AI can help an associate turn a messy evidence tracker into a focused draft questionnaire, but it should not decide what management must confirm or send questions on its own. Start with a banker-approved source set, classify each missing or inconsistent point, and ask the system to propose one question tied to a specific evidence reference. A banker then checks the source, wording, respondent, confidentiality treatment, and whether the question is already answered elsewhere. The controlled output is a reviewed questionnaire that preserves existing answers, flags uncertainty, assigns management owners, and records why each question remains open. Treat generated text as a preparation aid, not as evidence, advice, an approval, or a substitute for management input.
1. Define the evidence gap before asking for a question
A useful questionnaire begins with a gap register, not a blank prompt. For each issue, record the source document or workbook location, the relevant period, the fact that is present, the fact that is missing or inconsistent, and the consequence for the sell-side workstream. Typical categories include an unreconciled revenue total, a customer concentration figure that does not match the schedule, an operating metric with no definition, or an expense line that lacks supporting detail. The record should distinguish missing evidence from evidence that exists but has not yet been located.
Use stable labels such as financial reconciliation, operating metric, ownership, customer, employee, contract, or tax item. These labels help the reviewer see whether a proposed question addresses a real gap or merely restates a request already answered. They also make it easier to route a question to the right management respondent without turning the model into an authority on the transaction.
2. Build a controlled source-of-truth packet
Before using an AI tool, decide which records are authoritative for the draft. A practical packet may include the current due-diligence request list, the latest banker-approved financial schedules, the CIM or teaser version being used, prior management answers, and a change log. Preserve document names, versions, dates, page or tab references, and any known limitations. Do not let an unverified summary silently replace a financial statement or an approved answer.
The question-generation instruction should require a source citation for every proposed question. If the system cannot point to the missing line, conflicting values, or unanswered request, it should return a review flag rather than fill the gap from general knowledge. This is consistent with NIST's emphasis on data provenance, defined responsibilities, and documented risk management for AI systems [1].
3. Ask for questions, not conclusions
The task should be narrowly framed: identify unresolved evidence, preserve the answer already on file, and draft a neutral question for management. The system can propose wording such as, 'The monthly revenue schedule totals differently from the financial summary for the stated period. Which schedule should be treated as the controlling figure, and can you provide the reconciliation?' It should not infer which number is correct, state that a representation is inaccurate, or convert an unresolved issue into a valuation conclusion.
Require each proposed item to contain five fields: evidence reference, gap description, draft question, suggested respondent, and reviewer note. A sixth field can record the expected supporting document or explanation. Asking for this structure reduces the risk that polished prose hides an unsupported premise. It also makes a questionnaire usable by a banker who needs to approve each item quickly and trace it back to the working file.
4. Preserve existing answers and detect duplicates
The first comparison should be against answers already received. An answer may be complete, partial, conditional, or inconsistent with another record. The workflow should label those states instead of treating every imperfect answer as a new question. If a prior answer addresses the issue, retain it and propose only a clarification when a specific unresolved point remains.
Duplicate detection matters because repeated requests create avoidable friction with management and make the request list harder to control. Compare the proposed question with the request-list wording, prior correspondence, data-room filenames, and the latest answer log. A human should decide whether two differently worded items are genuinely distinct. The model may highlight likely duplicates, but it should not close or delete a request without approval.
5. Assign the respondent by knowledge, not convenience
A targeted question names the person or function most likely to answer it, while leaving the banker in control of the assignment. Revenue recognition or a financial reconciliation may belong with the finance lead. A customer contract question may require the commercial owner or counsel. A headcount or incentive question may require the people lead. The appropriate respondent depends on the business and process, so the suggested owner is a routing aid rather than a statement that the person has authority to make a representation.
Keep the question neutral and request the underlying support where needed. Avoid asking management to agree with an AI-generated interpretation. A better formulation asks management to identify the controlling record, explain a variance, define a metric, or provide the missing schedule. The banker can then decide whether the response is sufficient, whether a follow-up is needed, and whether counsel or another specialist should review the issue.
6. Set confidentiality and permission controls before the pilot
A bank should establish what may be entered into the selected tool before a pilot begins. Review the tool's terms, retention settings, access controls, and use of submitted content. Apply the firm's confidentiality policy and any client, engagement, data-room, or contractual restrictions. If an attorney is involved, legal-ethics guidance may impose additional, fact-specific duties. ABA Formal Opinion 512 explains that lawyers must assess disclosure and access risks and may need informed consent before entering client-related information into some generative AI tools [2]. That guidance should not be treated as a universal rule for every banker or jurisdiction, but it is a useful reason to involve qualified counsel where applicable.
Use the least sensitive workable input during an initial test. A redacted gap description may be enough to test classification and question structure. Do not assume that a private interface, a vendor promise, or a firm's internal access setting automatically creates privilege, confidentiality, or security. Record the permission decision, the approved users, and the categories of information excluded from the pilot.
7. Pilot on a bounded sample with stopping criteria
A pilot should use a small, representative set of already reviewed gaps, including at least one reconciliation, one operational definition, one duplicate, and one item that should be rejected because the answer is already present. Compare the generated output with the source packet. Check whether every question has a valid reference, preserves known answers, uses neutral wording, assigns a plausible respondent, and avoids invented facts.
Define stopping criteria before running the pilot. Stop and revise the workflow if the system invents a source, changes a number, merges distinct periods, exposes restricted text, produces questions without traceable evidence, or repeatedly routes items to the wrong owner. Also stop if the banker cannot review the output efficiently or cannot explain why an item remains open. NIST's generative AI guidance recommends additional human review, tracking, documentation, oversight roles, and retention of testing and validation history for appropriate uses [1].
8. Make banker review a hard gate
The accountable reviewer should inspect the evidence reference, the wording, the implied premise, the respondent, the requested support, and the confidentiality status. The reviewer can approve, edit, defer, merge, reject, or send the item back for more source checking. Keep the review state and reason in the working record. Approval means the banker approves the question for the next process step. It does not mean management has answered it or that the underlying fact has been verified.
Review intensity should match the risk of the task. A simple duplicate suggestion may need a quick comparison. A question about financial reconciliation, a material contract, a customer concentration figure, or a potentially sensitive personnel matter deserves closer source and wording review. ABA Formal Opinion 512 similarly ties the needed level of independent verification to the tool and task and states that professional responsibility cannot be delegated to the tool [2].
9. Produce the implementation outcome and keep the boundary clear
The deliverable is a banker-reviewed management questionnaire with evidence-gap references and assigned respondents. It should contain only approved questions, each linked to the source record that caused it, with unresolved issues clearly marked. Keep rejected drafts and correction logs in the appropriate internal work area rather than presenting them as management requests. When management responds, a human should compare the response with the cited evidence and decide whether the gap is closed, narrowed, or still open.
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. In this workflow, the software is user-operated: the banker remains responsible for permissions, question approval, communication, and conclusions. The practical test is not whether AI produced fluent text. It is whether each approved question is traceable, appropriately assigned, and supported by a review record.
Key Points
- Start with a versioned evidence-gap register and require a traceable source reference for every proposed question.
- Preserve existing answers and classify items as complete, partial, inconsistent, duplicate, or still open before drafting follow-ups.
- Use AI to propose neutral questions and routing suggestions, not to decide facts, representations, valuation conclusions, or approvals.
- Set confidentiality permissions and review tool terms before entering client, company, personal, privileged, or sensitive information.
- Pilot on a bounded sample and stop when provenance, wording, routing, privacy, or banker review quality fails.
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.