define a consistency task and assign its owners

The task is to identify statements that cannot readily be reconciled on their face, not to answer bidder questions or reconcile financial numbers. Define the review scope around specific legal and factual subjects, such as disputes, contract restrictions, ownership of intellectual property, or required permissions. Set the transaction's review date and name the entities covered. This prevents a broad prompt from treating every difference in wording as a defect.

Assign a supervising M&A attorney to approve the scope and assess legal significance. A diligence coordinator maintains the document inventory and queue. Business record owners verify facts within their responsibilities, while a reviewing attorney determines whether the evidence supports closure or further inquiry. An authorized person controls any eventual release of corrected material. These are recommended workflow roles, not a transfer of legal judgment to software. ABA Formal Opinion 512 discusses supervision and makes clear that AI does not replace the lawyer's responsibility for professional work. [1]

Clear permissions before entering transaction material

Before a real trial, have responsible counsel assess the governing jurisdiction's professional rules, client instructions, confidentiality obligations, and document-use restrictions. ABA Formal Opinion 512 is guidance based on the ABA Model Rules, not a blanket rule enacted identically everywhere. It describes a fact-specific assessment of access risks and calls for understanding the tool's terms, privacy policy, and related contractual terms. Its informed-consent discussion depends on the tools and uses described in the opinion. [1]

Translate that assessment into an approved input boundary. With appropriate technical help, check actual retention settings, use of inputs for training, service-provider access, matter separation, and who can see extracted text or queue entries. A document's existing data-room access does not itself settle whether a separate AI use is authorized. Use synthetic material first if the assessment is incomplete. Omit or redact material only where counsel approves that approach; redaction alone should not be treated as proof that disclosure risks are resolved.

Do not place privileged analysis or restricted personal information into the tool simply because it might explain a discrepancy. Keep those materials in their approved environment and route the issue to a reviewer with appropriate access. Neither an AI label nor this workflow guarantees privilege or security.

Build a versioned source inventory before comparing answers

Preserve the originals and give every approved document a stable identifier. Record the file version, document date, effective date where relevant, upload date, entity covered, and status: executed, draft, superseded, or unverified. Distinguish when a statement was made from the period it describes. A recently uploaded questionnaire can still report an earlier state of affairs. Identify who supplied or approved an answer without assuming that approval makes the underlying statement true.

For each extracted statement, retain the question or clause heading and enough surrounding text to preserve qualifications. Store a page, paragraph, question number, or other locator that opens the original passage. Check scanned text against the page image when recognition errors could change meaning, especially for negation and exceptions. Missing schedules, exhibits, amendments, or unreadable pages should produce an evidence-gap entry, not a reassuring consistency conclusion.

Treat original documents and verified business records as the evidence base, with counsel determining their significance. AI summaries are derivative aids, not replacement records. NIST's generative AI profile recommends grounding retrieval data and checking output sources and citations. This inventory applies those recommendations to a diligence comparison task. [2]

Require paired excerpts instead of an AI reconciliation

Give the tool a bounded instruction: compare only the approved sources for the named subject and return candidate inconsistencies with two exact excerpts and their locators. Require it to identify the entity, relevant time, and qualification attached to each statement. Ask it to label missing evidence explicitly and to distinguish copied source text from its own explanation. Do not ask it to select the true statement or infer legal consequences.

For an illustrative comparison, a questionnaire might state that no written dispute notices have been received, while an approved document contains such a notice. The useful output is the source pair and the unresolved question about scope and timing. It is not an allegation that someone gave a false answer. A reviewer still needs to establish whether both statements concern the same entity, subject, and period.

Open both originals before accepting a candidate into the reviewed queue. Verify that the quotations exist, the locators work, and surrounding language has not been omitted. NIST warns that generated explanations and citations can themselves be false, so a plausible rationale is not verification. [2]

Separate contradictions from time, scope, and interpretation differences

Use provisional classifications that guide review rather than announce a legal conclusion. A potential factual contradiction concerns the same proposition, entity, and period but contains incompatible statements. A time difference may reflect a resolved dispute or a later amendment. A scope difference may arise because one answer covers the parent company and another covers an affiliate. A qualification difference may turn on words such as written, material, or to the respondent's knowledge.

For an illustrative historical difference, an earlier questionnaire may describe an unresolved claim while a later answer reports none outstanding. A dated settlement record might explain the change, but counsel and the relevant record owner must verify that explanation. Do not simply discard the earlier answer because it is older. Preserve it and record why the later position is supported, including any remaining uncertainty.

Legal interpretation belongs in a separate classification. For example, an answer about assignment restrictions and a clause about change-of-control consent may raise different questions rather than contradict each other. The reviewer should inspect the complete contract, relevant amendments, transaction structure, and applicable law before determining the implications. This article does not decide whether any particular consent is needed. Where evidence is insufficient, keep the issue open instead of having AI invent a bridge between statements.

Make each queue entry actionable and permission-aware

A useful entry contains an issue identifier, subject, paired source identifiers, exact excerpts, locators, versions, and relevant dates. Add the provisional classification, a short explanation of the difference, the evidence still needed, an issue owner, a reviewing attorney, and a review deadline tied to the transaction's actual needs. Keep the AI suggestion separate from the human assessment so later readers can tell which is which.

Use clear states such as awaiting source verification, awaiting factual confirmation, awaiting legal review, resolved with correction, resolved without correction, and unresolved. Severity should be a reviewer-assigned assessment of the issue's relevance, not a model's unsupported risk score. Prevent duplicate pairs from obscuring separate underlying questions. Preserve access restrictions when sharing entries: copying a restricted excerpt into a broadly visible queue can expose the very information the document controls were intended to limit. ABA's confidentiality discussion includes risks from inappropriate access inside a firm as well as outside it. [1]

Resolve issues through records, not persuasive generated wording

The issue owner should return evidence rather than a bare assertion that the discrepancy is harmless. Depending on the subject, that could be the signed amendment, dated correspondence, approved register, or other underlying record. The coordinator attaches its identifier and locator. The reviewing attorney then records whether the pair was a true conflict, a supported time or scope difference, an extraction error, or still unresolved.

A closure record should identify the supporting evidence, rationale, reviewer role, decision date, and affected document versions. If a correction is required, track the revised passage and the related questionnaire or draft answer that must also change. Retain the superseded version rather than silently replacing the history. Reopen dependent entries when new evidence changes their basis.

Keep factual confirmation, legal review, and authority to disclose distinct. The software should not contact a buyer, negotiate an explanation, approve a disclosure, or release a revised answer. The supervising attorney and authorized transaction participants retain those decisions. ABA guidance expressly rejects reliance on AI alone for functions requiring a lawyer's personal judgment or participation. [1]

Pilot against manual review and specify stopping conditions

Start with a permission-cleared set that resembles the intended work, including questionnaires, amendments, draft answers, and difficult scans where applicable. Have a qualified reviewer independently identify conflicts and explain non-conflicts before comparing the tool's results. Include examples of entity differences, changed dates, exceptions, missing exhibits, and ordinary paraphrases. Measure both missed issues and false flags, alongside quotation accuracy and correct source pairing.

Set acceptance criteria before seeing results. Pause the trial if an excerpt is fabricated, the tool retrieves excluded material, a known consequential conflict is missed, or source locators cannot be reproduced. Suspend use if permissions or provider terms become uncertain. Repeated false flags should trigger a narrower scope or revised instructions rather than pressure on reviewers to close entries without investigation.

Document the failure, correct the affected process, and retest before restarting. Reassess after changes to the model, extraction method, instructions, permissions, or document mix. NIST recommends ongoing source verification and warns against extrapolating capability from narrow anecdotal tests. A successful trial supports only the tested task and conditions, not a claim that the full diligence record is complete or accurate. [2]

Keep the reviewed queue separate from the released record

Before relying on a revised document, the coordinator should check that affected passages were updated and the reviewing attorney should confirm that unresolved issues remain visible. An authorized human approves any release through the transaction's established process. The result is a source-paired inconsistency queue with accountable resolution and review records, not an AI certificate that diligence contains no conflicts.

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 queue described here is a recommended implementation framework, not an assertion that Acquivo provides a particular AI, integration, or audit feature. Attorneys supply professional judgment, and authorized users operate the process. Acquivo does not represent clients or provide legal advice.