AI-assisted recasting should create a reversible presentation map, not rewrite the source ledger. Keep each original account code, label, period, balance, and source location unchanged. Let AI propose a presentation category and explain the evidence it used, then have the accountable associate or senior reviewer approve, reject, or leave the proposal unresolved. Store the rule version and reviewer decision beside the mapping. Test that mapped totals still tie to the source trial balance, but do not treat that tie-out as proof that the classification is correct. Stop the pilot when source labels are missing, categories are ambiguous, or the model cannot show its basis. The deliverable is a reviewer-approved mapping table with unresolved items visible.
Start with the distinction between source data and presentation
Recasting changes how information is grouped for analysis or presentation. It should not silently change the underlying chart of accounts. A source account such as “Repairs - shop” can be mapped to a presentation category such as “Facilities and maintenance,” while the original code and label remain available for tracing. This distinction matters because financial statements describe different dimensions of a business. The SEC explains that a balance sheet presents assets, liabilities, and equity at a point in time, while an income statement presents revenue and expenses over a period. Those statements are related, but they are not interchangeable views of the ledger. [1]
The mapping is therefore an interpretation layer. It is not a new ledger, a restatement, a tax conclusion, or an audit opinion. If a proposed category would alter accounting meaning rather than organize presentation, send it to the appropriate accounting professional instead of allowing the AI workflow to decide.
Define the accountable review boundary before using AI
Assign one person to own the classification decision for the pilot. In a boutique investment bank, that may be the associate preparing the analysis, with a senior reviewer responsible for sign-off under the firm’s own procedures. The owner should define the permitted presentation categories, the source periods in scope, the accounting basis being used, and which decisions require escalation. A category dictionary should include a plain-language definition, inclusion examples, exclusion examples, and an “unresolved” option.
This is consistent with the general direction of official AI accountability guidance. GAO organizes accountability around governance, data, performance, and monitoring, and notes that AI inputs and operations may not always be visible. [2] NIST’s AI Risk Management Framework is voluntary guidance for incorporating trustworthiness considerations into the design, development, use, and evaluation of AI systems. [3] Neither framework makes a classification automatically correct. They support a control design in which a named reviewer can inspect the inputs, challenge the result, and stop the process.
Build a reversible mapping table
Use one row per original account, or one row per account-period combination where the same label can carry different meaning over time. At minimum, retain the source system or file name, account code, original label, period, source location, balance sign convention, proposed presentation category, AI rationale, confidence or uncertainty note, reviewer decision, reviewer date, mapping-rule version, and exception status. Keep the original label as a protected field rather than replacing it with the proposed category.
A useful decision field has three states: approved, rejected, and unresolved. “Unresolved” is not a failure to finish the spreadsheet. It is a controlled result that prevents a weak inference from becoming an asserted fact. A fourth field can record the next action, such as request a subledger description, inspect the general-ledger detail, or ask the controller for clarification. When the rule changes, create a new version and preserve the prior version. Do not overwrite the history of a prior reviewer decision.
Give the model constrained evidence, not an open-ended question
The AI task should be narrow: propose one category from an approved dictionary, identify the source text it relied on, and state what information is missing. Provide only the fields needed for classification. Depending on the engagement’s confidentiality requirements, that may mean using a controlled environment, redacting unnecessary names and customer details, or keeping the model away from raw documents and supplying a prepared account extract instead. Confirm the permitted handling method before uploading financial information. Do not assume that a tool’s output is confidential, privileged, secure, or suitable for every client file.
Prompt instructions should prohibit invention. The model should not create account descriptions, infer a tax treatment, merge accounts merely because their labels look similar, or convert an ambiguous label into a definitive category. Require a short rationale tied to the supplied label or source description. If the evidence is too weak, the expected answer is unresolved.
Check meaning with source and period tests
A label-only match is not enough. Compare the proposed category with the account’s normal balance, statement location, related accounts, and period behavior where those fields are available. Check whether the same source account changes meaning after a business event, reorganization, new product, or change in accounting policy. Review footnotes and accounting-policy descriptions when the presentation depends on a judgment that the ledger label cannot establish. The SEC’s guide specifically points readers to footnotes for significant accounting policies and complex judgments. [1]
Keep accounting-framework assumptions visible. IFRS 18, for example, adds principles for aggregation and disaggregation and requires a complete set of financial statements with comparative amounts for annual reporting periods within its scope. It is effective for annual periods beginning on or after January 1, 2027, with earlier application permitted. [4] That does not make one IFRS presentation category a universal answer for a U.S. private company, a management schedule, or another reporting basis. The reviewer must confirm the applicable basis and purpose before approving a rule.
Separate mapping review from tie-outs and other analyses
After classification, perform a deterministic check that the mapped balances aggregate back to the source population under the documented sign and period rules. Investigate missing accounts, duplicate rows, unexpected sign changes, and unmapped balances. This verifies coverage and arithmetic. It does not establish that an account belongs in the selected category. A perfectly balancing map can still contain a consistently wrong classification.
Keep this work separate from extraction testing, addback eligibility, and reconciliation of conflicting records. Extraction testing asks whether the source values were captured correctly. Addback analysis asks a different eligibility question. Conflicting-record reconciliation asks which source or version should control. Combining all three questions in one AI prompt makes reviewer responsibility unclear and makes an error harder to localize.
Use a pilot with explicit stopping conditions
Start with a small, representative sample that includes clear labels, recurring ambiguous labels, unusual accounts, and known exceptions. The reviewer should compare every proposed category with the source evidence during the pilot. Record false positives, false negatives, missing rationales, and cases where the same label receives inconsistent treatment. Do not turn a pilot observation into a performance promise or a universal accuracy rate.
Stop and escalate when the model cannot identify its evidence, when source descriptions are incomplete, when a category dictionary has overlapping definitions, when the same account receives materially different proposals without a documented reason, or when confidential-data handling has not been approved. Also stop when the reviewer cannot explain why an account was approved. A smaller map with visible unresolved items is safer than a complete-looking map that hides uncertainty.
Approve the version and preserve an audit trail of judgment
The final reviewer should approve the mapping version, not just the exported presentation. Retain the source snapshot, category dictionary, prompt or instruction version, model output, reviewer changes, unresolved queue, and tie-out results under the engagement’s document-control process. If the source data changes, rerun the affected rows or clearly mark the prior map stale. If a reviewer changes a rule, record the reason and affected accounts rather than editing history in place.
Monitoring is part of the control. GAO’s framework includes monitoring as one of its four complementary principles, and NIST presents risk management as an ongoing activity rather than a one-time approval. [2] [3] In practical terms, sample approved rows in later periods, watch for new labels, and review whether unresolved items are being closed with evidence. A reviewer remains responsible for the decision even when AI suggested the initial category.
What the completed deliverable should contain
The implementation outcome is a reviewer-approved account mapping table that preserves source labels and flags unresolved classifications. It should be accompanied by a short scope note stating the reporting purpose, accounting basis, periods, source files, category dictionary version, reviewer, and known limitations. The presentation schedule can show the proposed categories, but users should be able to trace each amount back to the original code and label without relying on memory or an opaque model response.
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, software is user-operated: the deal team still chooses the source, permissions, mapping rules, stopping conditions, and final human review.
Key Points
- Keep original account codes, labels, periods, balances, and source locations unchanged; add a presentation mapping as a separate layer.
- Require AI to propose from a reviewer-approved category dictionary and to identify the evidence and missing information behind each suggestion.
- Use approved, rejected, and unresolved decisions, with versioned rules and preserved reviewer history.
- Tie mapped totals back to the source population, but do not confuse arithmetic coverage with correct accounting classification.
- Stop the pilot when evidence, category definitions, confidentiality permissions, or reviewer rationale are inadequate.
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.