Claude Code's 94,273-yen "loss" was a 36,636-yen profit

The seller runs a one-person online car-accessories business and built an order-management system with Claude Code to compare sales against actual purchase costs per product. The trap was structural: a single order can contain multiple line items, while the purchase order for stock is placed once per order. Applying the full order-level purchase cost to each line item doubled the cost side and made a profitable product appear to be a 94,273-yen loss.

The same month brought a second, similar error. A calculation concluded that wholesale sales were almost in the red because it omitted the shipping fees received from customers; once that revenue was added, wholesale turned out to be 2〜9% profitable. In both cases the AI was confidently wrong, and in the first case the seller nearly acted on the bad number by raising prices or delisting the product.

The seller's response was to adopt three habits: asking what unit a figure was aggregated on, hand-checking one case with a calculator, and asking what income or costs are missing. The correction — match purchase cost and sales at the order level — was written into Claude Code's memory so the same mistake would not recur, which points to a practical role for the human as the one who verifies the numbers.