How Should M&A Attorneys Use AI to Prepare Draft Disclosure Schedules for Lawyer Review?

Review authority and confidentiality before drafting

Assign a supervising transaction lawyer to authorize the drafting task, an intake owner to control documents, and seller personnel to verify facts within their responsibilities. Define permitted work narrowly: extract facts, suggest provision matches, and prepare candidate text for review. Exclude autonomous disclosure, negotiation, approval, and legal advice. Counsel should assess the engagement, applicable professional rules, and client instructions before any matter information enters the tool.

ABA Formal Opinion 512, issued in July 2024, addresses generative AI under the ABA Model Rules, rather than establishing a rule for every jurisdiction. It calls for a fact-specific assessment of confidentiality risks and understanding the tool's terms and privacy policies. Its informed-consent discussion concerns uses that can expose representation information through self-learning tools; generic engagement-letter authorization is insufficient when informed consent is required. Counsel must determine the applicable local requirements. [1]

Have the firm's technology reviewer examine access, retention, training use, deletion terms, and service-provider access for the proposed configuration. Limit inputs to approved matter records and authorized users. If permissions or confidentiality protections cannot be established, stop ingestion and use a manual process. Redaction is a possible control, not proof that information cannot be identified or that privilege is preserved.

Build the schedule map from the actual agreement

Freeze a working agreement version and identify its date or version label. Counsel should create a map of each provision requiring a schedule response, including the exact clause text, schedule number, relevant defined terms, entity scope, time period, and any knowledge or materiality qualifications. Separate affirmative lists requested by the agreement from possible exceptions to representations. A generic diligence checklist cannot substitute for this map.

An original disclosure schedule filed on SEC EDGAR illustrates why agreement wording matters: it adopts undefined terms from its merger agreement, qualifies the schedules by specific agreement provisions, and says headings do not amend the contractual descriptions. That historic exhibit is a transaction example, not SEC guidance or a template for a current deal. [2]

Have counsel confirm the map before AI drafting begins. If a provision is ambiguous, create a counsel question rather than asking the model to resolve the legal interpretation. Preserve any agreed disclosure instructions with the agreement version, so later reviewers can distinguish a drafting assumption from a negotiated requirement.

Create a source registry that reviewers can reopen

The intake owner should register approved records with stable document identifiers, versions, dates, and access restrictions. Sources may include executed contracts and amendments, corporate records, relevant accounting records, and seller responses. Record whether an item is signed, draft, incomplete, or superseded. Keep original documents available to authorized reviewers; extracted text is a working copy, not the source of truth.

Check text extraction before using it. Missing pages, unreadable scans, omitted attachments, and misplaced table columns can change what a document appears to say. Preserve a page, section, or cell locator for each excerpt and check it against the original. Link amendments to the base agreement and flag missing referenced documents rather than summarizing the base agreement as though it were complete.

Limit generation to the approved source set. Instructions should require an explicit missing-information response when support is absent. NIST's July 2024 Generative AI Profile warns that models can produce false content and fabricated supporting citations; it recommends reviewing and verifying sources and citations during testing and ongoing monitoring. A plausible document reference is therefore not enough. [3]

Draft candidates as reviewable records, not final disclosures

For each candidate, require a unique record identifier, agreement version, provision number, proposed schedule placement, neutral factual wording, and supporting source identifiers with excerpts and locators. Add separate fields for missing facts, conflicting records, proposed cross-references, assigned seller verifier, assigned lawyer, and review status. Keep factual extraction separate from the model's suggested reason for matching a provision.

Use statuses such as candidate, needs facts, factual check complete, counsel review pending, rejected, and superseded. No model-generated status should mean approved. If the system offers confidence scores, do not let them replace source checks or legal judgment. A highly confident match can still use the wrong entity, agreement version, or reporting period.

Distinguish 'no responsive evidence located in the reviewed sources' from 'none.' The former describes the search result; the latter may become a substantive schedule statement. Where the sources are incomplete, create a question for the responsible seller verifier. Do not fill an empty schedule with a negative assertion merely because retrieval returned nothing.

Route factual questions to the seller and legal questions to counsel

An illustrative candidate might identify a restriction in a customer contract and propose placement under a contracts provision. The seller's commercial owner should confirm the correct counterparty, whether the contract remains active, and whether amendments or side letters exist. Counsel should read the restriction in context and assess its relevance to the transaction structure and actual representation. The model should not conclude that consent is required or that disclosure cures a breach.

Ask targeted questions with source context: whether a referenced amendment has been located, whether a termination notice was received, or which entity holds the agreement. Record the responder's role, response date, and supporting record where available. A seller response that conflicts with a document should remain flagged until the responsible humans reconcile it.

Seller verification and lawyer review are distinct checkpoints. A factually correct sentence can still omit a contractually requested detail, while legally appropriate wording can describe an outdated fact. ABA guidance says lawyers may use AI as a foundation for drafting but cannot abdicate professional judgment or responsibility for their work. [1]

Propose cross-references without assuming cross-disclosure

Allow one source fact to generate multiple proposed provision matches, but give counsel control over schedule placement and cross-references. Show why each additional provision might be relevant and preserve the supporting text. Avoid duplicating slightly different descriptions of the same fact across schedules without a consistency check.

The filed schedule example conditions cross-disclosure on matters including whether applicability is apparent on the face of the disclosure and whether another schedule would become misleading. Those are that exhibit's terms, not a universal rule. [2] Counsel should inspect the current agreement's cross-disclosure language and assess its effect under the governing law. Do not assume a data-room upload, another schedule entry, or an AI-created link satisfies a particular disclosure requirement.

Preserve changes instead of silently refreshing the draft

When an agreement provision changes, identify affected candidates and mark them for renewed counsel review. Preserve the prior clause, new clause, prior candidate wording, and reason for the change. Renumbering requires checking both schedule labels and internal references. A revised definition or threshold may require reevaluating entries even when their source facts have not changed.

When a new source arrives, retain the earlier record and mark its relationship to the replacement. Reopen seller verification if the new document changes the factual statement. Compare versions for newly added assertions, deleted exceptions, and unresolved questions that disappeared. Never carry forward human approval merely because the model labels a revision minor. The approval should relate to the specific text and agreement version reviewed.

Run a bounded trial with explicit stopping conditions

Before broader use, counsel should choose a limited sample with straightforward records and harder cases such as amendments, missing attachments, conflicting facts, and facts relevant to multiple provisions. Manually establish expected findings, then compare the AI queue for omissions, unsupported additions, incorrect provision matches, and source-location errors. Review documents with no candidates as well as those with proposed entries.

Set stop conditions before the trial: invented facts or references, material meaning changed by extraction, unauthorized record access, or a missed item that makes the task unreliable for its intended use. Pause the affected workflow, investigate, and retest before expanding it. NIST cautions against extrapolating capabilities from narrow or anecdotal evaluations. A successful sample does not establish reliability across all schedule topics. [3]

Document the tested task, source conditions, reviewer findings, and limits. Repeat relevant checks when the agreement, input format, tool configuration, or drafting instructions change. Keep every proposed disclosure subject to human review even after the workflow passes its initial trial.

Deliver a human-controlled review queue

The implementation outcome is a source-linked draft schedule review queue for counsel and seller verification. Before any external delivery, counsel should review wording, completeness, cross-references, unresolved questions, and consistency with the current agreement. The authorized seller representative confirms the facts through the agreed process. Keep internal comments, retrieval instructions, and unapproved alternatives out of any deliverable. Delivery remains a separate authorized human action.

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 workflow described here is a recommended implementation framework, not a claim that Acquivo provides particular AI controls or integrations. Software does not represent clients or replace counsel's legal advice. Apply the workflow only after the responsible professionals assess the matter's requirements; no AI draft guarantees confidentiality, privilege, legal sufficiency, or a transaction outcome.