Verification habits for finance

Ask AI to extract claims into a review aid, then verify each row yourself. Useful fields include:

• claim ID;
• exact claim text;
• claim type: source fact, calculation, definition, explanation, estimate, forecast, recommendation, or unknown;
• entity, scope, period, currency, and unit;
• named source and precise location;
• formula and inputs, if calculated;
• status: verified, unsupported, conflicting, stale, or not applicable;
• reviewer and review date;
• required correction or escalation.

The AI-created ledger is not evidence. It can miss a claim or cite the wrong location. Compare it line by line with the output, including headings, charts, footnotes, and implications.

Before checking numbers, confirm that the source itself is authoritative and current. A file named final_v7_revised.xlsx does not establish approval. Record system of origin, owner, extraction time, report parameters, version, and whether later adjustments exist.

Watch for mismatched:

• legal entities or consolidation scope;
• fiscal and calendar periods;
• currencies and exchange-rate dates;
• actual, budget, forecast, and scenario versions;
• gross and net presentation;
• cash and accrual basis;
• reported and constant-currency values;
• audited, unaudited, preliminary, and restated figures.

If two sources conflict, do not average them or choose the one that supports the narrative. Stop and resolve which source governs.

Use deterministic tools to recompute every material amount derived from inputs. Inspect formulas, not only displayed values. Verify signs, denominators, rounding, hidden rows, filters, merged cells, and excluded categories.

Common checks include:

• subtotal equals the sum of included rows;
• opening balance plus movements equals closing balance;
• assets equal liabilities plus equity, subject to presentation;
• cash-flow ending cash connects to the relevant balance-sheet amount;
• actual minus reference matches stated variance;
• percentage uses the correct denominator;
• weighted averages use correct weights;
• currency conversion uses approved rates and dates;
• totals are not duplicated across categories.

Rounding can create small differences. Define an accepted tolerance rather than letting the model call anything “approximately correct.” A tolerance is a policy or model-design choice, not an AI judgment.

Arithmetic can show what changed; it usually cannot prove why. Verify causal commentary against operational evidence or an accountable owner.

“Travel expense increased by $80,000” may be verified from accounts. “Because the sales conference moved earlier” requires event, timing, and classification evidence. “The conference generated future revenue” is a separate claim that may be unsupported.

Use explicit labels:

• Observed: directly supported by verified records.
• Reported: stated by management or another named source.
• Calculated: recomputed with a visible formula.
• Hypothesis: plausible but not established.
• Forecast: conditional on listed assumptions.
• Decision: made by an authorized person.

These labels stop a chain of inference from collapsing into apparent fact.

Verify context, not just digits. Check titles, footnotes, accounting policies, exclusions, subsequent events, and definitions. A metric can be numerically copied yet misleading if the model omits that it excludes a major cost or covers only one segment.

For charts, verify axis baseline, scale, period order, labels, and whether categories sum to the total. AI-generated chart descriptions can overstate visual trends or ignore that two axes use different scales.

For narrative, compare every adjective with an explicit criterion. “Material,” “strong,” “stable,” and “efficient” should not appear merely because they sound like finance commentary.

Match review depth to consequence. A low-stakes internal learning note may need a source check. A budget recommendation, board report, customer decision, filing, payment, or public disclosure requires established controls and qualified review.

A strong gate identifies:

• the exact artifact and version being approved;
• source files and claim ledger;
• deterministic calculation evidence;
• unresolved differences and assumptions;
• policy and disclosure checks;
• approver identity, role, and authority;
• allowed next action;
• timestamp and change invalidation rule.

Approval of draft version 3 should not silently apply to version 4. If numbers, assumptions, recipients, or attachments change, require re-review according to policy.

Asking the same model “Are you sure?” is not independent verification. Asking a second model can reveal inconsistencies, but agreement between models does not prove truth because they may share the same gaps or patterns.

Evidence comes from trusted records, reproducible calculations, authenticated systems, qualified interpretation, and accountable approval. AI can organize those checks, but it cannot certify its own answer.

Use a consistent sequence:

1. freeze the draft version;
2. inventory claims;
3. confirm source provenance;
4. reperform calculations;
5. inspect context and footnotes;
6. verify explanations and assumptions;
7. resolve or disclose conflicts;
8. apply privacy, policy, and authority checks;
9. obtain the required approval;
10. archive evidence and release only the approved artifact.

If the deadline makes these checks impossible, reduce the scope, label the artifact preliminary, or delay according to the accountable process. Urgency does not make unsupported claims safer.