AI in Accounting: The Questions Sellers Keep Asking

AI in accounting means software that reads your transactions and drafts a categorization, a reconciliation match, or a written explanation of a variance for a human to approve. For an ecommerce seller the practical question is narrower: can a model look at a settlement report, a bank feed, and an inventory ledger and tell you something you did not already know? Sometimes. The questions below are the ones sellers keep asking, answered without the vendor gloss.

Can AI do my ecommerce bookkeeping on its own?

No. It can draft most of the entries, and it can draft them faster than a person, but someone still has to own the mapping and sign off on the close. The failure mode is not that the model makes up a number. It is that the model categorizes an Amazon reimbursement as revenue for six months because nobody told it otherwise, and the error compounds month after month.

The IRS does not care which software drafted the entry. Publication 538 requires a business that must account for inventory to use an accrual method for purchases and sales, and it requires the method to be applied consistently year to year. A model that flips between cash and accrual treatment because two months of data looked different is a compliance problem, not a productivity gain.

What does AI do well in a seller’s books?

Three things, in rough order of how much time they save.

Categorization. Given a chart of accounts and a few hundred labeled transactions, a model will assign new bank lines to the right account with high accuracy and flag the ones it is unsure about. Bookkeepers who used to categorize every line now review exceptions.

Matching. Payout reconciliation is pattern matching at scale: this Shopify deposit equals the sum of these 214 orders minus these fees minus these refunds. Shopify’s own payout documentation describes the activity report that breaks a payout into transactions, fees, and the ending balance. A model reading that report next to the bank feed can propose the match and show its work.

Narration. Ask why gross margin dropped four points in July and a model that can query your data will list the three SKUs where landed cost rose and the one channel where refund rate doubled. That is the CFO-style question a small seller rarely gets answered, because nobody has the hours to run the analysis by hand.

What does it do badly?

Anything that depends on a fact outside the data it can see. A model does not know your supplier raised prices on the last container unless the bill is in the system. It does not know a marketplace changed its referral fee structure unless it reads the fee schedule, and it will apply last year’s rate if you let it. It also struggles with intent: a $3,000 transfer to your personal account might be an owner draw, a loan repayment, or a mistake, and the data alone cannot tell it which.

The other weak spot is inventory. Cost of goods sold at the SKU level requires a valuation method, real landed cost per unit, and a record of every unit received and shipped. A model can compute FIFO layers once that data exists. It cannot conjure the data from a bank feed.

Is an “AI CFO” a real product or a rebrand?

Both, depending on the vendor. The honest version is a layer that sits on top of clean, transaction-level books and answers questions about margin, cash, and inventory in plain language. The rebrand version is a chatbot bolted onto a dashboard that was already there.

The tell is what the product needs underneath it. ConnectBooks, which syncs Amazon, Shopify, Walmart, TikTok Shop, and eBay into QuickBooks Online, QuickBooks Desktop Enterprise, or Xero, has an AI CFO called Crunch in active beta. Its premise is that the AI can only be trusted because the books beneath it reconcile at the settlement and SKU level first. That ordering, books first and model second, is the right way to evaluate any of these products. A model reading bad books produces bad advice with a straight face.

Will AI catch fraud or errors my bookkeeper missed?

It catches anomalies, which is a narrower claim. If a vendor is paid twice, or a refund is issued for an order that was never placed, or a fee line appears that has never appeared before, a model trained on your history will flag it. Whether the flag is fraud, an error, or a legitimate one-off still takes a human to decide.

Where it helps most is the 1099-K reconciliation. The IRS page on Form 1099-K says marketplaces must report gross payments over $20,000 in more than 200 transactions and may report lower amounts too. Gross on the form will not equal net in your bank. A model that has the settlement detail can bridge the two and show every fee and refund in between, which is the exact worksheet an auditor asks for.

Does using AI for bookkeeping change what I owe?

No. Tax liability comes from the transactions, not the tool that recorded them. What changes is how defensible the records are. Better categorization and a documented reconciliation trail make an examination shorter. Sloppy automation that nobody reviewed makes it longer. For questions about method, nexus, or whether a specific expense is deductible, the answer is your CPA or the state department of revenue, not a chatbot.

How should a seller start?

Get the books to reconcile without AI first. Every marketplace payout should tie to a bank deposit, every fee should have its own account, and COGS should post from real unit costs. Once that holds for three consecutive months, turn on the assistive features one at a time: categorization suggestions, then match proposals, then narrative reports. Review the exceptions the model raises for a full month before trusting any of it unattended.

Ecommerce keeps growing as a share of retail. The Census Bureau’s second quarter 2026 release put online sales at $340.2 billion, 17.1 percent of total retail, up 12.2 percent year over year. More volume through more channels means more lines to categorize and more payouts to match. The sellers who benefit from AI are the ones whose books were already clean enough for a model to read.

What questions should I ask a vendor?

Four. Where does the model get its data, and can I see the source transaction behind any number it shows me? What happens when it is unsure, and does it flag or guess? Can I export the full ledger if I leave? And who reviews the model’s output before it posts to my accounting system? A vendor with good answers to all four is selling a tool. A vendor who deflects on the last one is selling a risk.