AI is good at the recording layer of ecommerce bookkeeping and unreliable at the interpretation layer. The line between them has nothing to do with how advanced the tool is. It runs along a simpler fault: where the right answer sits inside structured data a model can read, automation gets there faster than you do. Where the right answer depends on a judgment about one client's specific facts, no model gets there at all.
For a firm serving Amazon, Shopify, Walmart, TikTok Shop and eBay sellers, that fault line matters more than it does for a typical service client, because ecommerce data arrives pre-broken in ways general automation was never built to handle.
The four layers of an ecommerce close
Sort the work into four layers and the picture sharpens.
Capture. Pulling transactions in from banks, cards, marketplaces and payment processors.
Classify. Deciding which account each transaction belongs to.
Reconcile. Proving the ledger agrees with an outside source.
Interpret. Deciding what the finished numbers mean and what the client should do about them.
AI has taken real ground in the first two, made partial progress on the third, and almost none on the fourth. The fourth is what clients pay an advisor for.
Where it helps
Transaction classification is the clearest win. A model that has seen six months of a client's coding history will code a recurring bank line correctly at a volume no junior staffer matches, and it does that at two in the morning without a review queue backing up behind it.
Feed matching is second. Tying a deposit to an invoice, or a payout to a settlement, is pattern work, and pattern work is what these systems do.
Drafting is third and the most underrated of the four. Client emails, close notes, workpaper narratives, engagement summaries, a first pass at a management letter. Firms that got value out of AI early did it here, because the cost of a bad draft is a rewrite rather than a restatement.
Exception surfacing is fourth. A model that flags the entry unlike the last two hundred entries of its type does something a human reviewer does poorly at scale and does not enjoy.
Where it breaks
Four places, and every one of them is specific to ecommerce.
Settlement periods do not align with calendar months. An Amazon settlement that opens on the 27th and closes on the 9th carries sales from two accounting periods. Splitting it correctly requires a policy decision about orders shipped in one period and refunded in the next. There is no pattern to learn, because the client's policy did not exist until someone wrote it down.
Reserves are not expenses. Marketplaces withhold funds at period end, and that withheld amount is a receivable. Automation that reconciles to the bank deposit leaves the reserve out of the picture entirely. Revenue looks light one month and heavy the next, and the client asks you why.
Inventory valuation is a cost problem, not a coding problem. One SKU can carry four unit costs across four purchase lots, and the freight and duty allocated to each lot moves the number again. A model codes the freight bill to the right account without trouble. It cannot decide which units that freight belongs to.
Returns and unsellable returns behave in opposite directions. A returned unit that goes back into sellable stock and a returned unit that gets disposed have opposite effects on inventory value, and a reimbursement for a lost unit has to be matched to the write-off it offsets. Miss this and COGS runs understated for months with nothing flagging it.
One structural problem sits behind all four. If a marketplace payout lands in the ledger as a single deposit with no detail behind it, no model can reconstruct what it contained. The fix is a connection that decomposes the settlement before any automation runs on top of it. ConnectBooks does that across Amazon, Shopify, Walmart, TikTok Shop and eBay into QuickBooks Online, QuickBooks Desktop Enterprise or Xero, and that layer has to be right before an AI layer has anything honest to read.
What stays with the accountant
Four things, and they are the four worth billing for.
The chart of accounts. Splitting marketplace fees into referral, fulfillment, storage, advertising and refund administration is a design decision. A single "Marketplace fees" account is where margin analysis goes to die, and no tool will unwind it for you after the fact.
The cutoff policy. Write down how this client handles period straddles, reserves, inventory in transit, and advertising charged against a settlement that spans two months. Consistency matters more than which choice you make, because consistency is what makes periods comparable.
The exception queue. Automation produces a list of items it declined to handle. Someone owns that list on a schedule and has the authority to change a coding rule when a new pattern appears. Weekly works for most sellers. Monthly is too slow, because a miscoding repeats forty times before anyone catches it.
The conversation. Whether a strong month was growth or a channel promotion. Whether the client can fund the next purchase order. Whether the books would survive a buyer's diligence. Whether the inventory loan makes sense at that rate. No product answers these, and this is most of what a good advisor is for.
The framing that holds up
Machines do the recording. People own the policy. That description survives contact with the work better than "AI will replace bookkeepers" or "AI cannot be trusted," both of which are marketing positions rather than accounts of what happens during a close.
It also tells you where the analysis tools fit. Once settlements are decomposed, fees are split, COGS is per unit and the cutoff policy is written down, an analysis layer finally has something real to read. Crunch, the AI CFO built into ConnectBooks, answers questions against that reconciled data: why profit moved between two periods, which SKUs sit behind the change, what advertising and fee bands did to margin. It reads the books. It does not create them, and nothing else in this category does either.
For the sequence in a form you can hand to staff, the monthly ecommerce bookkeeping checklist runs the close task by task, from pulling marketplace reports through inventory reconciliation. Firms taking on more ecommerce work can list with the ConnectBooks EcomExpert directory, and the Accounting Portal handles firm-level access across multiple client accounts without chasing individual invitations.
The driver's seat is not a sentimental idea. It describes where the judgment calls live, and they have not moved.