A founder receives an AI-written summary: “Revenue fell 18% to $4,100. The store should buy more traffic to recover the gap.” It sounds like a commercial diagnosis. In the example below, the model has compared an order total with a provider transfer from the same cohort and called the difference a decline over time.
The remedy is not a more reassuring writing style. The report needs a short, inspectable explanation of the bridge between the figures before anyone acts on its conclusion.
Read the source labels before writing the narrative
Fictional example: One store has 100 paid, delivered, single-unit orders at $50 each. At the common cutoff, ten orders have received full $50 payment refunds. The simplified cohort has no discounts, partial refunds, taxes, shipping charges to customers or other adjustments. No earlier comparable period is supplied. These are teaching assumptions, not actual store or Aright results.
A reconciled provider statement for the same transactions includes $100 of payment fees and $300 in funds still held by the provider. The $4,100 transfer is after the ten refunds, those fees and that hold. There are no other provider inflows or outflows in this constructed statement.
| Source and event | Value | What it answers |
|---|---|---|
| Order records: 100 × $50 | $5,000 | Gross order value in the chosen cohort |
| Completed payment refunds: 10 × $50 | −$500 | Revenue reversed by those refunds at the cutoff |
| Cohort value after those refunds | $4,500 | Sales revenue retained under these stated assumptions |
| Provider fees | −$100 | A separately identified payment cost |
| Provider funds still held | −$300 | Money not transferred yet; not another operating expense merely because it is held |
| Provider transfer | $4,100 | Cash transferred under this statement's defined scope |
The bridge is $5,000 − $500 − $100 − $300 = $4,100. There is no unexplained $900 hole. There are three different components whose meanings matter.
The $4,500 retained amount is also not profit. Goods, fulfilment, marketing and other costs are absent. Nor is it a general accounting-revenue conclusion. The example deliberately freezes a simple cohort so management can see the difference between orders, refunds and cash.
Diagnose the generated sentence, not only the arithmetic
The percentage in the bad report comes from ($5,000 − $4,100) ÷ $5,000 = 18%. That calculation is correct as a difference between two values. The word “fell” is not supported: there is no earlier revenue period. Calling the whole difference a sales shortfall also bundles refunds, a fee and cash timing into one commercial explanation.
A source link would not cure that error. Both the order export and the provider statement could be genuine. The failed step is the transformation from two differently defined figures into a trend and then a spending recommendation.
Google Analytics' metric schema illustrates why definitions must travel with numbers: gross purchase revenue and purchase revenue after refunded transaction revenue are distinct fields. Neither is automatically a payment-provider transfer or a complete profit measure. Check the actual instrumentation and source definition before mapping a familiar dashboard label into the business report.
Replace the paragraph with a decision-ready explanation
A corrected management note could read:
What happened in this cohort: 100 delivered orders generated $5,000 of gross order value. Ten completed full refunds reversed $500, leaving $4,500 after those refunds at the stated cutoff.
Why the transfer is $4,100: The provider statement additionally deducts $100 of fees and retains $300. The retained funds are not yet available cash; withholding them does not itself create a second expense.
What is not established: Revenue growth or decline, overall profit, the cause of the refunds, or the value of buying more traffic. There is no comparable earlier cohort or full cost model here.
Next action: Investigate the refund cases and confirm the hold's conditions and timing from the actual provider record. Keep the held $300 out of available-spend assumptions. Bring a comparable cohort and the missing costs before making a growth or budget claim.
That note gives the founder something useful without pretending that every business question has been answered. It also separates actions: customer problems may need attention immediately, while a channel-budget decision needs additional evidence.
Make the evidence inspectable at the disputed step
The person checking this report should be able to find the 100 order identifiers, the ten refund records, their amount and completed state, and the provider statement entries supporting fees, holds and transfer. Check that these belong to the intended cohort and cutoff. An unmatched refund or a statement containing another month's transactions would change the bridge.
AI can assemble a proposed explanation and identify missing information. Reconciliation calculations should be independently reproduced, and the human owner should read the source records that matter to the decision. A system-generated confidence label is not that check. Do not ask the model to reconcile a mismatch by inventing “other fees.” Leave an unresolved item visible until its source is found.
If a provider later releases the $300, the cash position changes at that later time; it is not automatically another $300 of sales from the original orders. If a further refund arrives, update the cohort version and explain the change. A report should be correctable without silently rewriting the record that supported an earlier decision.
Keep explanation separate from recommendation
Suppose a valid earlier cohort did exist. A decline might still reflect a different product mix, acquisition period, return maturity or unavailable stock. The report would need to compare those conditions before deciding that traffic is the constraint. An AI summary of two totals cannot supply that causal explanation.
Conversely, if cash backing for existing commitments is inadequate, management need not wait for a perfect growth analysis to restrain spending. The same report can support a narrow cash decision while leaving the demand question open.
Start the next reporting review with the sentence most likely to change an action. Ask which figures support it, how those figures were transformed, what competing explanation remains, and what decision is actually justified. A readable bridge and an honest boundary are more valuable than a polished report whose citations conceal the wrong comparison.