Some finance tasks are merely tedious by hand. Nine of them are actively dangerous, because the manual version fails silently: it produces a number that looks reasonable, gets used in a decision, and is wrong in a direction nobody checks. Those nine are settlement reconciliation, fee categorization, per-unit COGS, inventory reconciliation, reimbursement tracking, sales tax data assembly, currency conversion, advertising allocation, and period-over-period variance investigation.
What follows is each task with its specific failure mode, because the reason to automate is rarely the hours.
1. Settlement reconciliation
How it fails: the deposit gets booked as revenue.
A marketplace payout is gross sales minus dozens of fee types, minus refunds, plus reimbursements and adjustments. Recorded as a single revenue line, the business shows a revenue figure that is materially understated and an expense structure that does not exist. Every margin calculation downstream inherits the error, and it never surfaces, because the bank balance agrees.
This is the most common serious bookkeeping error in ecommerce and it survives for years.
2. Fee categorization
How it fails: everything becomes one expense line, and diagnosis becomes impossible.
Referral, fulfillment, storage, long-term storage surcharge, removal, return processing, and advertising behave differently and respond to different decisions. Collapsed into “marketplace fees,” a fourth-quarter storage spike and a referral rate change look identical. You can see that margin fell. You cannot see why, which means you cannot act.
3. Cost of goods sold per unit
How it fails: the annual plug hides everything inside it.
Opening inventory plus purchases minus closing inventory gives a defensible yearly COGS and a useless monthly one. Product-level profitability becomes unknowable, and the blended average quietly subsidizes the worst performers with the margin of the best. The IRS covers inventory valuation and the requirement to apply methods consistently in Publication 538, and consistency is exactly what hand-maintained cost tables lose first.
4. Inventory reconciliation to the marketplace
How it fails: discrepancies get absorbed instead of investigated.
Units go missing in fulfillment centers, arrive damaged, get disposed of, or sit in transfer between locations. Reconciled by hand once a quarter, the gap becomes one adjusting entry with no explanation. Reconciled continuously, each discrepancy is a dated event with a cause and, frequently, a claim you can still file.
The manual version writes off money that was recoverable.
5. Reimbursement tracking
How it fails: claim windows close.
Lost and damaged inventory reimbursements have deadlines. A seller tracking them in a spreadsheet updated when someone remembers will miss some, and will never know how many, because an unfiled claim leaves no record. Reimbursements also need correct booking when they do arrive, since treating them as ordinary revenue overstates sales and distorts the return rate.
Check current claim policies and windows in your Seller Central help documentation, since these change.
6. Sales tax data assembly
How it fails: reconstruction under deadline pressure.
Marketplace facilitator rules mean the marketplace collects and remits in most states, but the seller still reports gross sales and deducts the facilitator-collected portion. Assembling that by hand in April, across several marketplaces and dozens of jurisdictions, is where errors enter.
Automate the data collection, monthly. Leave the filing judgment to a tax professional, and confirm current obligations with your state revenue department through the Federation of Tax Administrators directory, because thresholds and rules shift.
7. Currency conversion
How it fails: one rate gets applied to a whole month.
Sellers operating across Amazon regions frequently convert at a single month-end rate. The correct treatment applies the rate at transaction date and recognizes the gain or loss separately. A single blended rate buries real foreign exchange movement inside gross margin, so a product looks like it got cheaper to make when in fact the currency moved.
At any meaningful cross-border volume, this is not a rounding issue.
8. Advertising allocation
How it fails: the largest variable cost never reaches the product.
Ad spend posts as one lump and is spent per SKU at efficiencies that vary by an order of magnitude. Without allocation, product rankings are built on gross margin, which ignores the cost most capable of turning a good product into a losing one. Sellers reorder the wrong inventory on the strength of this, which is expensive in a way that compounds.
9. Variance investigation
How it fails: the question takes so long to answer that nobody asks it.
Profit is down eleven percent. Finding out why means comparing periods across sales, pricing, fees, returns, storage, advertising, and cost changes, then isolating which SKUs drove the movement. Done by exporting and pivoting, that is most of a day, so it happens quarterly at best and usually only after something has gone badly wrong.
This is the task most worth removing from human hands, because the manual cost is not the labor. It is the questions that go unasked. Tools have started addressing it directly: ConnectBooks includes Crunch, an AI analysis layer currently in active beta that answers questions about profit, margin, fees, and inventory against a seller’s own data. Others in the category approach the same problem through scheduled reporting or a data warehouse.
Whichever route, the standard to hold is that “why did margin move” should be answerable the same week it moves.
The pattern across all nine
Every one of these fails quietly. None of them throws an error, and the bank balance reconciles regardless. That is why they persist in businesses that are otherwise well run, and why the trigger to fix them is usually an unpleasant surprise rather than a routine review.
Two things worth knowing before you automate. Connecting tools to a disorganized chart of accounts produces wrong numbers faster, so fix the account structure first. And every automated feed needs a human reconciliation check on a schedule, because a sync that fails silently is indistinguishable from one that is working until the day the number has to be right.
If you are picking where to start, take numbers one and three. Settlement reconciliation and per-unit COGS carry most of the accuracy risk, and the other seven are improvements to a close that already ties out.
