Before accepting an AI-generated financial summary, check that each comparison uses the intended reporting period, unit and measurement basis. Verifying the digits alone is insufficient. A correctly extracted six-month revenue total is still the wrong denominator for a claim about three-month revenue growth.
For finance operations teams and managers reviewing business reports, the practical step is to require an intermediate table of values with their context before asking for narrative conclusions. Keep reported figures, derived calculations and unresolved comparability questions visible. Approve the comparison before polishing its wording.
This guide uses fictional financial records to demonstrate that review. It does not report a trial of an assistant, a data subscription or a financial-reporting system. The proposed checks concern preparation of an accurate summary; they do not establish a company’s financial health or explain what caused a reported change.
Give every value a period and a basis
Ask for more than a list of numbers and citations. For each value, retain the entity and scope, metric definition, start and end dates where applicable, currency, scale, source location and reporting version. A source reference should identify the table, row and column, not merely a document containing many similar values.
Distinguish amounts accumulated over an interval from balances measured at a date. Revenue for April through June answers a different question from cash at June 30. The distinction should survive copying into a spreadsheet, a chart-ready table and the final management brief.
The SEC’s current EDGAR API documentation illustrates why machine-readable data still needs contextual review. It separates facts by units and distinguishes annual, quarterly and instantaneous frames. It also warns that facts grouped into a calendar frame can have different reporting start and end dates because companies use different financial calendars.
A label such as “Q2” is therefore not enough to align two companies or two sources. Expand it to actual dates and confirm whether it means a calendar quarter, a fiscal quarter or a cumulative disclosure presented in a quarterly filing. If the required context is missing, retain the gap instead of inferring dates from the document title.
Work through the comparison before writing the sentence
Imagine a fictional services business whose approved training documents contain these figures. Every amount and date below is invented, and all revenue rows use the same entity scope and accounting basis.
| Record | Metric and period | Reported value |
|---|---|---|
| A | Revenue, January 1–March 31, 2026 | EUR 10 million |
| B | Revenue, January 1–June 30, 2026 | EUR 22 million |
| C | Revenue, April 1–June 30, 2026 | EUR 12 million |
| D | Revenue, April 1–June 30, 2025 | EUR 10 million |
| E | Revenue, January 1–June 30, 2025 | EUR 20 million |
| F | Cash balance at June 30, 2026 | EUR 9 million |
If the question is second-quarter year-over-year revenue growth, compare C with D. The calculation is (12 - 10) / 10 = 20%. If the question is first-half year-over-year growth, compare B with E: (22 - 20) / 20 = 10%.
Both answers are correct for their respective questions. Neither permits the statement “revenue growth was 20%” without specifying the period. Comparing C with E would produce a 40% decline arithmetically, but it would mix a quarter with a half-year. The calculation would not answer either intended growth question.
Ask the assistant to name the numerator, denominator and comparison dates before generating the sentence. Have the reviewer verify that selection against the source headings. A calculator can confirm the division; it cannot decide whether the denominator represents the intended business comparison.
Mark a derived quarter as derived
Suppose record C is absent. Under the example’s stated common basis, the second quarter can be derived from B minus A: EUR 22 million less EUR 10 million equals EUR 12 million. Record the two inputs and the subtraction alongside the result. Do not cite B as though the source directly reported the three-month amount.
Before doing this with real records, check that the two cumulative figures share a financial year, scope, currency and reporting basis. A later filing may revise the earlier comparative or change how a metric is presented. If the inputs cannot be reconciled, leave the derived quarter unresolved for the reporting owner.
Do not generalize subtraction to every measure. A cash balance at June 30 is not a cumulative flow to which the March balance should be added. Ratios, averages and per-share figures also need their own definitions; subtracting two reported averages does not generally produce the average for the intervening period.
In the fictional case, record F belongs in a balance comparison with another specified date. It should not be added to an earlier cash balance and labeled first-half cash. Preserve the type of measure as well as its number.
Normalize units without losing the original
A second fictional export might express record C as EUR 12,000 thousand. That is EUR 12 million, not a thousandfold difference in revenue. Keep the original value and unit, the normalization operation and the normalized amount in the review record.
Currency conversion is a separate operation. A euro amount and a dollar amount cannot be treated as comparable merely by changing the symbol. If a comparison requires conversion, the reporting owner must specify the appropriate basis and source. Do not have the assistant silently choose a rate or treat a conversion as an observed company disclosure.
The original FinQA paper, published at EMNLP 2021, separates supporting financial facts from executable reasoning programs. Its error analysis includes an example involving billion-to-million conversion even when relevant facts were retrieved. It also distinguishes evaluation of the numerical answer from evaluation of the calculation program.
That is historical research on a curated financial-question dataset and older models, not an error-rate estimate for today’s assistants. It informs a modest review principle: inspect the chosen inputs and operations alongside the final value. This article does not import the paper’s model-performance percentages into a workplace claim.
Keep the reporting version visible
A comparison may legitimately use revised historical figures or figures as originally reported. Those answer different information questions. Identify the chosen basis before mixing documents from different publication dates.
In her September 2018 CFA Institute practitioner article, Mohini Singh illustrates how a revenue-recognition change altered previously reported Microsoft comparatives. The example uses Calcbench to inspect original and revised disclosures. It is a historical demonstration of reporting-basis differences, not evidence about a current assistant or independent testing of a data product.
The current Calcbench function reference separately exposes calendar-period, fiscal-period and as-originally-reported retrieval functions. This is vendor documentation showing that such choices exist in that product; no function or subscription was tested here.
For an internal monthly pack, record which close or revision the figures come from. If a new version changes an input, recalculate the affected comparisons and revisit the prose. Retaining a previous percentage while updating its source table can leave the report internally inconsistent.
Review the claim that leaves the table
Before releasing the summary, check the claim in both directions. Can the reviewer reconstruct it from the selected sources and calculations? Does its wording preserve the period, unit, scope and revision basis that make it true?
For the fictional example, an acceptable sentence is: “Revenue for April–June 2026 was EUR 12 million, up 20% from the same three months in 2025.” That statement reports a comparison. The records do not show whether pricing, customer growth or another cause explains it.
Our chart-comparison guide covers choosing the measure a visualization should communicate. This review comes earlier: establish which reported values can support that measure. Keep prospective judgments separate too; the forecasting guide addresses claims about outcomes not yet known.
Start with one recurring financial summary. Require the contextual value table, expose every derived period and unit conversion, and have the reporting owner resolve incompatible inputs before narrative approval. A concise report can omit intermediate detail from its main page while retaining enough evidence for someone else to reconstruct the comparison.