Teardown

The fee column was tax: what a year of marketplace transactions actually showed.

Every report the owners pulled had a column called "fees." For two years the three of them argued about the fee rate using that column. It was a serious argument: the business runs a large volume through one big online marketplace, and the marketplace's cut is the biggest cost after what they pay their suppliers.

The column was tax.

The column was tax. It was the sales tax the marketplace collected from buyers and remitted on the sellers' behalf. The actual fee the marketplace kept had never been in their database at all. It sat in a report nobody had downloaded, one order at a time, in the marketplace's own seller portal.

I found this in the second week of a diagnostic engagement, and I want to walk through how, because the method matters more than the finding.

Where the real number lives

Marketplaces publish a transaction report: every order, refund, shipping label, claim and payout, as the platform itself recorded them. For this business it was about 700,000 rows covering twelve months, available to anyone with the main login in a few clicks. Nobody had pulled it, because nobody had needed to; the database "knew" the fees.

Matching that report to the company's own per-item records settled four arguments in one afternoon.

The four arguments one report settled

  1. 1
    The fee rate.The platform kept a little over five percent of item sales for the year. The database's column had been showing a figure roughly fifty percent higher, because it was tax. Pricing to suppliers, the case for a second sales channel, and the hiring plan had all been argued from the wrong column.
  2. 2
    One strange month.For about twelve days in one month, the platform charged a variable fee two to three times its normal rate, then dropped back. A large part came back later as fee credits. A five-figure amount did not. Nobody had noticed, because the monthly total looked only slightly high and nothing in the database flagged it.
  3. 3
    Refunds that stayed sold.The report held roughly four thousand refund rows for the year, around one percent of sales. Only a quarter of them were full refunds that the company's records showed as refunded. The rest were partial refunds on items that stayed sold: a goodwill credit for a condition complaint, shipping refunded when two items were charged separately, a lost parcel made good. Six figures a year of cost that had never had a number, because the company's system recorded only the sale.
  4. 4
    Money in equals money out.The payouts from the platform to the bank, summed, matched everything else in the report to within a few thousand dollars, which was the last few days of orders not yet paid out. That is the reconciliation anchor the bookkeeping did not have. It also meant that once the fee column was fixed, every other number in the chain could be checked against something the business did not produce itself.

Why this happens, and why it is not a software problem

The people who build the reports are not the people who negotiate the fees. Finance reports what the system produces. Operations fixes what customers complain about. The definition of "fee" belonged to nobody, so the system's label became the truth.

A new dashboard would have inherited the same column. Most of the AI pitches this company had received would have, too: an AI system trained on a database where "fee" means tax learns that fees are tax.

The fix was one sentence, written down and owned by one person: a fee is what the platform kept on the order, taken from the platform's own ledger, not from our database. Every report inherits that definition from the moment it is written, and the platform's report becomes the audit trail behind the per-item number.

What an owner can do with this on Monday

Three checks, each under an hour:

  1. 1
    Find the number everyone argues about.In most owner-led companies it is a margin, a fee, a rate of unpaid orders, or a cost per unit. Ask who owns its definition. If the answer is "the system," that is the place to start.
  2. 2
    Pull the outside record.Whatever platform, bank or carrier sits between you and your customers publishes a record of what it did. Download twelve months of it. Match it to your own records on the one key you both share (an order number, an item number, a date and amount). Where they disagree, your record is usually the one that is wrong.
  3. 3
    Write the definition and name the owner.One sentence per number, one name, one date each month when the number is read aloud with its definition. Then, and only then, decide whether any tool needs to change.

The engagement that produced this is diagnostic: no software was built to find it, and none was needed. What was needed was someone whose job for two weeks was to doubt the column.

Thirty minutes is enough to decide.

If a board, lender or buyer is about to ask what your company knows about its own numbers, that doubt is the first thing I bring. Thirty minutes is enough to decide whether it is worth a week.