When new financials arrive, do not ask an AI tool to overwrite the deal model. Freeze the prior version, identify the exact reporting period and source files, and create a new refresh version. A banker should map each changed input to its source, reconcile reported figures to the prior model, document recast decisions, and recalculate formulas before reviewing outputs. Test the model for broken links, unexpected variance, and changed assumptions. Keep the prior version available for comparison, and require a named human reviewer to approve the refreshed file for its stated use. AI can help locate and classify changes, but it should not certify valuation, approve disclosures, or replace review.
Start with a controlled refresh, not an overwrite
Treat the incoming financials as a new evidence set. Record the file name, issuer or company, reporting period, currency, units, publication date, receipt date, and whether the statements are preliminary, restated, reported, or management-adjusted. The source-of-truth record should be the received financial statement or other authorized company record, not an AI summary. This preserves the distinction between a newly reported fact and an interpretation of that fact.
Make a read-only copy of the last approved model and assign the refresh a new version identifier. Preserve the old workbook, source package, assumptions, outputs, and review notes. A simple naming convention such as deal code, reporting period, and revision number is useful, but the important control is that a reviewer can reconstruct what changed and which version was used. FINRA describes inventories, testing, monitoring, and parallel use of current and new models as useful considerations for AI applications [3].
Confirm the financial source and period
Before extracting values, compare the new statements with the prior source package. Check entity name, fiscal year-end, quarter, consolidation scope, accounting basis, currency, units, and whether comparative periods were restated. Do not treat a later-arriving file as automatically superior if it covers a different period or uses a different presentation. If the source conflicts with an earlier input, retain both records and explain the selection in the change log.
An AI assistant may help find candidate line items, but the banker should inspect the page, note, table, or approved data record supporting each material input. Capture the source locator beside the proposed value. For example, a revenue change should identify the statement and period, while a working-capital adjustment should identify the relevant balance-sheet lines and the rationale for including or excluding them. If the source is incomplete or ambiguous, stop the affected update rather than filling the gap from memory or an unverified estimate.
Build an input-level reconciliation
Create a reconciliation between the prior model input and the new source. At minimum, show the old value, new value, absolute change, percentage change when meaningful, source locator, treatment, and reviewer status. Separate reported changes from recast changes. A reported change comes from the new financial statements. A recast change reflects a deliberate analytical treatment, such as reclassifying a one-time item or aligning a management measure with the model’s stated definition.
Investigate material or surprising variances before changing downstream outputs. Check signs, decimal scale, thousands versus millions, currency, period alignment, and whether a line moved because of a restatement rather than current-period performance. Preserve the original reported amount even when a normalized amount is used. The model should make clear which cells are sourced, which are derived, and which are assumptions. That distinction makes later review faster and prevents an AI-generated extraction from becoming an unexplained hard-coded number.
Use AI for bounded extraction and comparison
Define the AI task narrowly. Suitable tasks include locating specified line items, comparing two versions of a statement, proposing a list of changed labels, or flagging a missing source locator. The prompt or operating instruction should identify the permitted source package, expected period, output fields, and a requirement to quote or point to the supporting record. Do not allow the tool to edit the production model, infer a valuation conclusion, select a recast without approval, or disclose material to an unapproved service.
Set a stopping condition before the trial begins. Stop and route to a banker when the source has conflicting periods, unreadable tables, nonstandard labels, unexplained totals, missing notes, or a proposed value that cannot be tied to a source locator. A useful pilot compares the AI’s proposed extraction with a human-checked sample and records errors. If the error pattern affects material inputs or cannot be explained, discontinue AI assistance for that refresh and use a manual process. NIST describes its AI RMF as voluntary and focused on incorporating trustworthiness into AI design, use, and evaluation, not as a certification of a particular tool [2].
Recalculate formulas and inspect dependencies
After approved inputs are updated, recalculate the model in a controlled environment. Check that formulas still reference the intended sheets and periods, that links are not broken, and that totals reconcile. Inspect the income statement, balance sheet, cash flow, debt schedule, working-capital calculations, tax assumptions, transaction adjustments, and valuation outputs that depend on changed cells. A formula recalculation is not a review of whether the formula remains conceptually appropriate.
Compare the refreshed outputs with the prior version and explain meaningful movements. A bridge can separate the effect of new reported results, approved recasts, changed assumptions, and formula or structural changes. Avoid presenting an output as reliable merely because the workbook calculated without an error. OCC guidance emphasizes model development and use, validation and monitoring, governance, and clear roles, while also noting that practices should reflect the organization’s risk profile and use case [1].
Test the refreshed model before relying on it
Run mechanical checks first. Confirm that required source fields are populated, balance-sheet checks balance, subtotals tie, dates advance as expected, units are consistent, and no error values or unexpected blanks remain. Then run reasonableness checks against the source package and the prior version. Look for changes that are possible but unexplained, such as a margin movement caused by a sign error or a debt balance change caused by a period mismatch.
Use targeted sensitivity or scenario checks where the changed inputs affect a decision-relevant output. The purpose is not to prove that a valuation is correct. It is to expose whether the model behaves as designed under plausible changes and whether a new input creates an unstable or unintuitive result. FINRA’s report discusses upfront and ongoing testing, new datasets and scenarios, monitoring, human review, and fallback plans for AI applications [3]. For a boutique bank, the scale of those controls should be proportionate to the model’s purpose, materiality, and reliance.
Maintain a versioned change log
The change log should connect every approved update to a person, source, decision, and model version. Useful fields include change ID, sheet and cell or named range, prior value, new value, source locator, change type, reason, AI involvement if any, preparer, reviewer, date, test result, and approval status. Record rejected suggestions when they could otherwise be mistaken for accepted inputs. Keep the log with the model package rather than in an isolated chat transcript.
Document assumptions separately from facts. If management provides an estimate or an adjustment is proposed for analytical consistency, label it as an assumption or judgment and identify who approved it. Do not let an AI explanation become the official rationale without human confirmation. A later reviewer should be able to answer four questions: what changed, why it changed, what evidence supported it, and who decided the refreshed version was fit for its stated purpose.
Require human release approval
The accountable reviewer should inspect the source reconciliation, changed inputs, recast decisions, formula checks, output bridge, exceptions, and open issues. The reviewer should be someone with appropriate responsibility for the deal model, not merely the person who ran the extraction. Approval should state the version, purpose, date, and any limitations. If a material issue remains unresolved, mark the model as not approved for the affected use and retain the stopping reason.
Do not use the refreshed model for external materials, a pricing discussion, or a transaction decision until the applicable internal review is complete. This article does not establish a legal or regulatory requirement for every adviser or transaction. Bank policies, engagement terms, professional obligations, and jurisdiction-specific rules may impose additional controls, so qualified counsel or compliance personnel should address questions in those areas. Human approval is a release control, not a guarantee that the model or valuation is correct.
Turn the workflow into a repeatable checklist
A practical checklist has five gates. First, source gate: confirm identity, period, basis, units, authority, and permissions. Second, input gate: map every changed value to a source locator and reconcile it to the prior version. Third, calculation gate: recalculate formulas and inspect dependencies. Fourth, testing gate: complete mechanical, variance, and targeted scenario checks, with stopping criteria for unresolved exceptions. Fifth, release gate: obtain reviewer sign-off and archive the prior and refreshed packages together.
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 any implementation, the platform is user-operated software, so the bank remains responsible for its source choices, permissions, review process, and approval decision. The useful outcome is not an AI-certified model. It is a traceable refresh in which updated financials, adjustment decisions, calculations, and human review remain connected.
Key Points
- Freeze the prior approved model and create a separately identified refresh version.
- Verify the new source’s entity, period, currency, units, accounting basis, and restatement status before extraction.
- Link every changed input to a source locator and distinguish reported facts from recasts and assumptions.
- Use AI only for bounded extraction or comparison, with permissions, error checks, and stopping criteria.
- Recalculate, test, log exceptions, and require accountable human sign-off before the refreshed version is used.
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.