Before AI Explains Your Sales, Make the Numbers Reconcile

Start with the decision the report must support

"Explain our sales" leaves too much room for interpretation. The assistant can produce a long account of product trends without answering the question that matters to the business. A more useful brief is specific: should we continue the current promotion, investigate a weakening category, or prepare a controlled increase in acquisition spending? Each question needs different evidence — a promotion decision needs product revenue after discounts, the cost of serving those orders, returns as they become observable and an appropriate comparison; a stock decision also needs availability, purchasing commitments and replenishment timing. A sales export alone cannot answer both.

Write the decision at the top of the report specification: the choice, the person who will decide, the review date and the information that could change the answer. This keeps the project from expanding into a dashboard for every possible question. Choose the comparison deliberately, too — last week is convenient, but a different promotion, stock position or holiday can make it a poor reference, and a prior-year comparison can carry its own calendar and assortment differences.

The first useful result from this exercise may simply be a missing-data finding. If delivery costs are unavailable by market, the team cannot confidently compare the contribution of two markets. That finding creates a concrete task for the reporting owner. A plausible estimate should stay labelled as an estimate, with its effect on the decision shown separately rather than quietly folded into the total.

Give every number a definition

A metric name is only the beginning of a definition. "Revenue" can describe ordered merchandise, invoiced sales, amounts after adjustments, or a total that includes delivery charges and taxes. Create a short metric dictionary shared by finance, commerce and marketing — the formula, included records, timing rule, currency, authoritative system and owner, plus exclusions. Otherwise two people can follow apparently identical instructions and still produce different answers.

The cost boundary deserves particular attention. Warehouse rent and a per-parcel fulfilment charge behave differently; a retained agency fee and a commission on each sale behave differently too. List exactly which costs your management figure includes, and avoid calling it net profit when overhead or financing remain outside it. Decide how delivery appears — customer delivery income alongside the full delivery cost, or a correctly calculated subsidy within an agreed model, but never both, or you deduct the customer's contribution twice.

Apply the same discipline to returns. A product refund that has already reduced net revenue must not become another deduction in a later "returns cost" field — return transport, inspection and unrecoverable product value are separate components. And record what "new customer" means before using acquisition metrics: a new email address is not necessarily a new buyer, and if your identifiers cannot reliably connect prior orders, document that limitation rather than presenting a precise figure the records cannot support.

Reconcile before you interpret

Consider a hypothetical retailer comparing two weeks, using the same management definitions throughout and excluding tax. The amounts below are invented for teaching purposes — they are not a benchmark or an observed case result.

Weekly management measure	Previous week	Current week
Net merchandise revenue	€90,000	€100,000
Add net delivery income	€3,600	€4,000
Revenue included in this model	€93,600	€104,000
Less cost of goods, payment fees, fulfilment and returns	€48,600	€56,000
Contribution before advertising	€45,000	€48,000
Less advertising spend	€21,000	€27,000
Contribution after advertising	€24,000	€21,000

Net merchandise revenue increased by roughly 11%. Contribution after advertising fell by €3,000, or 12.5%. The bridge is easy to inspect: revenue including delivery rose by €10,400, the specified non-advertising costs rose by €7,400 — leaving €3,000 more contribution before advertising — and advertising spend then rose by €6,000, leaving €3,000 less after advertising. This does not establish that the extra advertising caused the decline; the periods may contain a different product mix, returns from earlier orders or different customer behaviour. It gives the team an accurate description of the period and a reason to investigate.

The arithmetic changes the conversation: a higher sales total did not produce a higher amount remaining under this cost model.

Reconciliation is what makes that table trustworthy. Keep order, line-item, product and refund identifiers, and flag unmatched records for investigation instead of guessing a match. Separate missing values from genuine zeros — a blank product cost means the calculation is incomplete, while a zero cost asserts that none applies. Check what happens when two tables meet: attach an order's €2 payment fee to each of its three product lines and sum the result, and the report claims €6 of fees. Validate the expected relationship at every join — one-to-one, one-to-many or many-to-one — and compare row counts and totals before and after.

Keep a short control record for each transformation: the export's record count and totals, what filtering excluded and why, what joining left unmatched, and whether aggregation returns to the checked overall amount. Treat a reconciliation difference as an item to explain, not a figure to balance away quietly.

Turn checked numbers into a brief

The important separation is between calculation and interpretation. Store the formulas and transformation rules outside the conversation, run them on defined inputs, reconcile the outputs, and only then give the assistant an approved summary — a compact table of agreed measures, the metric definitions, a reconciliation result with its exceptions, and a short record of known changes such as promotions or stock interruptions. The full transaction history stays available for authorised investigation, but the first analysis does not need to read every field in every export.

Deterministic code repeats the same rule consistently. It can also repeat a wrong rule consistently.

A useful brief keeps three things visibly separate: what changed, what could explain it, and what the team should do next. An observation might say contribution fell by €3,000. An interpretation might suggest a higher share of costly-to-ship products. A proposed action might ask an analyst to compare contribution by shipment profile before the next promotion decision. Only the observation is established by the table — the rest needs its own check against the records.

Before scheduling anything, reproduce a previously reviewed period and investigate any difference instead of adjusting the expected answer until the process passes. Then test it against a duplicated export, a partial refund, a missing product cost and a stale input, and specify the correct behaviour for each: a duplicate should never raise revenue, a missing material cost should produce an explicit limitation, and a late input should block the report from declaring a complete period.

Once the process is dependable, a short checklist keeps it that way:

• Choose one recurring decision. Put the choice, owner and review date at the top of the brief.
• Agree the metric dictionary. Have finance, commerce and marketing confirm formulas, timing and cost boundaries before comparing results.
• Preserve the record identities. Keep the keys needed to connect orders, items and adjustments, and make unmatched records visible.
• Reconcile before interpreting. Check totals, row relationships, missing costs and duplicate inputs before drawing conclusions.
• Give AI a checked analysis pack. Supply compact tables, definitions and evidence references, and require observations, hypotheses and actions to stay distinct.
• Test normal and broken inputs. Reproduce a reviewed period, exercise real failure cases and keep rerun behaviour inspectable.

The strongest AI trading report may contain fewer conclusions than the first impressive demo. It knows which numbers reconcile, which comparisons are appropriate and which questions remain open — and that is what makes its next recommendation worth acting on.