Step 1: Identify the source data and deliverables
What to do: Build an inventory of source files, source URLs, dataset versions, analysis files, reports, spreadsheets, dashboards, PDFs, charts, assumptions, formulas, and external references. An independent AI auditor is useful when the agent that created the analysis should not be the only system checking it.
Success looks like: Every deliverable has a named source set and a clear version or reporting period.
Common mistake to avoid: Do not audit a report without preserving the exact inputs used to produce it.
Step 2: Recompute key financial figures independently
What to do: Recalculate totals, subtotals, growth rates, margins, ratios, forecast variances, budget-to-actual differences, category totals, vendor aggregations, and derived claims. A source-grounded financial analysis should reproduce the reported result from the cited inputs.
Success looks like: Independent sums and formulas reproduce the reported results or expose a documented difference.
Common mistake to avoid: Never accept a figure simply because it appears in a professionally formatted report.
Step 3: Trace every figure to its evidence
What to do: For every material number, record its source file, row or record, field or column, formula, comparison base, time period, units, currency, and status as reported, calculated, estimated, or inferred. For related processes, use financial reconciliation to make mismatches visible rather than burying them in narrative.
Success looks like: Another reviewer can reproduce the number without asking the original model author for clarification.
Common mistake to avoid: A citation to a whole workbook is not enough when the material value comes from one specific row or field.
Step 4: Test methodology and comparability
What to do: Check consistent periods, like-for-like categories, annual prepayments, cash spending versus accounting expense, forecasts versus budgets and actuals, ending versus average balances, and gross versus net figures. If the task involves multiple records, cross-check financial records before interpreting a trend.
Success looks like: Every comparison has a complete and appropriate denominator, scope, period, and accounting convention.
Common mistake to avoid: Do not call a forecast a legally approved budget revision without evidence of the formal change.
Step 5: Test the deliverable itself
What to do: Verify that tables include all rows, charts reference the complete range, formulas update when source rows change, dashboard totals match underlying tables, PDF and spreadsheet figures agree, and citations resolve to the correct evidence. You can also verify AI output against source documents before it reaches decision-makers.
Success looks like: Numerical accuracy and presentation accuracy both pass, including visualizations and warnings.
Common mistake to avoid: Do not judge a chart by appearance alone; inspect its source range and update behavior.
Step 6: Issue a verdict and preserve evidence
What to do: Mark each material item Pass when it reconciles, Partial when it needs qualification, Fail when it is incorrect or unsupported, and Remove or restate when the claim should not remain. Reusable rules can be captured in reusable AI workflows so recurring corrections become persistent audit rules.
Success looks like: The final report contains a verdict, calculation, source reference, limitation, and reviewer-readable evidence trail.
Common mistake to avoid: Do not replace a failed claim with a more precise-looking estimate unless the estimate is explicitly supported.
