A general-purpose chatbot can explain your settlement report, write the formula you need, and draft a reconciliation checklist. It cannot close your books. The blocking problems are not intelligence problems: the file is often too large to upload intact, the data required to compute cost of goods sold is not in the file at all, results change between runs, and nothing gets posted to your ledger at the end. Those limits do not improve with a better prompt.
That is worth saying plainly, because the failure is quiet. The output looks like accounting. It is formatted like accounting. Whether the numbers are right is something you have to check by hand, which removes most of the reason you tried.
A settlement report from a marketplace is not a summary. It is a transaction ledger where a single order generates several signed rows: item principal, item tax, commission, fulfillment fee, promotional rebate, and more, each tagged with an amount type and description. A 14-day settlement for a seller doing roughly $430,000 in gross sales runs somewhere around 200,000 rows.
Take a real-shaped example. 214,880 rows, 62.4 MB as a CSV, covering two weeks:
Net deposit: $255,928.14.
Revenue for the period is $433,311.58, being principal plus shipping charged. Sales tax collected and remitted nets to zero and is not revenue. The gap between the deposit and the revenue is $177,383.44. Every dollar of that gap is a fee, a refund, a rebate, or a pass-through, and each belongs in a different account.
OpenAI's own File Uploads FAQ sets a hard limit of 512 MB per file, caps text and document files at 2 million tokens, and states that CSV files and spreadsheets cannot exceed approximately 50 MB depending on row size. A 62 MB settlement file sits on the wrong side of that line.
So you split it. Now you are managing four partial files and reconciling their subtotals, which is the spreadsheet work you were trying to avoid, plus a new opportunity to lose rows at the seams.
When the tool runs Python over the file, arithmetic is reliable. When it reads a sample of rows and reasons about them in prose, arithmetic is not. The interface does not always make the difference obvious, and the prose answer reads exactly like the computed one.
The tell is a total that is close but not exact. A figure of "approximately $255,900" is not a reconciliation. Reconciliation means the settlement total agrees to the penny with the bank deposit, or you know why it does not.
Ask for a fee breakdown and a model will produce clean categories. Some of those categories will be its own groupings rather than the marketplace's. Storage and long-term storage may collapse into one line. A refund commission credit may be netted against commission instead of shown separately. Digital services fees may be dropped into "other."
The output is tidy and internally consistent. It just does not match the report, which means it will not tie out next month either.
Of the $26,510.88 in refunded principal above, a substantial share belongs to orders that shipped and were recognized in the prior settlement. Matching a refund to its original order requires the earlier file. The model only has what you uploaded, and it will not tell you it is missing context, because it has no way to know the earlier file exists.
The same problem hits reserves, deferred transactions, and anything the marketplace holds back and releases later.
This is the one that ends the exercise. A settlement report contains prices, fees, and quantities. It contains no unit cost, no landed cost, no inventory layers, no freight, no duty. Cost of goods sold cannot be derived from it under any prompt.
A model asked for gross margin from a settlement file will do one of two things. It will tell you the data is not present, which is correct and useless. Or it will apply an assumed cost percentage, produce a margin figure, and mention the assumption in a sentence you may not read. The second behavior is where sellers get hurt, because the number that survives into the spreadsheet is the number, not the caveat.
Run the analysis twice and totals can differ, categories can regroup, and edge-case handling can change. There is no version control and no audit trail. If you cannot reproduce a number, you cannot defend it, and a lender or an acquirer will ask you to.
At the end, you have an answer in a chat window. Your general ledger is unchanged. No journal entries, no clearing account cleared, no COGS booked, no inventory relieved. Whatever you learned still has to be entered by a person, and the entering is most of the work.
That 214,880-row file contains order-level detail. Uploading it to a consumer account means thinking about retention, about whether the content can be used to improve models, and about what your own policy says. OpenAI publishes different data-usage terms for consumer products than for its business offerings. Read them before uploading customer-level financial records, and check what your accountant or your acquirer expects.
None of the above makes the tool worthless. It makes it a different tool than the one the headline suggests.
Explaining the report. Paste twenty rows and ask what a specific amount description means and which account it belongs in. This is fast, accurate, and better than searching a help center.
Writing the query. Ask it for the pandas code, the SQL, or the pivot configuration that groups a settlement by amount description and nets the signs correctly. You run the code yourself, on the full file, and you can read what it does.
Drafting process. A month-end reconciliation checklist for a four-channel seller, a memo explaining a variance, a first draft of a costing policy for your CPA to correct. Good use of a language model.
Thinking out loud. Working through whether a rebate is a contra-revenue item or a selling expense, then taking the reasoning to a professional rather than acting on it. Useful, as long as the professional stays in the loop.
The reason a chatbot cannot close ecommerce books is that the books need data the chat window never receives: cost layers, inventory movement, channel identity, and a connection that writes entries into the ledger.
That is a pipeline, not a prompt. ConnectBooks pulls settlement detail from Amazon, Shopify, Walmart, eBay, and TikTok Shop, decomposes each payout into its transaction and fee types, applies FIFO costing against tracked inventory, and posts into QuickBooks or Xero. The Amazon accounting integration is the specific mechanism for the file described above. Once that exists, profit by SKU and by channel becomes a report rather than a project, because cost and revenue are attached to the same records.
ConnectBooks has also announced Crunch, an AI CFO intended to answer questions against that reconciled data rather than against a raw file. It is not open to customers yet, and there is a waitlist ahead of its release describing what it will do.
Hand a general chatbot a settlement report and you will get a confident, well-organized answer within a minute. Check it against the report and you will usually find something regrouped, something dropped at a file boundary, or a cost assumption you did not supply. Sometimes all three.
Use it for the parts it is good at: explanation, code, drafting, and thinking. Do not use it as the thing that produces the number you act on. The distinction between a tool that helps you understand your books and a tool that keeps your books is not a matter of degree, and it is covered further in what AI bookkeeping software actually automates.
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
