A question to an AI financial assistant needs five things stated explicitly: the period, the scope, the measure, the cost basis, and how you will check the answer. Leave any of them implicit and the software fills the gap with a default you did not choose. The number that comes back will be internally consistent and answer a slightly different question than the one you asked.
This is not a quirk of one product. It is what happens when a system with a large space of reasonable interpretations gets an underspecified request.
Language models are sensitive to how a request is written in ways that have nothing to do with meaning. Sclar, Choi, Tsvetkov, and Suhr, in "Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design," presented at ICLR 2024, found performance differences of up to 76 accuracy points across formatting variations that preserved the semantic content of the prompt, evaluated on LLaMA-2-13B. The sensitivity persisted with larger models, more examples, and instruction tuning.
That research is about benchmark tasks, not accounting, and the finding should not be over-extended. What it does establish is that phrasing is a real variable rather than a stylistic preference. In a finance context, where the same words map to several defensible definitions, the risk compounds: ambiguity in the question plus sensitivity in the system produces answers that are hard to reproduce.
Reproducibility is the property you need. A number you cannot get twice is a number you cannot put in front of a lender.
State the exact dates and the basis. "Last month" is ambiguous across a calendar month, a 4-5-4 retail month, and a settlement period that straddles month end.
Better: "for orders placed between March 1 and March 31, 2026." Better still, add whether you mean order date, ship date, or settlement date. Those three produce different revenue for the same month, and marketplace payouts almost never align to a calendar boundary.
Which channels, which SKUs, which locations. "Total sales" invites the assistant to include or exclude sales tax collected, shipping charged to customers, gift card redemptions, and cancelled orders, each of which is a defensible reading.
Say what is in and what is out: "across Amazon and eBay only, excluding sales tax collected, including shipping charged to the customer."
Revenue, gross profit, contribution, and net all get called "profit" in conversation. Define the one you want by naming what it subtracts.
"Contribution, meaning gross revenue less refunds, less all marketplace and payment fees, less landed COGS, less advertising attributed to those SKUs." Long, and worth it.
State the costing method and whether landed cost is included. FIFO and weighted average produce different numbers on the same units, and a business that has been buying at rising prices will see a material gap.
"Using FIFO cost layers, with freight, duty, and inbound handling included in unit cost."
Name the check before you read the answer. This is the part everyone skips and the one that catches errors.
"The fee total should reconcile to the sum of the settlement reports for that period. Show me the settlement lines behind it."
A seller asks: what was my profit on the Halo Series last month.
The Halo Series is 14 SKUs sold on four channels. Depending on what the assistant assumes, four defensible answers exist for March.
Reading one. Revenue less product cost at supplier invoice price, all channels, order date basis. Result: $84,120, a gross margin of 47.9 percent.
Reading two. Same, but with landed cost included, so freight and duty of $2.31 per unit are in the unit cost across 9,400 units. Result: $62,406, a margin of 35.5 percent.
Reading three. Landed cost, plus all marketplace and payment fees of $41,880. Result: $20,526, a margin of 11.7 percent.
Reading four. Landed cost, fees, and $12,340 of advertising traced to those SKUs, with $3,910 of refunds netted and 182 sellable returned units credited back to inventory at $9.86. Result: $6,070, a margin of 3.5 percent.
Four numbers, all correct, spanning a factor of thirteen. The seller who asked "what was my profit" and received $84,120 will make a reorder decision that the $6,070 figure would have prevented.
The fix is not a better model. It is a question that says which of the four you meant.
Fill in the brackets and the ambiguity largely disappears.
"For orders placed between [start date] and [end date], across [channels], for [SKU or product family], calculate [measure, defined by what it subtracts]. Use [FIFO or weighted average] cost layers with landed cost [included or excluded]. Exclude sales tax collected. Then show the settlement lines that support the fee total, and tell me if any data for this period is missing."
That last clause matters more than the rest. A finance tool that says the data is incomplete is doing its job. One that produces a confident number from a partial month is not.
Comparative. "Compare contribution per unit for SKU 4417 on Amazon versus eBay for Q1 2026, using landed FIFO cost, and show the fee breakdown for each." Channel comparisons are where most sellers find their real problem, and they depend on eBay fee detail being posted at transaction level rather than as a lump.
Diagnostic. "Gross margin fell from 38 percent in January to 31 percent in March. Decompose the change into price, cost, fee, refund, and mix effects, in dollars."
Threshold. "List every SKU whose contribution per unit was negative in March after fees, landed COGS, and advertising, sorted by units sold."
Cash. "For the next 60 days, show expected marketplace payouts by date against scheduled supplier payments and inbound freight, using actual payout timing rather than order dates."
Check the tie-out first, before reading the number. If the fee total does not reconcile to the settlements, nothing else on the screen is worth reading.
Ask the same question a second time, in different words, and see whether the answer matches. If it does not, the underlying data is ambiguous or the assistant is guessing, and both are worth knowing.
Then ask for provenance. Any figure should trace back to the orders, settlement lines, and cost layers behind it. If you cannot drill from a margin percentage to the transactions that produced it, you cannot defend that percentage to an accountant, a lender, or a buyer.
None of these questions work against summary journals. They need each payout broken into transactions and fee types, cost layers per unit including freight and duty, channel identity carried through to the ledger, and returns dispositioned rather than netted against sales. That foundation is what ConnectBooks builds, and it is what makes SKU and channel level profit answerable at all.
Preparing for that is its own piece of work, covered in more detail in how to prepare your books for an AI CFO.
ConnectBooks has announced Crunch, an AI CFO designed to answer questions against reconciled seller data, with a waitlist open ahead of its release. Whichever assistant you end up using, the discipline is the same: specify the question, name the check, and refuse to accept a number you cannot trace.
Running an e-commerce business comes with plenty of challenges, but ConnectBooks is here to make your life easier. With real-time insights, seamless integrations, and detailed tracking of your profitability and inventory, you can stay ahead of the game. Whether you’re selling on Amazon, Shopify, Walmart, TikTok or eBay, ConnectBooks helps you manage your finances with 100% accuracy and confidence, so you can focus on growing your business.
Ready to level up? Start making smarter, data-driven decisions every step of the way. Try ConnectBooks Free Today or Schedule a Demo
