Every AI accounting tool, from every vendor, produces answers whose quality is capped by the data underneath. In ecommerce that cap is low by default, and not because sellers are careless. It is low because the data arrives from five or six systems that disagree with each other about what a sale is, what a unit costs, and what a product is called.
Model quality stopped being the limiting factor in this category some time ago. Data quality never stopped being it.
Gartner puts a number on the general version of this problem in its published data quality research, estimating that poor data quality costs organizations at least $12.9 million per year. That figure came from large enterprises already buying data quality software, so it is not a proxy for a seller doing eight figures. What it establishes is that the cost is structural rather than incidental, and that it is borne by companies with dedicated data teams. A five-person ecommerce operation has no such buffer.
A marketplace payout arrives in the ledger as one deposit. Revenue, refunds, commission, fulfillment, storage, advertising, and tax pass-through are all inside that single number and none of them are visible. Everything downstream inherits the loss, permanently, because the detail is not recoverable from the bank feed.
This is the most common and most consequential defect, and the fix is a connection that pulls settlement detail rather than a smarter model reading the bank line.
Settlement periods do not align with calendar months. Orders ship in one period and refund in the next. Reserves are withheld at a period boundary and released later. Advertising is charged against a settlement that straddles two months. Inventory sits on a container across a quarter end.
Each of these is a cutoff decision. Made inconsistently, they produce period-over-period comparisons that measure the bookkeeping rather than the business.
Two failures live here. The first is landed cost left out, so freight and duty sit in operating expenses and gross margin reads several points better than it is. The second is method drift, where FIFO is nominally the policy but a manual adjustment at quarter end effectively applies an average.
Either one makes every per-unit profitability figure unreliable, which makes every pricing and reorder decision built on it unreliable too.
The same physical product carries a different identifier in every system. This is the defect sellers underestimate most, and it is worth a worked example.
A channel that is not connected, a warehouse whose movements are tracked in a spreadsheet, a supplier whose invoices arrive by email and never reach the system. Missing data does not announce itself. It shows up as a total that is simply smaller than reality, and no tool flags an absence it has no reason to expect.
One product. Here is what the systems call it.
Nothing here is wrong in the sense of being a typo. Each was created reasonably, at a different time, by a different person, under a different platform's naming constraints. The consequence appears in the reporting.
Annual units across channels: 14,820 on Amazon, 3,110 on Walmart, 2,405 on Shopify, 980 on eBay. Total 21,315 units. Annual contribution after fees, returns, and advertising: $118,400, $27,800, $31,240, and $9,800 respectively, or $187,240 in total.
On a properly aggregated contribution ranking, this product is the second-largest contributor in the catalog. On a ranking built from the raw identifiers, it appears four separate times, none of them in the top five, and the largest single line is the Amazon alias at $118,400. Three of the four lines look like minor products that could be discontinued without much thought.
The inventory consequence is worse. True demand is 1,776 units a month. With a 45-day supplier lead time, coverage before safety stock needs to be about 2,664 units. Split by alias, each channel view shows a modest gap: the Amazon alias needs roughly 1,853 units against 1,500 on hand, the Walmart alias needs 389 against 300. Two small flags, each easy to defer, and a real aggregate shortfall of several hundred units heading into a lead time you cannot compress.
An AI analysis tool reading this data will do exactly what it was asked and rank the four lines separately, because from its point of view they are four products. It will not tell you they are the same thing. It has no basis to know.
Three reasons, and none of them are about model capability.
Speed. A wrong number produced in four seconds gets used more times before anyone questions it than the same number produced over two days in a spreadsheet.
Fluency. Output arrives formatted, ranked, and explained. Presentation quality reads as data quality even though the two are unrelated.
Absence of friction. Building the number by hand forces you to touch the inputs, and touching the inputs is how errors get noticed. Remove the manual step and you remove the accidental review that came with it.
None of this is an argument against the tools. It is an argument for fixing the inputs first, because the tools remove the last informal control that was catching input problems.
Run these on last month. Each is answerable in under an hour with a properly structured ledger and impossible without one.
Settlement tie-out. Pick one payout per channel. Do the ledger entries net to the deposit, to the penny?
Fee separation. Are commission, fulfillment, storage, returns processing, and advertising in separate accounts, or collapsed into one line?
Inventory tie-out. Does inventory on the balance sheet equal the physical count at every location, valued at landed cost?
Landed cost presence. Is freight and duty in unit cost, or in an operating expense account below the gross margin line?
SKU identity map. Does one internal SKU map to every channel identifier for that product, in one place?
Channel tagging. Can you produce a profit and loss by channel without exporting anything?
Cutoff documentation. Is there a written rule for reserves, cross-period refunds, and in-transit inventory?
Any answer of no is a specific, fixable defect with a known remedy. That is better news than it sounds, because the alternative diagnosis, "our numbers feel wrong," has no remedy at all.
The sequence matters because later steps depend on earlier ones.
Connect channels at settlement level first. ConnectBooks pulls settlement detail from Amazon, Shopify, Walmart, eBay, and TikTok Shop into QuickBooks or Xero, so a Shopify payout reaching QuickBooks Online arrives decomposed rather than as a lump sum.
Split the fee accounts second, because fee-level detail is useless if it lands in one bucket.
Establish inventory third: physical count, valued at landed cost, tracked by warehouse with FIFO applied consistently and transfers held as in transit until received.
Map SKU identity fourth. One internal identifier, every channel alias attached to it. This is the step most often skipped and the one that quietly breaks reporting for years.
Write the cutoff rules fifth, and apply them the same way every month.
Only then does the analysis layer produce anything trustworthy. SKU and channel profit reporting is the first thing that becomes possible, and it is a reasonable test of whether the previous five steps actually landed.
ConnectBooks has announced Crunch, an AI CFO built to answer questions against this reconciled data, and there is a waitlist open ahead of its release. The order of operations does not change. Whatever arrives in this category, from anyone, will be as good as the data it reads and no better.
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
