Before AI: Five Checks When Finance Reports Disagree

Hypothetical example. The company, figures and roles in this article are invented to illustrate the checks.

A midsized company with a US parent and a UK subsidiary has closed Q3. FP&A wants an AI tool to draft first-pass commentary on revenue against budget for the management report. This is one of the tasks covered in Six Finance Reporting Tasks Worth Automating. The proposed AI input file, a general ledger export, shows Q3 revenue of $12.86M. The management report shows $12.40M.

The $12.90M budget covers external-customer revenue for the same two entities and the same quarter as the management report. Measured against the management report, revenue is $0.50M below budget. Measured against the export, it is $0.04M below. If the wrong figure goes in, the commentary tells a different story from the one the board will see.

The checks below trace the gap to one item:

Neither figure is wrong; they answer different questions. The export shows revenue before consolidation eliminations. The management report shows consolidated revenue from external customers. The right input is the one prepared on the basis the output requires, which is not necessarily the newest file or the one from the system you trust most. The five checks help you find that basis and decide whether the data supports the task. The decision covers this one task and dataset, not all of the company’s data.

One prerequisite comes first: confirm that this data may be used with this AI tool under your company’s policies. Customer-level revenue detail may carry restrictions.

1. Metric definition

Compare the written definition behind each figure:

• what it includes and excludes
• its recognition basis (recognized revenue, billings or bookings)
• whether it is gross or net of discounts
• how it treats sales between group companies

Look for a documented definition, such as a reporting policy, a glossary or a report footnote, and for the account range each figure actually uses. “What this report has always shown” is not a definition; it is a finding.

The person accountable for the metric resolves the definition, often the Controller for reported revenue. Building a file does not by itself give someone that authority, although the preparer may also be the accountable owner. AI can draft commentary on a defined metric, but it does not approve the definition. If the metric the commentary discusses has no agreed definition, the task should not go ahead.

In the example: Both figures are recognized revenue from the ledger, so the recognition basis is not the cause. The management report’s definition specifies revenue from external customers, and nobody had checked the export’s account range against that wording.

2. Scope and period

Compare:

• entities and consolidation level
• period and cutoff
• data version (pre-close, locked close or post-close adjustments)
• currency and translation rates
• units and sign conventions

The evidence is each extract’s parameters: entity list, period, run date, rate table and ledger view. Preparers confirm their own parameters, and the consolidation owner confirms which scope fits the task. A mismatch usually means rebuilding the extract rather than adjusting numbers by hand.

In the example: Both figures cover the same quarter and the same locked close, in US dollars at the consolidation’s rates. The scope is what differs. The export sums both entities’ ledgers before eliminations. The management report is consolidated: this company eliminates the US parent’s Q3 sales of services to the UK subsidiary, so consolidated revenue includes outside customers only.

3. Source and traceability

Trace each figure back to its source. That means the ledger or subledger, the mapping from accounts to report lines, transformations, adjustments such as eliminations and top-side journals, and manual overrides. Useful evidence includes the mapping table, elimination journals, query steps, and a log of manual adjustments with the preparer and reason for each. In a file built with Power Query, transformations appear in the Applied Steps list. Values typed directly into a worksheet need separate review. Each file’s owner traces their figure, and the Controller confirms that the adjustments are approved.

A figure nobody can trace, or an override nobody can explain, should not feed the task. COSO’s 2026 guidance on generative AI lists data quality, source and completeness as a risk: unverifiable or incomplete inputs produce unreliable outputs that are hard to detect once they spread (COSO).

In the example: The export maps every revenue account to “Revenue.” That includes the intercompany revenue account, which holds $0.46M, all in the US parent’s ledger. As a result, the export puts the parent at $9.41M, against $8.95M from external customers. The management report figure traces to the consolidation workbook, where an elimination entry removes the same $0.46M. Neither figure carries manual overrides.

4. Reconciliation and exceptions

Build a bridge from one figure to the other, with support for each reconciling item. Record who accepted it, and list any open exceptions with an owner and a due date. The preparer builds the bridge, and the metric owner accepts it or requires a correction. Material items escalate through your existing review. There is no universal tolerance for an unexplained remainder. Whether it matters depends on what the output will say and on the thresholds your close and review already apply.

In the example: The bridge above closes with nothing unexplained. The UK subsidiary’s recorded intercompany purchases agree to $0.46M, and the Controller accepts the bridge.

How you act on a finding like this matters:

• One-off manual correction. For example, removing the intercompany line from this quarter’s export. This can work, but it needs documentation and review, and someone has to remember it next period.
• Repeatable correction. Fix the mapping, the query or the extract, or build the input from the consolidated view. Then check each period that the input total equals consolidated revenue. COSO describes similar controls for data transformation: rules written in plain language, named owners, and reconciliations between raw and transformed data.
• AI assistance. AI can suggest where to look. It does not establish or approve the correct figure.

Excluding the intercompany account works here only because the investigation established that it explains the entire difference. In other cases, excluding intercompany accounts will not necessarily reproduce a full consolidation.

5. Context required for the task

Agreement between the figures makes the input consistent, but not sufficient. Compare what the output needs with what the input holds, and keep two kinds of evidence apart:

• What changed: movement by entity, product line, customer group or region. This can be calculated from the data.
• Why it changed: a delayed contract, lower volume, a price change or currency translation. This evidence lives in deal records, operating data and the knowledge of business owners. Matching totals say nothing about it.

FP&A defines what the output needs. Business owners confirm the causes. The reviewer who signs off confirms the commentary is supported. If any part of the requested output asserts causes the input cannot support, cut it or hold it. NIST’s Generative AI Profile calls the relevant risk confabulation, which it describes as “confidently stated but erroneous or false content.” It also notes that outputs can include made-up logic that appears to justify an answer (NIST AI 600-1).

In the example:

The input supports two statements: revenue is $0.50M below budget, and $0.35M of the shortfall is in the UK. It supports nothing about why. It has no product or customer detail and no notes from business owners. If budget and actual used different GBP/USD rates, the input also cannot separate currency translation from lower sales in local currency.

Working table

Three possible decisions

In none of these decisions does AI choose the revenue figure. For a wider view of where your finance team stands: AI Readiness Assessment