Ops & Supply

How to reconcile amazon seller settlements with AI

Ops & Supply3 AI tools7 steps6 friction points

Want an agent to handle this on your live data? See the Starch version →

Amazon Seller Central deposits a settlement into your bank account every two weeks — but the number it lands with rarely matches what you expected. Each settlement report bundles together product sales, FBA fulfillment fees, referral fees, advertising charges, storage fees, reimbursements, and adjustments into a single flat file. Reconciling that file means matching each line to your orders, your ad spend, and your accounting records so you know whether you actually made money on each SKU.

The workflow feels like a natural fit for AI because the hard part is pattern-matching and arithmetic across structured data, not creative judgment. You have a flat file, a chart of accounts, and a set of rules — things like 'referral fee should be 15% of sale price for this category' or 'this reimbursement maps to this original order.' An LLM can read a CSV, apply categorization logic, flag anomalies, and draft journal entries. That's why operators keep reaching for ChatGPT or Claude when settlement day rolls around.

General-purpose AI tools — ChatGPT, Claude, Gemini — can meaningfully help here. They can parse a pasted settlement report, classify fee types, calculate net proceeds per SKU, compare expected versus actual fees, and produce a draft reconciliation summary or journal entry memo. The capability is real. The catch is everything that surrounds the raw language model: getting your data in, keeping outputs consistent, and making the workflow repeatable the next time a settlement drops.

Skip the manual work → See the Starch version

Ops & Supply3 AI tools7 steps6 friction points

AI walkthrough

How to do it with AI today

A practical walkthrough using ChatGPT, Claude, and other off-the-shelf LLMs — what they're good at, what you'll have to do by hand.

Tools that work for this

ChatGPTClaudeGemini

Step-by-step

1 Download your settlement report from Seller Central as a flat file (CSV). Also pull your corresponding order report and advertising spend report for the same settlement period — you'll need all three to do a complete reconciliation.

2 Open ChatGPT or Claude and paste in the first 200–300 rows of your settlement CSV along with a plain-English description of your fee structure: referral fee percentage by category, FBA tier rates you're expecting, and any active promotions that should appear as adjustments.

3 Prompt the LLM to group all line items by transaction type — sales, FBA fees, referral fees, advertising, storage, reimbursements, and other adjustments — and produce a summary table showing total dollars in each bucket.

4 Cross-reference the fee summary against your expected rates. Ask the LLM to flag any referral fee that deviates more than 1% from the expected rate for that ASIN's category, and any FBA fee that doesn't match the published size-tier rate for that SKU's dimensions.

5 Paste in your advertising spend report and ask the LLM to reconcile ad charges in the settlement against your Sponsored Products spend for the same period, noting any discrepancy between what Seller Central charged and what the ads console reported.

6 Ask the LLM to draft journal entry line items mapping each transaction type to your chart of accounts — revenue, COGS, FBA fulfillment expense, advertising expense, and reimbursement income — formatted for manual entry into QuickBooks or your accounting system.

7 Copy the output into a spreadsheet, do a final sanity check that the net settlement amount matches your bank deposit, and file the reconciliation document with a date stamp.

Prompts you can copy

Here is my Amazon settlement report CSV. Group all line items by transaction type and produce a summary table showing total dollars in each bucket: sales, FBA fees, referral fees, advertising, storage, reimbursements, and adjustments.

My referral fee should be 15% for grocery ASINs and 8% for health and beauty. Flag every row in this settlement where the referral fee percentage deviates from the expected rate by more than 1%.

Compare the advertising charges in this settlement report to the Sponsored Products spend data I'm pasting below. List any line where the amounts don't match and note the dollar discrepancy.

Based on this settlement summary, draft journal entry line items mapping each fee type to the following chart of accounts: [account list]. Format them as debit/credit pairs ready for manual entry into QuickBooks.

The settlement shows $3,240 in FBA fees. My shipment had 180 units across three size tiers. Here are the published FBA rates for each tier. Calculate whether the fees charged match what I should have been billed and flag any overcharge.

Reality check

Where this gets hard

The walkthrough above works — until your numbers change, the LLM hallucinates, or you have to re-paste everything next month.

No live connection to Seller Central or QuickBooks — every run starts with downloading files manually, trimming them to fit the context window, and pasting them in fresh.

Settlement reports routinely exceed 1,000 line items; most LLM context windows handle only a few hundred rows reliably before truncation silently drops transactions from the analysis.

Output formatting drifts between sessions — the journal entry structure you carefully prompted last settlement period isn't guaranteed to look the same this period, so downstream paste-in to QuickBooks requires re-checking every time.

Nothing persists between runs — the categorization rules, fee-rate expectations, and account mappings you defined last month live only in your chat history, not in a reusable workflow.

The LLM has no way to verify whether the settlement amount actually hit your bank account — you're manually cross-referencing Plaid or your bank statement separately, outside the AI session.

Ad spend reconciliation requires juggling three separate reports in the same prompt; coordinating context across the settlement file, ads console export, and order report in one session is fragile and easy to get wrong.

Tired of the friction?

Starch runs the whole workflow on live data — no copy-paste, no hallucinated numbers, no re-prompting next month.

See the Starch version →

Starch alternative

The same workflow on Starch

Starch is an agentic operating system — the layer where an agent builds and runs the software your reconciliation workflow depends on, connected to your live Seller Central data, bank accounts, and accounting records, so the work happens continuously rather than as a monthly manual session.

The Amazon Seller Dashboard (live in the App Store) pulls your settlement and order data directly — no CSV downloads, no copy-paste. Your reconciliation surface reflects the actual settlement figures, not last week's export.

Connect Plaid once and Starch syncs your bank transactions on a schedule, so the agent can automatically verify that each settlement deposit matches what landed in your account — flagging discrepancies without you running a separate bank check.

Connect QuickBooks from Starch's integration catalog and describe your chart of accounts in plain English; the agent drafts and posts reconciling journal entries against real invoice and payment data, not a manually pasted summary.

Describe your fee-rate rules once — 'flag any referral fee that deviates more than 1% from category rate, and any FBA charge that doesn't match the published size-tier rate' — and Starch runs that logic against every settlement automatically, surfacing exceptions as alerts rather than requiring you to re-prompt.

The Amazon Channel Manager app (coming soon — request beta access) will add per-SKU true profitability tracking that layers advertising spend, FBA fees, storage, and referral fees into a single margin view, so reconciliation and profitability analysis happen in the same surface.

The workflow persists and runs on schedule — next settlement period, the same reconciliation runs automatically against fresh data, outputs land in your dashboard, and exceptions surface as notifications rather than a task you have to remember to do.

Get closed-beta access →

Toolkit