Step 1: Gather and Preserve the Source Records
Collect income statements, balance sheets, cash-flow statements, general ledgers, trial balances, receivables and payables ledgers, bank statements, invoices, tax forms, budget workbooks, filings, models, scans, and supporting documents. Preserve filenames, versions, reporting periods, fiscal dates, sheet names, row and column labels, units, currencies, formulas, and audit-trail tabs.
What success looks like: Every input can be identified and retrieved without guessing which version or period was used.
Common mistake to avoid: Do not overwrite the original export or silently change missing values before the audit begins.
Step 2: Define the Cross-Check Rules
Write down the tests before looking at the result. Arithmetic checks include assets equaling liabilities plus equity, revenue less expenses equaling operating income, annual totals equaling monthly totals, and variance equaling actual minus budget. Record-to-record checks compare statements with ledgers, bank balances, customer balances, debt schedules, tax forms, and dashboards. For broader workflows, document the rules in reusable AI workflows so repeated corrections become persistent audit rules.
What success looks like: Another reviewer can reproduce the scope, formulas, comparison periods, and acceptable tolerances.
Common mistake to avoid: Avoid changing the definition of a ratio after seeing which result looks favorable.
Step 3: Ask the First AI to Prepare the Analysis
Use the primary AI to extract values, normalize dates and units, identify relevant fields, calculate totals and ratios, and create the requested workbook, report, dashboard, or presentation. Require citations to the source file, sheet, row, column, and field for each important figure. The output should also list assumptions, exclusions, and missing records. If the work includes recurring reporting, connect the extraction with automated financial reporting practices that preserve traceability.
What success looks like: The deliverable explains where each material number came from and how it was calculated.
Common mistake to avoid: Do not accept an attractive summary that has no source references or treatment of missing data.
Step 4: Use a Separate AI Auditor
Send the original records and the primary AI’s output to an independent auditor, instructing it not to rely on the first system’s conclusions. The auditor should re-extract key figures, recompute totals and ratios, compare recalculated values with the delivered output, test formulas, verify labels and periods, check charts, and identify missing, duplicated, stale, or unsupported data. An independent AI auditor changes the review from checking everything to examining what is actually flagged.
What success looks like: Each tested item receives a pass, fail, or requires-review status with evidence for exceptions.
Common mistake to avoid: Do not give the auditor only the summarized answer; it needs the source records as well.
Step 5: Reconcile the Financial Statements
For the balance sheet, verify Assets = Liabilities + Equity. One supplied example reports 2025 assets of $619.0B, liabilities of $275.5B, and equity of $343.5B, which matches because $275.5B + $343.5B = $619.0B. The same dashboard matched 15 of 16 years, while 2010 could not match because liabilities were unavailable. For the income statement, compare revenue, expenses, operating income, and net income; for cash flow, compare operating cash flow, net income, and capital expenditures. A structured three-statement modeling review keeps those relationships together.
What success looks like: Statement totals reconcile, and any unmatched period has a documented reason rather than a forced adjustment.
Common mistake to avoid: Never convert an unavailable liability or cash-flow value into zero to make the equation balance.
Step 6: Cross-Check Growth Rates and Working Capital
Compare receivables, revenue, inventory, accrued liabilities, and operating cash flow over the same periods. The supplied red-flag data shows FY2025 receivables growth of +19.1%, revenue growth of +6.4%, and a growth gap of +12.6 percentage points, with receivables of $39.8B. Allowance coverage is allowance balance divided by receivables balance; FY2025 was reported as $0 divided by $39.8B, or 0.00%. Accrued liabilities were unavailable for FY2018–FY2022, so those periods should remain unavailable.
What success looks like: Growth gaps and coverage ratios use aligned periods, and missing years appear as gaps rather than invented values.
Common mistake to avoid: Do not interpret a zero allowance or a growth gap without reviewing the underlying schedules and accounting policy.
Step 7: Compare Cash Flow With Earnings
Calculate operating cash flow divided by net income, then calculate operating cash flow minus net income. The supplied FY2025 figures show an OCF-to-net-income ratio of 1.00x and a dollar difference of -$528.0M. The largest absolute difference was $24.5B in FY2024, while OCF was unavailable for FY2014–FY2016. In the General Mills example, 2025 EBITDA was $3.30B, free cash flow was $2.29B and down 9.3%, five-year average operating cash flow was $3.06B, and average capital expenditures were $0.64B.
What success looks like: The bridge between earnings, operating cash flow, capital expenditures, and free cash flow is numerically supported.
Common mistake to avoid: Do not treat a high cash-flow ratio as automatically positive without checking timing, one-offs, and missing periods.
Step 8: Verify Budget and Forecast Records
Recalculate monthly cumulative budgets, actual year-to-date spend, variances, variance percentages, and the largest positive and negative gaps. In the supplied 2026 cost-center plan, projected cumulative quota was $22,490.45, placeholder actual YTD spend was $21,365.92, and portfolio variance was -$1,124.53, or -5.0%. Medical Care had the largest listed year-end gap at -$359.31.
What success looks like: Every dashboard amount matches its source workbook and the placeholder status of actual spending is clearly labeled.
Common mistake to avoid: Do not present placeholder actuals as finalized actual spending.
Step 9: Cross-Check Ratio Analysis
Use identical definitions and reporting periods when comparing liquidity, leverage, efficiency, and profitability. In the supplied Apple versus Microsoft table, Microsoft led current ratio at 1.35x versus 0.89x, quick ratio at 1.57x versus 0.70x, and net margin at 36.1% versus 26.9%. Apple led asset turnover at 1.16x versus 0.46x, ROA at 31.2% versus 16.5%, ROE at 151.9% versus 29.6%, and ROIC at 115.4% versus 44.1%. For a repeatable financial ratio analysis, verify the numerator, denominator, units, and period behind every ratio.
What success looks like: A comparison distinguishes margin strength, return intensity, liquidity, leverage, and efficiency rather than reducing them to one score.
Common mistake to avoid: Do not compare ratios calculated from different fiscal dates or inconsistent definitions.
Step 10: Validate Longitudinal Financial Records
Compare each year with the preceding year and test whether reported changes are mathematically supported. The supplied Apple FY2015–FY2025 data reports FY2025 revenue of $416.2B, net income of $112.0B, operating cash flow of $111.5B, R&D of $34.6B, share repurchases of $90.7B, dividends of $15.4B, and capital returned of $106.1B. Capital returned represented 95.2% of operating cash flow, while total assets were $359.2B and total liabilities were $285.5B.
What success looks like: Year-over-year changes, trends, and allocation percentages can all be recalculated from cited annual values.
Common mistake to avoid: Do not infer a trend from a chart until the underlying annual observations have been checked.
Step 11: Check for Missing, Placeholder, or Unsupported Data
Label zero, blank, not applicable, not reported, unavailable, placeholder, estimated, and interpolated values separately. The supplied examples include placeholder actual YTD spending, unavailable liabilities in 2010, unavailable OCF for FY2014–FY2016, and unavailable accrued-liability balances for FY2018–FY2022. Missing values should display as dashes and be omitted from trend lines. Document extraction across PDFs, scans, Word files, presentations, and spreadsheets can help locate evidence, but extracted evidence still requires validation.
What success looks like: A reader can distinguish an actual zero from a value that was never reported.
Common mistake to avoid: Never interpolate or silently replace missing financial data without explicit approval.
Step 12: Test the Charts and Deliverables
Confirm that titles match the data, axes use correct units, fiscal-year labels are accurate, percentages are not shown as dollars, totals match tables, legends identify all series, and missing periods are visibly marked. Check dual-axis scales and filters for accidental exclusions. Useful visuals include receivables versus revenue growth, allowance coverage, accrued-liability swings, OCF versus net income, red-flag intensity, revenue and margin trends, capital returns versus OCF, balance-sheet structure, and ratio comparisons. For teams building interactive financial dashboards, chart validation should be an explicit audit rule.
What success looks like: Every visual can be traced back to a table, formula, source period, and documented treatment of gaps.
Common mistake to avoid: Do not let a polished chart imply precision that the source records do not support.
Step 13: Produce an Evidence-Based Audit Report
The final report should state the scope, source files, reporting periods, rules, formulas, recalculated values, original values, differences, tolerances, missing records, assumptions, exclusions, flagged exceptions, citations, test status, and recommended follow-up actions. A defensible review is complete, cited, reproducible, traceable to source files, clear about uncertainty, and focused on exceptions. The goal is to stop being your AI’s quality control for every line and instead review what the independent process has actually flagged.
What success looks like: A reviewer can reproduce the conclusion and act on each exception without repeating the entire analysis.
Common mistake to avoid: Do not issue a pass verdict without preserving the evidence trail and unresolved limitations.
