What Human Review Checkpoints Should M&A Attorneys Set for AI-Assisted Work?

Assign an owner and a permitted use to every checkpoint

Start with a defined task, such as extracting assignment provisions from a selected contract set for a draft diligence schedule. Identify the supervising M&A lawyer, the person preparing the extraction, any specialist needed for interpretation and the lawyer authorized to release the resulting work. In a smaller practice, one lawyer may perform several roles, but the decisions should still be distinct. An instruction to 'have someone review it' leaves both the scope of review and the authority to proceed unclear.

Record what passing each gate permits. A fact-checked extraction may support an internal issue list without being ready for client advice or a buyer response. Likewise, an approved draft is not authority to send it to every recipient. ABA Opinion 512, issued July 29, 2024, addresses competence and supervision under the ABA Model Rules, rather than establishing a uniform rule for every jurisdiction. Counsel should check applicable professional rules and matter restrictions before implementing this workflow. [1]

Gate one: approve the inputs and confidentiality choices

Before upload, the matter lawyer should approve the document set and its purpose. Use a source manifest identifying the current version, executed status where relevant, associated amendments and missing attachments. Exclude unrelated matters and documents the team lacks authority to use. Have the lawyer review the prompt too: a prompt can disclose client information even when no document is attached. If the input set is incomplete, narrow the task instead of asking the model to fill gaps.

The firm’s technology or security lead should help the lawyer examine the proposed tool’s terms, retention, training use, access controls and sharing arrangements. Decide whether to exclude confidential material, use an appropriately checked redacted set, or permit limited information under reviewed arrangements. Redaction is not proof that a matter cannot be identified. Opinion 512 calls for a fact-specific confidentiality assessment and discusses informed consent for self-learning tools presenting disclosure risks. It does not make every AI use subject to identical consent requirements. [1]

Document any required client permission and applicable restrictions before processing. California’s State Bar guidance separately emphasizes scrutiny of confidentiality protections and client instructions; that guidance is California-specific. [2] Hold the input gate if permissions, terms or access boundaries cannot be established. Neither a lawyer’s approval nor a vendor assurance guarantees security or privilege.

Gate two: verify extracted facts against the actual records

For contract extraction, require each material entry to identify the document version, clause or page and supporting text. The preparing reviewer should compare the output with the original record, checking parties, dates, definitions, exceptions and cross-references. A supervising lawyer should resolve material discrepancies before the extraction supports a diligence conclusion. Keep 'not located in the reviewed set' distinct from 'does not exist.' Missing text, an unreadable scan or an absent schedule should remain an open issue.

As an illustrative check, a summary stating that a contract requires consent on a change of control should be tested against the entire relevant provision, definitions and amendments. The text might address assignment, contain an exception or depend on transaction structure. This example is a review problem, not a conclusion about enforceability. Do not turn an extracted phrase into a definitive consent requirement without lawyer analysis.

NIST’s generative-AI risk profile recommends reviewing and verifying sources and citations in both testing and ongoing monitoring. [3] Decide review coverage based on the task and consequences. Opinion 512 recognizes that prior testing on a smaller contract subset can inform verification, but leaves the lawyer responsible for the work. [1] Critical fields warrant direct source checks within this recommended framework; a fluent summary is not evidence.

Gate three: reserve legal characterization for qualified counsel

Separate a verified fact from its legal significance. An AI-produced issue list can flag a clause for review; it cannot decide how that clause affects the client’s proposed sale. The matter lawyer should assess governing law, transaction structure, the complete contractual record and client objectives. Refer unfamiliar specialist issues to counsel competent in that area rather than treating the model’s answer as a substitute for expertise.

When the output cites authority, open and read the actual statute, regulation, decision or official guidance. Check jurisdiction, effective date, current status and whether the proposition follows from the source. Review relevant contrary authority and distinguish binding law from commentary. A real citation can still support the wrong proposition. NIST identifies false citations and fabricated reasoning among generative-AI risks. [3] California guidance likewise stresses that scrutiny must extend beyond merely eliminating false results and that lawyer judgment cannot be delegated. [2] Record the lawyer’s reasoning, qualifications and remaining questions before a legal characterization moves into advice or drafting.

Gate four: review drafts against instructions and transaction documents

The drafting lawyer should supply a bounded assignment: prepare a draft using an approved precedent and verified facts, identify missing instructions and do not invent deal terms. Preserve the starting document and compare the output with it. Check defined terms, internal references, schedules and consistency with the current transaction documents. For a purchase-agreement draft, inspect operative obligations and their exceptions, not just whether the prose reads well.

The reviewing lawyer should test the draft against the client’s instructions and the agreed drafting position. Does it introduce a representation no one authorized, remove a qualification, assume a consent has been obtained or describe an open issue as settled? Mark every substantive change for disposition. Where tax, employment or another specialist issue falls outside the reviewer’s competence, obtain the appropriate review rather than letting polished wording conceal an unresolved question.

ABA Opinion 512 permits AI to serve as a foundation for legal work, but says lawyers cannot rely solely on it for functions requiring professional judgment, including client advice and negotiation. [1] In this workflow, AI prepares drafts only. The responsible lawyer adopts, revises or rejects them and conducts any authorized negotiation personally.

Keep corrections traceable before internal reliance

Use a correction record that links each material issue to the output version, source record, reviewer and disposition. Distinguish an incorrect extraction, a missing document, an unsupported legal proposition and a drafting departure. Preserve the reason for the correction, not merely a clean replacement. This lets the supervising lawyer determine whether a defect affects other summaries or drafts produced from the same inputs.

Before a team member relies on a diligence schedule or advice draft, confirm that corrections have reached the current version and that unresolved issues remain visible. If the AI used a superseded agreement, revisit downstream conclusions that depended on it. Label intermediate material as unreviewed, fact-checked or lawyer-approved for a specified use. These are recommended workflow labels, not legal certifications. The preparer should escalate disagreements to the supervising lawyer; the model must not decide which reviewer’s position prevails.

Gate five: authorize each external release separately

Assign the matter lawyer a separate release decision for client advice, buyer communications, data-room materials and other external work. Review the exact version, intended recipients, attachments and purpose. Check whether confidential information, internal analysis, comments or unresolved assumptions remain. Confirm any client approval and recipient-access restrictions required for that particular disclosure. A document approved for the client is not automatically approved for a counterparty.

Separate preparation permissions from sending permissions. During a trial, keep AI-assisted work in a restricted draft area and have an authorized human perform any approved sharing through the firm’s established process. If a court submission arises, counsel must separately check applicable court requirements. Opinion 512 addresses review of analysis and citations before submission and notes that local AI-disclosure requirements can vary. [1] Do not treat a 'reviewed' label as permission for autonomous outreach, disclosure or approval.

Trial the gates with explicit stopping and escalation conditions

Trial one bounded task before expanding its use. Select synthetic, public or otherwise properly authorized material resembling the intended document conditions. Have a qualified lawyer establish a reference result, then compare the AI output for omissions, unsupported conclusions, version confusion and inaccurate source references. Evaluate the reviewer’s ability to detect and correct defects, not just the model’s response. A successful test on clean contracts does not establish reliability for scans, unfamiliar document types or different legal questions. NIST cautions against extrapolating capabilities from narrow or anecdotal assessments. [3]

Set stopping conditions before the trial begins. Pause the task if confidential information appears outside the approved boundary, a material source cannot be traced, a critical provision is repeatedly omitted, or no competent reviewer is available. Route confidentiality concerns to the matter lawyer and designated security lead. Route legal uncertainty to appropriate counsel. Resume only after the responsible humans resolve the cause, review affected outputs and approve the limited next use. Reassess when the tool, settings, input types or task scope changes. NIST recommends ongoing monitoring and prompt escalation by people able to evaluate reported issues. [3]

Keep software operation separate from legal responsibility

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, provide legal advice or replace counsel’s judgment. The checkpoints described here are a recommended implementation framework, not claims that Acquivo supplies particular AI functions, integrations or audit controls. Counsel should establish the gates within the actual tools and permissions available. The practical outcome is a mapped chain of accountable human decisions, from input selection to a specifically approved use or sharing, with no AI authority to approve work or resolve legal questions.