Bankers can keep AI-assisted deal narratives balanced by treating the model as an editorial drafting aid, not as the source of truth. Start with an approved fact set that separates historical actuals, management projections, assumptions, definitions, and known risks. Require the same metric definitions and period labels in the CIM, teaser, and management presentation. Have a named banker review every material claim against its source, test whether qualifiers survived the draft, and log contradictions rather than silently resolving them. A controlled pilot should stop when the system invents support, drops a material caveat, changes a definition, or exposes confidential content. The final decision to revise, approve, or disclose remains with the responsible human team, subject to the transaction’s applicable professional and legal requirements.
Treat the approved fact set as the source of truth
A balanced narrative begins before anyone asks an AI system to write. Create a dated, versioned fact set for the transaction and identify the record that controls each field. That record may be an approved financial schedule, management-approved operating data, a diligence response, or a counsel-reviewed disclosure draft. The important point is authority: the model should transform approved material, not fill gaps from general knowledge or an earlier deal.
Separate actual performance from projections at the data level. Give each figure a period, unit, currency, definition, and status such as actual, estimate, management projection, or scenario assumption. Store the source location and the person who approved its use. This prevents a historical margin from being blended with a projected margin or a management target from being narrated as an achieved result. It also makes a later challenge answerable without relying on memory.
Use an explicit unknown value when support is missing. A blank or review flag is safer than an AI-generated explanation. The banker can decide whether to obtain support, remove the claim, or describe the uncertainty.
Build a claim register before drafting
The implementation outcome should be a banker-reviewed register of claims, qualifiers, and contradictions. Each row can contain the proposed sentence, claim type, source record, period, actual-versus-projected status, relevant definition, required qualifier, risk or dependency, reviewer, and disposition. Claim types might include historical result, forecast, customer concentration, market statement, operational capability, or transaction-process description.
Write the claim in a form that can be checked. A statement such as ‘strong growth’ is difficult to review because it has no defined period or baseline. A statement that identifies the measure, comparison period, and source is easier to challenge. The register should also capture what the draft must not imply. For example, a forecast should not be presented as a commitment, and an adjusted metric should not be given a label that suggests it is a reported GAAP measure.
Assign one accountable human reviewer for each material group of claims. The reviewer is not approving the AI tool. The reviewer is approving the underlying statement for the intended document and audience. Keep unresolved rows open until the source, qualifier, or wording is fixed.
Use AI for controlled transformations, not judgment
A useful prompt asks the system to reorganize approved facts into a specified format while preserving numbers, labels, qualifiers, and source markers. It can propose a shorter paragraph, identify repeated definitions, or compare two drafts. It should not decide that a risk is immaterial, convert an assumption into a fact, select a buyer-facing emphasis, or invent a rationale for an unexplained variance.
Give the model a narrow task boundary. The input should contain only the approved material needed for that task, and the output should be required to flag missing support instead of completing the thought. Prohibit autonomous outreach, negotiation, disclosure, and approval actions. A human should copy any accepted language into the controlled working document after review, rather than allowing a model to publish or distribute it.
For confidential material, follow the firm’s information-classification and vendor-approval process. A trial should use synthetic or redacted content until the responsible security and confidentiality owners confirm that the selected environment is permitted. Do not assume that a tool’s general security statement creates privilege, confidentiality, or regulatory compliance for a particular transaction.
Reconcile the CIM, teaser, and management presentation
Cross-document consistency is more than matching headline revenue. Compare the same metric across every document, including its definition, period, unit, scope, and whether it is actual or projected. Then compare the surrounding explanation. A growth figure may be numerically identical while one document attributes it to volume and another attributes it to pricing. Those are different narratives and require source review.
Create a reconciliation pass that checks names, dates, segment labels, customer concentration language, headcount, adjusted measures, debt descriptions, use-of-proceeds language where relevant, and the status of material diligence items. For each mismatch, record whether it is a genuine difference in audience or an error. If a teaser needs brevity, shorten the explanation without deleting a qualifier that changes the reader’s understanding.
Use a stable glossary for terms such as recurring revenue, adjusted EBITDA, backlog, churn, and organic growth. The glossary should state the calculation basis or point to the approved definition. If management changes a definition, update the fact set and re-run the comparison across all affected documents.
Preserve risk, uncertainty, and downside context
Balance does not mean adding a generic risk paragraph at the end. Put the relevant limitation beside the claim it qualifies. If a projection depends on a renewal assumption, say so near the projection. If a margin improvement depends on hiring, pricing, capacity, or a one-time cost, identify that dependency. If the source is management’s estimate, label it as such and retain the applicable assumptions.
FINRA Rule 2210 provides a useful benchmark for communications by its members: communications must be fair and balanced, provide a sound basis for evaluating facts, and not omit a material fact or qualification when the omission would make the communication misleading. It also calls for balanced treatment of risks and potential benefits and restricts performance predictions and projections, subject to specified conditions and exceptions. The rule is not a universal checklist for every private transaction or adviser, so the responsible team must determine which requirements apply. [1]
A practical review question is: what would a reasonable reader need to understand the claim’s uncertainty? The answer may be a range, an assumption, a dependency, a time boundary, or a statement that the information has not been independently verified. Do not manufacture precision to make a narrative sound complete.
Keep actuals, projections, and adjusted metrics distinct
A narrative should show the bridge between reported history and forward-looking expectations. Label the last historical period, the forecast period, and the source of each. If the forecast uses a management case, do not describe it as a consensus view or an independently validated outlook. If a chart combines actuals and projections, make the transition visible in the title, legend, and nearby text.
SEC staff guidance explains that a non-GAAP measure can be misleading when adjustments omit normal recurring expenses, are inconsistent between periods, exclude charges but not gains, or are not clearly labeled. The staff also warns against giving an adjusted measure undue prominence over its comparable GAAP measure. Those points support a review control for deal materials: preserve the approved calculation, name the adjustment, show the comparable reported measure where required, and check that the same treatment is used across periods and documents. [2]
Do not let AI normalize an awkward source. If management uses a bespoke KPI, retain its approved name and definition, identify the period, and explain the limitation. A cleaner label can create a false impression of comparability.
Use neutral language and test the total mix
The FASB Conceptual Framework describes faithful representation as complete, neutral, and free from error. It explains that completeness includes the descriptions and explanations needed to understand what is depicted, while neutrality means the information is not slanted, weighted, emphasized, deemphasized, or manipulated to produce a favorable or unfavorable response. It also says an estimate can be faithfully represented when it is clearly described as an estimate and its limitations are explained. [3]
Apply that principle at sentence level. Replace unsupported intensifiers such as ‘unmatched’ or ‘certain’ with a description of the underlying fact. Avoid selecting only the favorable period when the approved fact set includes a relevant deterioration or one-off benefit. Ask whether the order of the paragraphs, chart scale, headline, and footnote placement could cause a reader to miss a material limitation. Balance is a design and editing issue as well as a numerical one.
For possible errors or omissions, use a reasonable-reader lens. SEC staff guidance describes materiality as an objective assessment of whether information would significantly alter the total mix available to a reasonable investor, and says the assessment should consider both quantitative and qualitative factors rather than operate as a mechanical test. That is a review concept, not individualized legal advice, and transaction-specific judgments should go to qualified counsel or other appropriate professionals. [4]
Pilot with stopping conditions and human review
Run a limited pilot on a non-public working copy or redacted sample. Measure process quality through review findings, not a promised accuracy percentage or time-saving claim. The pilot should test whether the system preserves source markers, actual-versus-projected labels, qualifiers, metric definitions, and document-specific audience constraints.
Stop the pilot and escalate if the system invents a source, changes a number, merges actuals with projections, drops a material risk, produces contradictory definitions, reveals restricted content, or cannot show which input supported a sentence. A repeated failure is evidence that the task or tool is not ready for that workflow. Do not expand access merely because a few outputs look polished.
The banker’s review should compare accepted language with the claim register and source records. A second reviewer may be appropriate for material financial, legal, or confidentiality judgments, depending on the firm’s procedures and the transaction. Keep the final approval outside the model and record the document version that was approved.
Turn the workflow into a repeatable checklist
Before release, confirm that every material claim has a source, owner, period, and status. Confirm that actuals and projections are visibly separated. Recheck definitions and units across the CIM, teaser, and management presentation. Confirm that risks and dependencies sit next to the claims they qualify. Review adjusted measures for labels, reconciliations, and consistent treatment. Search for unsupported superlatives, implied guarantees, unmarked estimates, and language that could be read as a price recommendation or outcome promise.
Then document the disposition of exceptions. A contradiction should be resolved by the responsible source owner or clearly presented as an unresolved issue, not silently harmonized by the model. A missing qualifier should be restored or the claim removed. A source that cannot be confirmed should remain a review item. The checklist is complete only when the named banker can explain why each material statement is present and why its wording fits the intended document.
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 workflow owner remains responsible for source selection, permissions, review, and approvals.
Key Points
- Start from a dated, approved fact set that separates actuals, projections, assumptions, definitions, and unknowns.
- Maintain a banker-reviewed claim register with source locations, qualifiers, contradictions, owners, and dispositions.
- Compare the CIM, teaser, and management presentation for matching definitions, periods, units, labels, and risk context.
- Keep adjusted metrics clearly labeled and consistently calculated, with the comparable reported measure treated appropriately.
- Use a limited pilot with explicit stopping conditions for invented support, dropped caveats, changed definitions, or confidentiality failures.
Sources and references
Source retrieval dates record research access, not article publication. Read our editorial and date policy.
- [1] FINRA Rule 2210, Communications with the Public · Accessed
- [2] SEC Corporation Finance Interpretations, Non-GAAP Financial Measures · Accessed
- [3] FASB Conceptual Framework for Financial Reporting, September 2024 · Accessed
- [4] SEC Assessing Materiality: Focusing on the Reasonable Investor When Evaluating Errors · 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.