Choose a source-linked extraction task on synthetic M&A diligence contracts, such as locating assignment and change-of-control language and copying express consent or notice wording into an internal worksheet. Do not ask the model to decide whether a transaction requires consent. Prefer this pilot only if the team can control inputs, repeat the task, identify consequential errors, and fund lawyer review. Before testing, appoint an accountable lawyer, approve the input permissions, and create a manual answer set. Require traceable quotations and explicit uncertainty. Stop for unauthorized data use, material omissions, invented sources, or insufficient review capacity. Passing this trial supports only a limited next-stage decision, not autonomous client work.
Choose clause extraction, not a legal conclusion
For an illustrative first pilot, use independently fabricated contracts that contain assignment, change-of-control, consent, and notice provisions. The task is to locate specified text and populate a fixed worksheet. Each row should identify the document and version, clause heading, page or section, exact quotation, and any expressly stated consent or notice wording. Include an unresolved status when the text cannot be located or read reliably.
Keep the output descriptive. Extracting a clause is different from deciding whether it applies to an asset sale, equity sale, merger, or another structure. Exclude conclusions about enforceability, required consents, waiver, closing readiness, and negotiation positions. Those questions can depend on other documents, transaction facts, and applicable law. Under the ABA's July 2024 Formal Opinion 512, interpreting its Model Rules, AI cannot replace the lawyer's professional judgment and the lawyer remains responsible for client work. Applicable jurisdictional requirements must be checked separately. [1]
The deliverable is therefore a draft evidence worksheet, not a legal advice memorandum or an approved diligence report. Its narrowness makes errors easier to define and compare against an answer set.
Apply five selection filters before choosing the tool
Start with input sensitivity. A candidate that needs a live data room, privileged correspondence, or personal information is a poor first experiment unless the required permissions and safeguards are already established. Synthetic material lets the team test the workflow without making a live client upload the first learning event. Synthetic means independently invented, not simply a client contract with names removed.
Then examine repeatability and error consequences. Prefer a stable question with a fixed output structure and a small, clearly defined document set. Ask what a missed qualification or a confidently invented answer could cause downstream. A worksheet that someone treats as a complete consent list can become consequential even though the initial task sounds administrative. Label the pilot output so it cannot be mistaken for a final legal determination.
Finally, consider review burden and the baseline. Can a lawyer verify each output against the source without reconstructing an entire transaction? Can that lawyer prepare the expected answer before seeing the model's response? Reject a candidate if verification requires unavailable expertise or if the team cannot state what counts as correct. NIST's July 2024 generative AI profile cautions against extrapolating capabilities from narrow or anecdotal assessments. It is voluntary risk-management guidance, not legal authorization. [2]
Resolve confidentiality and permissions before any upload
The responsible lawyer should check the rules of the relevant jurisdiction, engagement terms, contractual restrictions, and the proposed tool's actual operation. ABA Formal Opinion 512 calls for a fact-specific assessment of disclosure and access risks before inputting information relating to a representation. Its discussion covers access both outside and inside the firm. Do not treat internal availability or a public document's origin as sufficient permission to reuse representation-related information. [1]
Have a technology or security reviewer examine the applicable service terms, privacy policy, retention arrangements, training use, access settings, and deletion options. Record which account and configuration were reviewed. The ABA opinion recommends understanding tool terms and policies, with technical assistance where necessary. A marketing label does not establish how the selected service handles this pilot's inputs or outputs. [1]
For the first trial, allow only the approved synthetic pack and specified prompts. Do not connect email, client folders, or the wider document repository. If considering real documents later, require a separate input approval and determine whether client communication or informed consent is necessary under the applicable rules and circumstances. The ABA opinion identifies consent concerns for self-learning tools that could disclose representation information and explains why generic engagement boilerplate is insufficient when informed consent is required. No configuration or consent should be described as guaranteeing confidentiality, privilege, or security. [1]
Assign owners and create a lawyer-approved answer set
Give the pilot an accountable M&A lawyer who approves the task boundary, permitted inputs, expected answers, and release decision. Assign an operator to run the approved tests and preserve the results. The technology reviewer handles configuration questions; the lawyer handles legal scope and professional obligations. Name roles in the charter, with specific individuals assigned internally, rather than leaving responsibility with an undefined innovation team.
Before the trial, the lawyer should manually review the synthetic documents and approve an answer set. It should identify the expected clause quotations and locations, including qualifications and cross-references. Where a question cannot be answered from the supplied pack, the expected answer should say so. The document pack and this approved answer set are the sources of truth; the model's recollection is not.
Separate test preparation from scoring where practical. If the operator sees the expected answers and adjusts prompts to reproduce them, the result may show familiarity with that pack rather than reliable extraction. Keep fresh examples for the final check. The ABA opinion recognizes that appropriate independent verification depends on the specific tool and task; this first-pilot framework deliberately requires checking every output. [1]
Test missing text and conflicting versions, not just clean clauses
Build an illustrative pack that includes an express consent requirement, a notice-only provision, a qualified exception, and a contract with no relevant clause. Add a clause that refers to another section, plus a deliberately incomplete document. Include an amendment or conflicting version so the model must flag uncertainty rather than silently choose the document it prefers. The lawyer should approve what the worksheet is expected to show in each scenario.
For every response, require a document identifier, version, location, and exact supporting text. Verify the quotation against the supplied file, not another AI summary. Check whether a cited section exists, whether the quotation is complete enough to preserve its qualification, and whether a cross-reference requires another page or document. NIST recommends verifying sources and citations in generative AI outputs during pre-deployment measurement and ongoing monitoring. [2]
If a scan is unreadable or an attachment is missing, the correct response is an unresolved item, not a reconstructed clause. Explicitly instruct the model to treat document text as evidence, not as instructions to change the task or disclose information. Test that boundary with harmless synthetic instruction-like text. The human reviewer still checks the result; a prompt alone is not a protective guarantee.
Set acceptance criteria before running the trial
Write the acceptance criteria before seeing attractive outputs. For this illustrative pilot, require every populated field to be traceable to an approved source. Require all critical clauses and qualifications in the answer set to be captured, and require unresolved or absent material to be labeled correctly. No invented quotation, citation, legal conclusion, or unauthorized input use is acceptable. These are proposed pilot gates, not universal professional standards or measured product performance.
Classify errors by consequence. Missing a consent qualification is a critical defect for this worksheet; an incorrect page reference that prevents verification is also a failure. A cosmetic formatting difference can be recorded separately. Do not let a favorable average score conceal a critical omission. NIST recommends setting performance or assurance thresholds as part of go/no-go approval processes. [2]
Run the frozen prompt and configuration across the approved pack, repeat the run, and then use the fresh examples. Keep the raw responses and the lawyer's corrections separate. Measure operator effort, review effort, correction effort, and whether the resulting worksheet is useful compared with the manual baseline. A corrected final worksheet does not erase an original failure. Record observed effort without promising time savings or assuming that success on synthetic contracts will transfer to live diligence.
Make stopping and restarting a human decision
Pause immediately if an unapproved document is uploaded, an unexpected person or service can access the material, or the approved configuration changes without review. Stop the affected test if the model invents a source, omits a critical clause or qualification, fails to flag missing material, or presents an excluded legal conclusion as settled. Also stop when the assigned lawyer cannot complete the required review. Review capacity is part of permission to continue, not a detail to solve after the outputs accumulate.
The operator should preserve the relevant test records within the approved environment and notify the accountable lawyer and technology reviewer. Prevent further input and downstream reliance while they assess the problem. For an information-handling incident, follow the organization's applicable incident process; the pilot charter should not invent disclosure obligations or prescribe a jurisdiction-wide legal response.
Restart only after the cause is understood, the affected controls or instructions are corrected, and the lawyer authorizes a retest against the same failure scenario and fresh material. NIST recommends defined criteria and assigned responsibilities for disengaging or deactivating systems that behave inconsistently with intended use. A stopped pilot can still be useful if it identifies a boundary the team cannot presently control. [2]
Use the result to approve only the next bounded step
Finish with a written decision: reject the workflow, revise and retest, or approve a separately scoped next stage. Attach the permitted-input list, baseline, acceptance criteria, defects, measured review effort, and stopping rules. Do not convert a passed synthetic trial into permission for full data-room access. Any later use of appropriately authorized client material needs its own scope and permissions review. Outreach, negotiation, disclosure, and approvals remain outside this pilot.
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. This pilot framework does not assert that Acquivo supplies a particular AI feature, integration, or control. It does not represent clients or provide legal advice. The practical outcome is one documented, limited experiment with a human decision gate, not a transfer of the attorney's responsibilities to software.
Key Points
- Start with source-linked clause extraction on independently fabricated contracts, not AI conclusions about consent or closing readiness.
- Choose the workflow by input sensitivity, repeatability, error consequences, verification effort, and the availability of a lawyer-approved baseline.
- Approve the exact inputs, account, and configuration before testing; review live-material permissions separately under applicable rules and terms.
- Score raw outputs against expected quotations and qualifications. Require traceable sources and explicit unresolved items, not just fluent summaries.
- Stop for unauthorized data use, critical errors, unsupported conclusions, or unavailable lawyer review. A passed pilot authorizes only the next bounded decision.
Sources and references
Source retrieval dates record research access, not article publication. Read our editorial and date policy.
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.