M&A attorneys should train lawyers and staff through supervised exercises tied to the deal tasks each person may perform. Start with invented training documents, an approved tool, and explicit input restrictions. Require participants to trace outputs to original records, identify omissions, refuse to fill missing facts, and escalate uncertainty. Give staff document-handling exercises and lawyers substantive verification exercises, with an accountable lawyer reviewing work before use. Tool access or course attendance alone should not establish readiness. ABA guidance connects supervision with practical and ethical AI training; applicable duties depend on governing rules. [1] Use a limited pilot with defined permissions, stopping conditions, and human approval before any external disclosure.
Assign training owners before assigning AI tasks
The matter lead should define the permitted task, its source records, and the lawyer accountable for the resulting work. A supervising lawyer should teach and assess the exercises. A technical owner should explain the actual tool configuration and access controls. Where the firm lacks specialist expertise, it should seek appropriate technical or professional guidance rather than asking trainees to assess unfamiliar risks alone. ABA Formal Opinion 512 discusses both technological understanding and supervisory training under the ABA Model Rules. It is guidance interpreting those rules, not a uniform rule for every jurisdiction. [1]
Before live-matter training, have qualified ethics counsel review the applicable professional rules, client restrictions, engagement terms, and any disclosure or informed-consent requirements. Those questions depend on the tool, information, task, and governing jurisdiction. [1] Define separate responsibilities for staff preparing documents, lawyers analyzing provisions, and the matter lead approving use. A person permitted to create an internal index is not thereby permitted to interpret a contract or send its contents to a buyer.
Teach permitted inputs through a classification exercise
Give each participant an illustrative packet containing an invented contract excerpt, an internal-style training note, and a sample access restriction. Ask them to decide what may enter the approved tool, what needs escalation, and what should stay outside it. The correct answer should identify the permission basis, not just remove a name. In actual matters, a business description or combination of details may still expose sensitive information. ABA guidance calls for a fact-specific assessment of disclosure and access risks before submitting representation information. [1]
The technical owner should demonstrate the approved account, accessible data sources, retention settings, sharing controls, and any enabled connections. Unresolved questions about input reuse, third-party access, or deletion should go to the designated reviewer before live use. California's 2026 guidance emphasizes understanding data handling rather than relying on generalized marketing assurances; that guidance addresses California professional obligations. [2] Teach three practical choices: use invented material, use a specifically approved limited input, or do not use AI for the task. Neither a paid account nor anonymization alone proves that an input is permitted.
Give staff a document-indexing exercise with narrow permissions
For administrative staff and paralegals, start with an invented document set and an authoritative training index maintained by the supervisor. Ask participants to draft an index showing document identifier, version, date as displayed, and whether referenced attachments are present. The task is organization, not a conclusion about legal sufficiency. Require comparison against the actual files, including unreadable pages and duplicate versions, rather than treating the AI-generated index as the source of truth.
Build in a missing attachment and a document with an ambiguous date. A satisfactory response identifies the gap and routes it to the supervising lawyer. It does not infer a signature, declare a document complete, or overwrite the original index. Staff should practice working only within authorized folders and keeping outputs in the designated review location. California's current guidance specifically addresses training and supervision of lawyers and nonlawyers using AI-enabled tools. [2] The exercise translates that concern into observable document-handling behavior without expanding anyone's professional role.
Train lawyers to verify contract summaries against complete sources
Give lawyers an illustrative contract-extraction task, such as identifying language about assignment and consent. Require the document identifier, version, section, exact supporting passage, and any relevant definitions or exceptions. The reviewing lawyer should compare the output with the complete approved source, including amendments and schedules, before accepting it. Finding a sentence that resembles the summary is not enough if another provision changes its meaning.
Ask the trainee to separate extracted text, a tentative interpretation, and an unresolved legal question. Whether a clause affects a particular transaction remains a question for qualified counsel considering the documents, transaction structure, and applicable law. AI output should not substitute for that judgment. ABA Opinion 512 discusses independent verification appropriate to the task and tool, with responsibility remaining with the lawyer. [1] For the supervised exercise, require review of every extracted claim so the trainer can observe the participant's method, not merely the final answer.
Make missing facts and conflicting versions part of the test
An exercise should reward stopping when the record is insufficient. Ask a trainee to prepare a draft response about a consent requirement from an intentionally incomplete mock contract. The prompt can instruct the tool to identify supporting text and leave unsupported facts unresolved. The participant must still check the response. If the relevant schedule is absent, the useful result is a documented question for counsel, not a confident assertion that no consent is required.
Run a second exercise with conflicting mock versions. The participant should identify the conflict, preserve both sources, and ask the document owner which version controls. Include an inaccurate section reference in a draft output and require the trainee to detect it. NIST's generative AI profile recommends verifying sources and citations during testing and ongoing monitoring. [3] If the work includes legal research, lawyers should open the underlying authority, confirm jurisdiction and relevant dates, and check whether it supports the proposition. A plausible citation or the model's own reassurance is not verification. [1]
Practice output handling and the human release gate
Teach participants to distinguish an AI-assisted draft from reviewed work. Keep the original records unchanged. Record the sources used, unresolved questions, corrections, and responsible reviewer in the firm's approved working record. Label unreviewed material clearly so another team member does not mistake it for an approved analysis. ABA Opinion 512 suggests identifying AI-produced stored materials so future users understand their potential fallibility. [1] Choose a record format the team can actually maintain without assuming any particular software provides it.
Use an illustrative buyer-question exercise: a draft answer contains a supported statement and a fact not found in the packet. Staff should flag the unsupported statement, not send the draft. The supervising lawyer reviews substance and disclosure permissions; the authorized human sender checks the intended recipient and attachments. Keep outreach, negotiation, approvals, and external disclosure outside autonomous AI operation. California's 2026 guidance highlights the risks of broad system access and autonomous external transmission in the context of California duties. [2] This training framework uses a deliberate human release gate regardless of whether a tool offers automated sending.
Assess the person separately from the tool
A readiness check should observe a participant performing the permitted task under supervision. Ask them to explain why the input is allowed, locate the source supporting each material statement, identify gaps, and demonstrate escalation. Assess judgment and handling as well as correctness. A polished output produced by chance does not show that the user can catch the next error. NIST distinguishes human proficiency testing from testing AI capabilities. [3]
Keep the participant's authorization narrow: identify the task, tool configuration, permitted input category, reviewer, and restrictions. A lawyer who performs contract extraction successfully may still need separate practice for legal research. A staff member who handles document indexing appropriately has not been authorized to interpret legal obligations. When an exercise exposes a serious handling failure, provide targeted instruction and repeat the affected exercise before allowing that task. Record attendance separately from observed readiness. Neither the record nor a completed exercise should be presented as a certification of competence or compliance.
Run a supervised pilot with explicit stopping conditions
After the training exercises, the matter lead can consider a restricted live pilot only if the specific inputs and use have been approved. Choose a defined task, such as drafting an internal document index from permitted records. The supervising lawyer should review the pilot output against those records before it affects the matter. Track omissions, unsupported claims, source mismatches, and handling failures separately from stylistic edits. Do not extrapolate success on a narrow test to every deal workflow; NIST expressly cautions against broad capability claims from narrow or anecdotal assessments. [3]
Stop the pilot if someone submits unapproved information, accesses an unauthorized matter, cannot reproduce a material source, bypasses the reviewer, or encounters unresolved data-handling changes. Preserve the relevant working records under firm policy and notify the matter lead and technical owner. They should determine containment, correction, and any further professional review before restarting. NIST recommends defined go/no-go thresholds and a plan to halt deployment presenting unacceptable risk. [3] The purpose is controlled learning, not a promised reduction in review work or a guarantee of accurate output.
Refresh training when the workflow changes
Repeat relevant exercises when the tool's material settings, connected sources, permitted tasks, or client restrictions change. A new ability to access additional folders or transmit material deserves separate review, not automatic acceptance because the team recognizes the interface. California's 2026 guidance calls for periodic reassessment and updated training as AI systems and uses evolve. [2] The matter lead should keep permissions current and give staff a clear route to pause work without improvising a workaround.
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. It is user-operated software with no success fee. The exercises described here are a recommended training framework, not a claim that Acquivo supplies particular AI controls, integrations, or certifications. Acquivo does not represent clients or provide legal advice. Lawyers remain responsible for their professional work, and users retain decisions and external actions.
Key Points
- Assign a matter lead, supervising lawyer, and technical owner before trainees handle live deal information.
- Use invented documents to practice input permissions, missing-fact recognition, and escalation without exposing a client's matter.
- Give staff document-handling exercises and lawyers substantive source-verification exercises within their respective responsibilities.
- Assess observed judgment separately from tool performance, course attendance, or access to an account.
- Keep a human release gate for external communications and pause the pilot when permissions, source reliability, or review controls fail.
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] State Bar of California: Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law, 2026 revision · Accessed
- [3] NIST AI 600-1: Artificial Intelligence Risk Management Framework, Generative Artificial Intelligence Profile, July 2024 · 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.