Bankers should test AI financial ingestion on a representative, source-linked pack before using any extracted number in analysis. Include spreadsheet tables, PDF tables, scanned statements, negative values, units, footnotes, formulas, blank cells, and deliberately difficult layouts. For every control value, record the source file, page or sheet, cell or region, period, currency, and unit, then compare the output with a manually checked control total. Log omissions, sign changes, unit changes, and unsupported inferences as exceptions. Keep extraction accuracy separate from accounting judgment, and require a named human reviewer to accept, reject, or quarantine each result. If a critical field is missing, ambiguous, or fabricated, stop the pilot rather than silently filling the gap.
Start with a test pack, not a live deal file
The first test should answer a narrow operational question: can the workflow reproduce values that a banker can locate and verify in the source record? Build a small but representative pack from permitted materials. It should contain at least one clean spreadsheet, a PDF with selectable text, a scanned statement, a page with a dense table, and a document containing footnotes or appendices. Use redacted or synthetic materials when client confidentiality or the trial terms do not permit production documents.
Do not choose only easy pages. Deliberately include a negative expense, a parenthesized loss, a percentage, a currency symbol, a thousands or millions label, a subtotal, a blank cell, a formula cell, and a value that wraps across lines. Include a period label and a comparative column. The point is not to create a benchmark that looks impressive. It is to expose the failure modes that could change a model, a normalization, or a conclusion.
Define the source of truth and its locator
Before running extraction, create a control register. Each row should identify the source filename and version, page number or worksheet, cell or table region, line label, reporting period, currency, unit, and the expected value. For an image, record the page and bounding region or another reproducible visual locator. Preserve the original record rather than copying only the number into a separate sheet.
This discipline follows the logic of structured financial disclosure. The SEC describes Inline XBRL as both human-readable and machine-readable and notes that tagged points can expose context such as definitions and reporting periods [1]. That context is part of the value being tested. A value without its period, unit, sign, and label is not a safely ingested financial fact.
Test tables, signs, units, and periods separately
A single pass or overall accuracy score can hide material errors. Test the dimensions independently. For tables, check that row labels stay aligned with the correct column and that a continuation page is not read as a new table. For signs, test minus signs, parentheses, credit balances, and losses. For units, test dollars versus thousands or millions, percentages versus decimal fractions, and per-share amounts versus totals.
For periods, verify fiscal year, quarter, year to date, and comparative columns. Record whether the source is reported, adjusted, or pro forma if that distinction appears in the document. Do not let a model infer a period from a filename when the page itself supplies a different period. A result that has the right digits but the wrong period or unit is an exception, not a pass.
Make footnotes and formulas first-class test cases
Footnotes often qualify a number without changing its visual prominence. Add cases where a footnote explains a reclassification, an accounting policy, a discontinued operation, a segment boundary, or a non-GAAP adjustment. The extraction output should preserve the note reference or link it to the note text. It should not turn a qualified value into an unqualified input.
For spreadsheets, test both displayed values and formula provenance. A displayed subtotal may depend on hidden rows, a selected range, a linked workbook, or a manually overridden formula. Decide whether the workflow is extracting the displayed result, the formula, or both. Recalculate control totals independently, and flag any formula that cannot be traced or reproduced. Extraction can report what the file contains. It cannot decide whether an accounting treatment is appropriate for a transaction analysis.
Use control totals and source-linked exceptions
For each test document, select manually checked control totals such as a statement subtotal, balance-sheet equality, or a sum of visible components. The reviewer should calculate the check from the source values using a deterministic method and store the basis. Then compare the extracted values without rounding away a difference. A control total is a diagnostic, not proof that every line is correct. Two compensating errors can still produce the same total.
Log every exception with the source locator, expected value, extracted value, error type, severity, reviewer, and disposition. Useful error types include omission, duplication, transposed digits, sign inversion, unit mismatch, period mismatch, broken formula, lost footnote, and unsupported inference. Keep a separate field for the accounting question. That prevents a judgment call from being mislabeled as an extraction defect and prevents an extraction defect from being rationalized as judgment.
Prohibit silent gap filling
The safest behavior for a missing cell or unreadable scan is an explicit missing or needs-review status. It is not a zero, an estimated value, or a value copied from a nearby period unless a human-approved rule says so and the result is labeled accordingly. The SEC warns that its own financial statement datasets are derived from registrant-provided structured data, may contain extraction or compilation errors, omit some available information, and are not substitutes for the full filings [2]. That is a useful boundary for any ingestion workflow: a convenient structured output still needs source review.
Set stopping conditions before the pilot. Stop and quarantine the output if a critical control value is fabricated, if a sign or unit changes without an exception, if source locators are absent, if a scanned page is unreadable, or if the process cannot preserve the original file and version. Also stop when a reviewer cannot tell whether a value came from the source or from an inference. Do not proceed merely because most rows look plausible.
Assign the human review and acceptance record
Name the accountable reviewer by role, not by an informal assumption. A banker or analyst who owns the workpaper can verify source fidelity, while an accounting specialist may be needed for a classification or recasting judgment. The acceptance record should state what was tested, which source version was used, which exceptions remain open, what disposition was chosen, and whether the output is approved for a defined downstream use.
NIST’s Generative AI Profile recommends comparing output with known ground truth, using human oversight and automated evaluation, documenting upstream data sources, and testing data and content flows [3]. Those practices fit ingestion because they make review an explicit control rather than a vague expectation. Human acceptance does not mean the reviewer endorses every accounting conclusion. It means the reviewer accepts a stated extraction scope and its known limitations.
Run a bounded pilot before expanding the workflow
Begin with a defined document class and a defined downstream use, such as preparing a reviewable historical-data workpaper. Do not combine spreadsheets, scans, covenant calculations, and valuation adjustments into one pass. Track which layouts were tested and which were not. Re-run the same pack after any material change to the model, prompt, parser, template, or document-handling step.
The pilot should end in one of three states: accepted for the stated scope, accepted with documented exclusions and mandatory review, or rejected pending remediation. An accepted pilot is not a guarantee for a new issuer, a new language, a new statement format, or a new use. When the source class changes, expand the test pack and repeat the controls.
Turn the result into a repeatable banker workpaper
The practical deliverable is a financial ingestion test pack, a source-linked exception log, and a human acceptance record. Store the pack with a version identifier and a short change note. Keep the control register close to the extracted output so a later reviewer can reconstruct what was compared. Preserve rejected examples because they reveal where the workflow should stop.
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 context, the useful implementation question is still operational: what source was used, what did the extraction return, which exceptions remain, and who accepted the output for the next step? A software workflow should support those decisions without replacing the accountable professional.
Key Points
- Build a representative test pack with spreadsheets, PDFs, scans, difficult layouts, signs, units, footnotes, formulas, and missing cells.
- Record a source locator, period, currency, and unit for every control value, not just the extracted number.
- Use manually checked control totals as diagnostics, while keeping extraction defects separate from accounting judgments.
- Treat missing, unreadable, ambiguous, or fabricated values as exceptions and stop critical uses rather than silently filling gaps.
- Require a named human reviewer and an acceptance record that states scope, open exceptions, and permitted downstream use.
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.