📋 Overview
Your deposit comes in lower than you expected, and the settlement file has thousands of rows across a dozen fee types. Scanning it by eye catches the obvious problems and misses the slow ones: a size tier that quietly changed, a referral rate that no longer matches the category, a surcharge that started last period and will keep running.
This article covers a repeatable way to check each settlement with AI — what to export, what baseline the AI needs to compare against, a prompt pattern that produces checkable output instead of a summary, and what to confirm in Seller Central before you open a case.
🧾 Why Fee Surprises Hide in the Settlement File
A settlement report is transaction-level, not order-level. One sale can produce several rows — product charge, shipping charge, tax collected, referral fee, fulfillment fee — and the file also carries rows that have no order attached at all: storage charges, advertising charges, coupon fees, reimbursements, service fees, and manual adjustments.
Two structural things defeat eyeballing. First, volume: the fee that matters is often a few cents per unit repeated across thousands of units, not one large line. Second, period boundaries: a refund, reimbursement, or fee correction can land in a different settlement than the sale it relates to, so the file never balances order-by-order.
The goal of an automated check is not to re-add Amazon’s math. It is to answer three questions quickly: which fee types appeared that I have never seen before, which per-unit fees moved away from what I expected, and which single lines are large enough to be worth opening.
⚙️ Build the Check Once, Run It Every Settlement
1. Export the settlement flat file, not the summary
In Seller Central, go to Payments and open the Reports Repository, where completed settlement reports are listed by period. Download the flat-file (tab-delimited) version rather than the summary view — the summary shows totals by category, which is exactly the detail you need to drill into.
A good result: a file with columns for transaction type, order ID, SKU, quantity, amount type, amount description, and amount. Save the prior period’s file alongside it. Period-over-period comparison is where most of the signal is.
2. Write a fee expectation table before you prompt anything
An AI tool cannot flag an “unexpected” fee if you have never told it what you expect. Build a small table — one row per active SKU — with the fields you already know:
- Expected referral fee percentage for the product’s category. Referral fees range from roughly 5% to 45% depending on category, so use your own category’s published rate rather than a blanket assumption.
- Expected FBA fulfillment fee per unit, taken from Fee Preview rather than from memory.
- Current list price and any active coupon or promotion.
- Units on hand and roughly how long the oldest units have been in fulfillment centers.
This table is the part of the workflow that keeps working after any particular AI tool changes. Update it whenever you reprice, change packaging, or launch a promotion.
3. Make the tool compute with code, not read rows
Language models are unreliable at adding thousands of numbers read as text. They are much better at writing a few lines of code that do the adding. Use an assistant mode that executes code on an uploaded file — most major assistants offer some version of a data-analysis or code-execution mode, and supported file sizes and row limits change often, so check the current limits in the vendor’s own documentation before you plan around them.
If your settlement file is too large for the tool, a spreadsheet pivot table on amount description gives you the same totals, and you can paste that much smaller summary into the chat instead.
4. Run a two-pass prompt: arithmetic first, causes second
Mixing calculation and explanation in one request is how you end up with a confident, wrong narrative. Ask for numbers first:
You are checking one Amazon settlement period for unexpected fees. Files attached: current settlement flat file, prior settlement flat file, and my fee expectation table. Use code to compute every figure — do not estimate, and do not explain causes in this pass.
1. List every distinct amount type and amount description in the current file with row count and total value. Mark any combination that does not appear in the prior file as NEW.
2. For each fee type, show total this period, total last period, and the difference in both currency and percent.
3. For each SKU, compute referral fee as a percent of product sales and flag any SKU outside the expected range in my table.
4. For each SKU, compute the average fulfillment fee per unit and flag any SKU that differs from my expected value.
5. List the twenty largest negative line items that are not product refunds, with order ID or description.Output as tables only. Where a figure is missing from the file, say so rather than inferring it.
Then run a second prompt on the short flagged list only: “For each flagged item below, give the two or three most likely explanations and the exact Seller Central report or page I would check to confirm each one.” Keeping the passes separate means the explanations never contaminate the totals.
5. Triage flags into four buckets before acting
Most flags are not errors. Sort each one into timing (the charge relates to a different period), your own change (you repriced, ran a coupon, changed pack size), an Amazon-side change you can verify (a new surcharge tier, a size tier reclassification), or a possible error worth a case. Only the fourth bucket needs action, and you should confirm at least one flagged line manually in Seller Central before filing anything.
💡 Pro Tip: Keep a running fee dictionary — a single file listing every amount description you have ever seen in your settlements, with a one-line plain-English note on what it is and whether it is normal for your business. Paste it into the prompt each time. It turns the vague “this looks odd” into a specific “this has never appeared before,” which is the flag worth your attention.
🔍 The Flags That Usually Pay for the Exercise
- A fulfillment fee per unit that stepped up without a product change. Usually a size tier or dimensional reclassification after a remeasure. Compare against Fee Preview, and if the recorded dimensions look wrong, open a case with Seller Support asking for the unit to be remeasured.
- Referral fee percentage off your expectation. Check whether the product’s category assignment changed, or whether a price change moved it across a category-specific rate band.
- Aged inventory surcharges appearing for the first time. These apply in tiers based on how long units have been stored — 181–270 days, 271–365 days, 12–15 months, and 15 months and over — so a surcharge that starts small will grow unless the units move. Cross-check ages in the Inventory Ledger Report under Reports > Fulfillment > Inventory Ledger.
- Coupon and promotion costs larger than planned. Coupons carry both a per-coupon charge and a percentage of coupon-attributed sales, so cost scales with redemption rather than being fixed. Confirm the current structure and amounts in Amazon’s published fee schedule before budgeting.
- Adjustments and reimbursements with no order ID. Worth reading individually — these are where corrections, chargebacks, and one-off service fees sit.
⚠️ Where These Checks Go Wrong
Confusing a smaller deposit with a fee problem
Deposit size and fee totals are different questions. Amazon moved North American sellers to releasing funds seven days after confirmed delivery, so a period with slower deliveries or a shipping delay can disburse less without a single fee changing. Check your fee totals against the prior period before assuming you were charged something new.
Asking the model to reconcile order by order within one settlement
Because fees, refunds, and reimbursements cross period boundaries, per-order reconciliation inside a single file produces long lists of “missing” charges that are simply in the neighbouring settlement. Reconcile by fee type across periods, and only drill to the order level for a specific flagged line.
Filing a case for every flag the AI returns
A flag is a variance, not a finding. Cases built on unverified AI output get closed, and the time you spend reopening them erases the savings. Verify the underlying data in Seller Central, keep one fee type per case, and attach the specific rows and your expected figure.
Changing the export each time
If the columns or the date range shift between runs, period-over-period comparison stops working and every fee type looks new. Fix the export format and run the same one every settlement, even when you are curious about something else.
❓ FAQs
Should I upload the settlement summary or the flat file?
The flat file. The summary groups charges into totals, which hides the per-SKU and per-line detail the check depends on. If file size is a problem, pivot the flat file in a spreadsheet first and upload the pivot.
Can an AI tool file the dispute for me?
Treat drafting and filing as separate. AI is useful for assembling the case text — the fee type, the affected SKUs, your expected figure, and the variance — but a person should read the evidence and submit it. Submitting claims you have not verified wastes support cycles and can leave genuine errors unresolved.
How many settlement periods do I need before this is useful?
Two gives you a usable period-over-period comparison. Several months of files make the “new fee type” check far stronger, because seasonal and quarterly charges stop registering as surprises. Download and archive each settlement file yourself rather than relying on how long it stays available.
What if the fee is correct but higher than my model assumed?
That is a pricing and packaging decision, not a dispute. Feed the actual per-unit fees back into your cost model, then look at the levers that change the fee itself: pack dimensions and weight for fulfillment fees, storage duration for aged inventory surcharges, and redemption caps for promotion costs.