Measure an AI-assisted M&A task as a controlled comparison, not as a claim that the tool is faster or better. Define one repeatable task, preserve a comparable baseline, and record preparation time, tool expense, reviewer time, corrections, omissions, escalations, and later rework. Assign a human reviewer and use the matter’s approved source-of-truth records. Record which information was permitted in the tool and which was withheld. Compare quality and total observed effort, not just drafting time. Set a stop rule for material errors, confidentiality uncertainty, or review burden that exceeds the pilot’s tolerance. Professional duties and client-consent requirements vary by jurisdiction, so counsel should apply the applicable rules before expanding use. [1]
Start with a narrow M&A task and a fair baseline
A useful pilot begins with a task that can be described without ambiguity. Examples include extracting defined terms from a draft purchase agreement, organizing diligence requests by topic, preparing a first-pass issue list from supplied documents, or comparing two versions of a disclosure schedule. The task owner should state the expected input set, output format, reviewer, and completion condition before anyone uses the tool.
The comparison should be against a similar piece of work performed without the AI step, or against a deliberately repeated sample using the same instructions and source set. Avoid comparing a clean, familiar matter with a novel, document-heavy matter. Record differences such as document volume, urgency, deal stage, language, and reviewer seniority. A baseline is useful only when its limits are visible.
Do not make the baseline a race. If the conventional process normally includes a second-person check, preserve that check in the comparison. Otherwise the pilot may appear cheaper only because necessary review has been omitted. The objective is to observe the complete work path that the firm would actually accept.
Define the cost model before collecting numbers
Matter cost is broader than a subscription invoice or a model’s usage charge. Create fields for preparation time, prompt or instruction design, file preparation, tool charges, export or transfer work, primary review, secondary review, corrections, follow-up research, and rework after delivery. If internal billing conventions matter, keep them separate from operational time so an accounting choice does not hide effort.
Use actual observed amounts where available and label estimates clearly. A simple record might show minutes by role, direct tool expense, number of source documents, output length, and the point at which the work was accepted. Do not convert every minute into a claimed financial saving. Different matters may have different value, urgency, staffing, and fee arrangements.
The ABA’s Formal Opinion 512 discusses the relationship between generative AI use, reasonable fees, and time spent on input and review. It also distinguishes compensable work from time spent learning how to operate a tool in many circumstances. That guidance is tied to the ABA Model Rules and does not replace the rules, opinions, engagement terms, or client instructions applicable to a particular matter or jurisdiction. [1]
Count reviewer effort as part of the work
The accountable reviewer should log time spent checking citations, definitions, calculations, cross-references, missing documents, extracted facts, and changes from the source record. Record whether the reviewer accepted the output, edited it, rejected it, or sent it back for another pass. A short generated draft can require substantial checking when the consequences of an omission are high.
Separate review categories rather than using one quality score. For example, mark factual mismatch, unsupported inference, omitted item, formatting defect, confidentiality issue, and workflow error. This gives the team a way to see whether rework comes from the model, the prompt, the source set, or an unclear acceptance standard.
The reviewer should have enough subject-matter knowledge and matter context to make the acceptance decision. AI output is not the source of truth. The approved agreement, diligence tracker, board materials, client instruction, or other designated record remains the reference against which the output is checked.
Measure rework and omissions, not only first-pass speed
Define rework as an additional human action required because the first output was incomplete, incorrect, unusable, or insufficiently supported. Log the trigger, the person who found it, the corrective action, and the extra time. Include rework discovered after handoff when the pilot can observe it. A correction made during the first review and a correction found during signing preparation may have very different consequences, even if both are one line in a tracker.
For an extraction task, test a known sample of source items and record false inclusions, false exclusions, and unresolved items. For a comparison task, reconcile the output to the approved redline or version-control record. For a drafting task, check defined terms, party names, dates, monetary figures, conditions, schedules, and references to governing documents. These checks should be selected by the task owner before the pilot begins.
Do not claim that a low observed error count proves reliability. The sample may be too small, the source set may be unusually clean, or the reviewer may have corrected problems before they were recorded. Report the task scope, sample, review depth, and unresolved uncertainty alongside any result.
Protect matter information before the pilot starts
Create a data classification decision before uploading anything. Identify whether the workflow uses public information, anonymized examples, internal work product, client confidential information, personal data, privileged material, or transaction terms that should remain outside the tool. Decide which fields can be transformed, redacted, or replaced with placeholders and who approves that choice.
Record the tool setting, account owner, retention terms reviewed by the firm, access permissions, export path, and deletion or shutdown process that the team can actually verify. Do not treat a vendor statement, a setting, or a contract label as a guarantee of privilege, confidentiality, security, or compliance. If the team cannot establish an acceptable information path, stop the pilot or use a non-sensitive test set.
The ABA identifies confidentiality, competence, and client communication as relevant considerations when lawyers use generative AI. The exact professional obligation depends on applicable law, ethics rules, client instructions, and the facts of the engagement. Counsel should make that assessment before using client information, rather than trying to repair the decision after an upload. [1]
Assign owners, permissions, and stopping conditions
Name a task owner who defines the use case, a human reviewer who accepts or rejects output, and a records owner who preserves the pilot log. Permissions should follow the minimum access needed for the task. A user who can run an extraction does not automatically have authority to disclose information, send a client communication, change a signed record, approve a transaction step, or negotiate a term.
Write stopping conditions in advance. Examples include an unresolvable confidentiality question, an omitted item in a high-consequence field, an unsupported citation, repeated reviewer escalation, unexplained output variation, or review effort that exceeds the team’s agreed tolerance. A stop rule is not a failure of the pilot. It is evidence that the workflow needs a different source set, instruction, reviewer, tool setting, or decision not to expand.
Keep the pilot advisory and reversible. AI may prepare a candidate list or draft for review, but a human remains responsible for legal judgment, client communication, negotiation, approval, and disclosure. The workflow should make the handoff explicit instead of allowing an unreviewed output to become a matter record.
Use a measurement sheet that supports an audit trail
A practical measurement sheet can have one row per task instance. Include matter or test-set identifier, task definition, source-set description, baseline method, AI method, user, reviewer, timestamps, preparation minutes, tool expense, review minutes, correction categories, omissions, escalations, final disposition, and notes on confidentiality controls. Store the approved source reference separately from the generated output so the comparison can be reconstructed.
NIST describes its AI Risk Management Framework as a voluntary resource for incorporating trustworthiness considerations into the design, development, use, and evaluation of AI systems. Its generative-AI profile is intended to help organizations identify distinctive risks and select risk-management actions aligned with their goals. For a legal pilot, that supports a disciplined cycle of defining the use, mapping risks, measuring observed results, and managing the decision. It is a framework, not a legal safe harbor or a required certification. [2] [3]
Keep raw observations and conclusions separate. The log can say that a reviewer spent a particular amount of time on a particular test set. The decision memo can then explain whether the observed tradeoff justifies another controlled trial. Avoid converting a single matter’s result into a firm-wide promise or a conclusion about legal outcomes.
Decide whether expansion is justified
Expansion should require more than a favorable first-pass impression. Review the complete cost path, error and omission pattern, reviewer workload, data controls, and fit with the task’s risk. Compare like with like and preserve the reasons for excluding incomparable observations. If the tool reduces preparation time but increases checking time, report both. If it creates a new correction category, treat that as a workflow finding rather than hiding it in an average.
A sensible next step may be a narrower use case, a larger controlled sample, a different permission model, or no expansion. Ask whether the team can repeat the controls with ordinary matter pressure. If measurement depends on one unusually experienced user or an informal review habit, the result may not transfer. If the team cannot identify an accountable reviewer or a reliable source of truth, expansion should wait.
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, so the measurement plan should identify the human who controls each review, permission, and matter decision rather than assuming the platform replaces professional judgment.
Key Points
- Compare a clearly defined AI-assisted task with a genuinely comparable baseline, including the review steps the team would ordinarily perform.
- Track preparation, direct tool expense, reviewer effort, corrections, omissions, escalations, and later rework as separate observations.
- Use approved source-of-truth records and assign a qualified human reviewer who accepts, edits, rejects, or escalates the output.
- Decide in advance what information may enter the tool, who approves that choice, and what retention and access facts can actually be verified.
- Stop the pilot for material errors, confidentiality uncertainty, unsupported output, or review burden that exceeds the agreed tolerance.
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.