How Should M&A Attorneys Trace AI-Summarized Financial Figures to Original Deal Sources?

Assign responsibility before asking AI to summarize figures

Give AI a bounded task: extract identified financial figures from an approved document set and draft descriptions that preserve their source labels. Do not ask it to decide which earnings adjustments are acceptable, resolve conflicting statements or certify the business's performance. Assign a register owner to maintain source connections, a financial reviewer to validate accounting judgments, and a legal reviewer to check the proposed language and disclosure permissions. Identify separately who is authorized to release the materials.

ABA Formal Opinion 512, issued in July 2024, explains that lawyers need an appropriate degree of independent verification or review of generative AI output and cannot surrender professional judgment to a tool. The opinion is based on the ABA Model Rules, not a uniform rule binding every jurisdiction. Counsel should check the rules and guidance applicable to the engagement. [1] This workflow addresses financial source integrity, not valuation, investment-banking recommendations or transaction-specific legal advice.

Establish the original records and their approved versions

Ask the finance lead to identify the records that support each category of figure. These may include issued financial statements and their notes, approved management accounts, a quality-of-earnings report, or a reconciliation prepared and reviewed for the transaction. A draft marketing presentation is not a substitute for those underlying records. If a reconciliation is the immediate source, preserve the connection from its starting amount to the original statement or ledger-supported schedule.

Record each document's title, date, version and status, including any audited, reviewed or unaudited description actually present. Copy that description accurately without upgrading it. Preserve superseded versions for the internal comparison rather than overwriting them. Where a replacement changes a figure, ask the financial reviewer to explain the difference and identify affected draft passages. Keep deal-source citations inside the authorized review workspace; internal traceability does not require exposing confidential source files to every recipient of a teaser.

Build a register that another reviewer can actually use

Create one register entry for each distinct figure and use occurrence links when it appears in multiple places. Capture the exact draft wording and location, the source amount as displayed, the metric's original name, the relevant period, entity scope, currency and scale. Add the source document version and a precise locator: page and table row for a statement, or workbook, sheet and cell for a spreadsheet. A filename alone leaves the reviewer to reconstruct the evidence.

Keep separate fields for the proposed displayed amount and any transformation, such as rounding or a unit conversion. Record whether the figure is reported, adjusted, forecast, pro forma or otherwise defined by its source. For adjustments, link the approved reconciliation and its explanation. Add a responsible reviewer, review date, unresolved issue and status. Useful statuses distinguish extracted but unchecked, source matched, awaiting financial validation, disputed and approved for the specified use.

For illustration, a row for adjusted EBITDA should point to the approved reconciliation's ending amount, its underlying starting measure and the draft sentence where it appears. That row is incomplete if the citation resolves but the reconciliation period or included businesses remain unclear.

Check the period, currency and units before the amount

Start with the column header and notes surrounding the figure. Record actual start and end dates rather than relying on shorthand such as current year. Distinguish a fiscal year from a calendar year, a quarter from year-to-date results, and historical results from projections. For last-twelve-month figures, identify the cutoff date and the approved schedule used to assemble the period. Do not let AI infer missing months or relabel an annualized partial period as a completed year.

Check the currency and whether amounts are shown in whole units, thousands or millions. Preserve the original scale in the register even when marketing materials use a different scale. Have any conversion calculated outside the language model using a reproducible formula, with the financial reviewer confirming the basis. Specify the rounding convention before comparing displayed figures. A rounded draft amount may match its source, but the register should retain the unrounded input available in that source.

Read footnotes before accepting an apparent match. A comparative column may have been restated, or a heading may apply only to part of a table. Send ambiguous presentation questions back to the source owner rather than choosing whichever reading fits the draft.

Confirm which entities and operations the figure includes

Match the reporting entity to the business described in the materials. Distinguish consolidated group results, a subsidiary, a division, a carve-out and a management-defined combination. Ask the financial reviewer to confirm included entities, consolidation eliminations and exclusions. A legal sale perimeter and a financial reporting perimeter can require a documented bridge; do not assume that a shared trading name establishes identical scope.

Use an exception field for acquisitions, disposals, discontinued operations or changes in the reporting group. For an illustrative carve-out review, a group revenue figure cannot be treated as the sold division's revenue simply because both records have the same year label. Require the approved allocation or carve-out schedule and retain its assumptions. Counsel can check that the narrative describes the scope consistently, but the financial professional should validate allocation and consolidation judgments. If that validation is unavailable, leave the affected figure unresolved rather than creating a scope explanation with AI.

Preserve reported and adjusted labels with the reconciliation

Do not shorten adjusted EBITDA to EBITDA, rename a management-defined measure as net income, or treat a forecast as reported performance. Preserve the source's definition and identify each adjustment in the linked reconciliation. Ask the financial reviewer to confirm the starting measure, signs, classification and supporting records for each adjustment. A correctly copied total does not establish that its underlying accounting judgment is appropriate.

The SEC's non-GAAP interpretations explain, within the relevant US securities-reporting framework, that inappropriate labels and inconsistent adjustments between periods can make measures misleading. They also explain that extensive disclosure does not necessarily cure a materially misleading measure. [2] These are useful review prompts, not a conclusion that every private-company teaser or confidential information memorandum is subject to identical SEC requirements. Transaction communications and forecasts can involve specific conditions and exceptions that counsel must assess. [2]

Compare adjustment definitions across periods and across the teaser, memorandum and approved schedules. If the definition changes, retain the explanation and ask whether comparative wording needs revision. AI should flag discrepancies for review, not invent an add-back justification or decide that an expense is nonrecurring.

Test whether the citation supports the actual claim

Reopen every material source locator used in the draft. Confirm that the document exists, the locator points to the relevant record, and the record supports the entire sentence. A citation to annual revenue does not support a claim about recurring revenue, organic growth or profitability. NIST's July 2024 Generative AI Profile identifies fabricated citations as a confabulation risk and recommends reviewing and verifying output sources and citations during testing and ongoing monitoring. It is voluntary risk-management guidance, not a compliance certification. [3]

Recalculate derived margins, changes and reconciliation totals with deterministic tools using the approved inputs. Retain the formula and denominator definition in the register. Compare the underlying precision before evaluating a rounded result. Escalate conflicting source amounts with both versions visible; do not replace one silently. After correcting a figure, check every occurrence and surrounding claim. Removing a mistaken amount is insufficient if the draft still describes the business as growing on the basis of that amount.

Run a permitted trial with explicit stopping conditions

Before entering deal information, have the legal and technology reviewers examine the proposed tool's terms, retention arrangements, training use and access controls. Decide what information may be supplied, who can see inputs and outputs, and what deletion or retention obligations apply. ABA Opinion 512 calls for a fact-specific confidentiality assessment and understanding of relevant tool terms. Whether disclosure or informed consent is required depends on the applicable professional rules and circumstances; generic engagement boilerplate is not a substitute where informed consent is needed. [1] No setting should be described as guaranteeing confidentiality or privilege.

Start with public or synthetic financial records, then use only specifically permitted deal records. The financial reviewer should prepare a manually checked comparison set that includes difficult headers, adjustment labels and entity changes. Compare each AI extraction against that set and document errors. NIST recommends defined go/no-go criteria and cautions against extrapolating capabilities from narrow anecdotal testing. [3]

For this workflow, stop the trial if AI invents a source, loses a period or unit, changes entity scope, omits an adjustment label, or exposes information outside the permitted context. Pause the affected workflow, investigate and retest before resuming. A successful trial does not remove human review of material figures or authorize external disclosure.

Freeze the reviewed version and require human release

Create a figure-to-source register with checked periods, units, entity scope, and adjustment labels. Before release, have the register owner confirm that every material figure has a resolved status and the required financial and legal reviews. Have the authorized deal operator approve the recipient, document version and disclosure scope. Keep unresolved figures out of approved materials. If a source or draft changes afterward, reopen the affected entries and their dependent claims. AI should not approve materials, contact recipients or decide what may be disclosed.

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 with no success fee. The register and review gates described here are a recommended implementation framework, not a claim that Acquivo supplies a particular integration or verification feature. Accountable professionals retain their review responsibilities, and authorized users control release.