Before sharing an AI-assisted teaser, confidential-information memorandum, management presentation, or buyer email, the banker responsible for release should verify every material statement against an identified source, reconcile every number to the approved financial model or record, confirm that the current version and audience are authorized, and document unresolved issues. The reviewer should also check whether the drafting tool and prompt handled confidential information under an approved policy. AI can organize or rephrase material, but it should not decide what is accurate, disclosable, or approved. A discrepancy in a number, source, period, definition, or permission should stop release until the authorized banker and, where appropriate, the client or counsel resolves it.
Start with a defined release owner and version
An AI draft needs a human release owner before anyone asks the model to improve it. Name the banker who can approve the specific document and the client-side person who must confirm factual or commercially sensitive points. The owner should identify the document type, transaction, intended recipients, confidentiality status, and version date. This is a control boundary, not an administrative detail. The person who generated prose is not automatically the person authorized to distribute it.
Keep the working file separate from the release copy. Record which source pack, financial model, management questionnaire, and prior approved document the draft used. A reviewer should be able to answer, “Which version did this sentence come from?” without relying on the model’s memory or an unexplained prompt. NIST’s Generative AI Profile frames trustworthy AI as a lifecycle concern involving design, use, and evaluation, which supports treating the draft and its review trail as separate stages. [1]
Trace each material statement to a source
Review factual claims in three passes. First, mark statements that describe the company, products, customers, markets, operations, ownership, litigation, debt, or transaction terms. Second, attach a source locator such as a dated management response, general ledger report, contract summary, filing, board-approved forecast, or diligence schedule. Third, compare the wording with the source and label whether it is reported fact, management representation, estimate, or banker interpretation.
An AI system may combine two nearby facts into a statement that neither source actually supports. It may also turn a qualified management comment into an unqualified claim. A reviewer should ask whether the sentence preserves the source’s period, scope, uncertainty, and attribution. If the source is missing, stale, contradictory, or unavailable to the reviewer, remove the claim or hold the document. Do not repair a gap by asking the model to “fill in” a plausible detail.
Reconcile numbers before reviewing the narrative
Numbers require a check independent of prose quality. Compare revenue, gross profit, EBITDA, cash, debt, working capital, headcount, customer concentration, and other decision-relevant figures to the approved source of truth. Confirm the fiscal period, currency, units, reported or adjusted basis, and whether the figure is historical, forecast, run-rate, or management estimate. Recalculate percentages, margins, growth rates, totals, and bridge subtotals with a deterministic spreadsheet or calculator rather than trusting a model-generated result.
Run a cross-document comparison. The same metric should not appear with different definitions in a teaser, CIM, management presentation, and buyer email. If a number changed because the underlying period or definition changed, state that distinction. If two source records conflict and the difference could affect buyer interpretation, stop release and preserve both candidates until the authorized owner resolves the conflict. The review gate should block unresolved numerical or factual discrepancies rather than silently selecting the more favorable figure.
Check narrative logic, qualifiers, and omissions
After the source and number checks, read the material as a buyer would. Does the headline claim follow from the evidence? Does “recurring” have a documented definition? Does “customer” mean an account, location, contract, or revenue relationship? Are concentration, seasonality, churn, debt restrictions, pending disputes, or material dependencies omitted from a sentence that implies broad strength? The goal is not to make the document negative. It is to prevent a polished sentence from overstating what the underlying records show.
Check every comparative or superlative phrase, including “leading,” “unique,” “high growth,” “defensible,” and “low risk.” Those descriptions may be reasonable opinions in one context and unsupported factual claims in another. Replace them with attributable evidence or an appropriately qualified description. Separate management’s view from the banker’s analysis. AI should not be allowed to convert an assumption into a fact or a draft conclusion into a promise about valuation or buyer interest.
Verify confidentiality, permissions, and the recipient list
Before release, confirm that each recipient is covered by the applicable confidentiality process and that the intended disclosure is permitted for that stage of the sale process. Check the document label, watermark, access list, forwarding restrictions, and any client instructions. For an issuer transaction, Regulation FD is one example of a jurisdiction-specific disclosure issue: the SEC explains that the rule addresses selective disclosure of material nonpublic information by publicly traded companies and certain issuers. Its application depends on the facts, so the deal team and qualified counsel should determine the appropriate disclosure path rather than relying on a generic checklist. [3]
The drafting tool deserves the same scrutiny. Before confidential deal information enters a vendor or model, confirm the approved account, retention terms, permitted use, access controls, deletion process, and whether the vendor can use submitted information for another purpose. FINRA’s discussion of generative AI emphasizes understanding where data goes, protecting records, reviewing vendor contracts, planning for failures, and retaining ultimate responsibility even when a function is delegated. [2] If those answers are not available, use a non-confidential sample or stop the AI step.
Test the draft for source drift and prompt risk
Use a small, controlled review set before applying AI to the full material. Ask it to extract claims and numbers with source locations, then compare the output with the underlying records. Test edge cases such as negative values, units in thousands versus millions, fiscal years that do not match calendar years, adjusted versus reported earnings, and a sentence containing both fact and opinion. A useful trial is not evidence that every later output will be correct. It only shows whether the selected workflow exposes errors that a human can detect.
Stop the trial when the system invents a citation, loses a qualifier, changes a period, merges entities, produces irreconcilable numbers, or reveals confidential text outside the approved boundary. Also stop when the reviewer cannot reproduce the result from the recorded source set. NIST identifies source and citation verification as part of managing generative AI risks, so the practical test is whether the workflow makes verification possible, not whether the prose sounds fluent. [1]
Use a human approval gate for each release
The final reviewer should sign off on five questions: Are material claims sourced? Are material numbers reconciled? Are assumptions and qualifiers preserved? Are the tool, document, and recipients permitted? Is this exact version approved for release? A “yes” should be supported by a short checklist or review record that names open issues and their disposition. Approval should be tied to a file hash, version number, or other unambiguous identifier so an edited draft cannot inherit approval accidentally.
Escalate issues that require substantive judgment. The banker may need the client to confirm an operating fact, the model owner to explain a forecast, or qualified counsel to assess a disclosure or confidentiality question in the relevant jurisdiction. Do not ask AI to make that decision, send the material, contact buyers, negotiate language, or grant approval. The human owner remains accountable for the release decision and should be able to reject the draft without pressure to preserve its wording.
Make the checklist operational, not ceremonial
A practical gate has a clear status for every item: verified, corrected, qualified, unresolved, or not applicable. It names the source record, reviewer, date, document version, and next action. Unresolved items are release blockers when they affect a material number, factual assertion, confidentiality boundary, recipient permission, or required approval. Minor style edits can wait. A material discrepancy cannot.
After a release, preserve the approved copy and the evidence used to approve it under the firm’s retention policy. If the source pack changes, reopen the affected claims rather than assuming the old review still applies. FINRA’s guidance highlights model monitoring and business continuity concerns, which supports maintaining a non-AI fallback for critical drafting or review work. [2] The objective is controlled use of assistance, not dependence on a particular tool.
Apply the gate consistently across buyer-facing materials
The same control logic should cover a teaser, confidential-information memorandum, management presentation, diligence response, buyer list note, and email that summarizes the opportunity. The depth of review can reflect the document’s purpose, but the release owner, source trace, number reconciliation, permission check, and human approval should not disappear because the item is short. A concise email can still contain a material forecast, customer fact, or disclosure that changes a buyer’s understanding.
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. The platform is user-operated. A boutique bank can use a documented review gate as its operating framework, while authorized bankers and clients retain responsibility for source judgments, permissions, and release decisions.
Key Points
- Assign a named banker as release owner and tie approval to the exact document version.
- Trace every material statement to a dated source and preserve attribution, period, scope, and uncertainty.
- Reconcile numbers independently, including units, fiscal periods, definitions, formulas, and cross-document consistency.
- Confirm confidentiality, tool permissions, retention terms, recipient authorization, and any jurisdiction-specific disclosure review.
- Stop release when a material number, fact, citation, permission, or approval remains unresolved.
- Keep human reviewers responsible for substantive judgments, approvals, and distribution decisions.
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.