AI helps accountants most in the work clients never see, and breaks in the work clients care about most. Drafting, coding, summarizing, and first-pass matching have gotten meaningfully faster. Inventory valuation, period cutoffs on marketplace data, and anything requiring a judgment call about a client's specific facts have not improved at all, because they were never pattern-recognition problems.
For a practice serving ecommerce sellers, that split has a sharp edge. The automatable part is roughly the part you were already discounting. The non-automatable part is the part that goes wrong on every new client you take over.
The most widespread real use of AI in accounting firms right now is writing. The Karbon State of AI in Accounting Report 2025, based on responses from more than 500 accounting professionals across six continents, found that firms using AI reported saving an average of 18 hours per employee per month by automating routine communication work, with 63 percent citing email drafting and 40 percent citing meeting summaries.
That is a real result and a modest one. It is also the least glamorous claim in the category, which is probably why it is the most reliable.
For a client with a few thousand monthly transactions across bank, card, and marketplace feeds, a classifier trained on that client's history handles the recurring majority. The staff time freed is the time spent coding items that were never in doubt.
Payments to invoices, deposits to payouts, credits to returns. Some of this is rules-based rather than model-based, and the distinction only matters when you ask what lands in the exception queue.
Explaining a fee type, summarizing a state's guidance, producing a first draft of a memo. Fast, and easy to verify against the primary source, which you should do anyway.
Structure, headings, tie-out schedules, a first pass at variance commentary. The model produces the shape and you supply the judgment.
This is the failure that matters most in ecommerce and the one automation is worst at. Costing requires knowing what arrived, at what landed cost, into which location, and what shipped. None of that lives in a bank feed, and a classifier cannot infer it.
A settlement period rarely aligns with a calendar month. Orders ship in one period and refund in the next. Reserves are held and released across boundaries. Advertising is charged against a settlement rather than invoiced. Every one of those creates a cutoff decision, and every one of them is a decision rather than a pattern.
Intercompany eliminations, transfer pricing between a US entity and a foreign supplier, currency remeasurement on inventory purchased in a different currency. Automation executes a policy here. It does not choose one.
Marketplace facilitator rules, physical presence created by a third-party warehouse, product taxability by state. These are legal positions with facts underneath them. A model that produces a confident answer about a client's nexus is producing a liability, not an analysis. Point clients to a qualified professional or the relevant state's department of revenue.
Whether a client should take the inventory loan. Whether the books are ready for a buyer. Whether the owner understands that a strong month was a channel promotion rather than growth. No tool does this, and it is most of what a good advisor is for.
A practice takes over an importer doing $4.12 million in revenue across three marketplaces. The prior bookkeeper used AI-assisted coding. The bank is reconciled. Categorization is consistent. Nothing in the trial balance looks alarming.
The freight and duty on twelve inbound containers, $412,800 for the year, was coded to an operating expense account called "Freight and shipping" rather than capitalized into inventory. It was coded that way consistently, month after month, with high confidence, because that is how the previous bookkeeper had coded it and that is what the model learned.
The year purchased 186,000 units. Freight and duty per unit works out to $2.2194. Ending inventory is 41,204 units.
A 7.8 point overstatement of gross margin, an understated balance sheet, and net income overstated by $91,448.60. Every SKU-level profitability number the client had been using to make pricing and reorder decisions was wrong in the same direction, which is the worst kind of wrong because it is invisible in comparisons.
No confidence score flagged any of it. The automation was accurate to its training and its training was the mistake.
Two things follow for a practice.
First, the compliance floor is dropping. Work that used to be billable because it took hours now takes minutes, and clients will eventually notice. Practices that priced on time spent recording transactions have a repricing problem coming.
Second, the diagnostic work is getting more valuable, not less. Someone still has to catch the freight coded to the wrong side of the gross margin line, and that catch is worth more to the client than a year of clean categorization.
The Karbon report noted the same tension from the other side: only 37 percent of firms surveyed were investing in AI training for their teams, against 85 percent expressing optimism about AI in accounting. Enthusiasm is not a capability.
A costing policy per client, written down. FIFO or weighted average, when it is applied, what is included in landed cost. If it is not documented, the automation will pick up whatever the last person did.
Settlement-level data feeds. Every marketplace payout should arrive in the ledger decomposed into transaction and fee types. This is a connection problem, and it precedes everything else. ConnectBooks does it across Amazon, Shopify, Walmart, eBay, and TikTok Shop into QuickBooks Online, Desktop, and Enterprise or Xero.
A chart of accounts that separates fee types. If commission, fulfillment, storage, and advertising collapse into one expense line, no downstream analysis can separate them either.
A named exception owner per client. The automation produces a queue. Someone reviews it on a schedule. Without this, the queue becomes a backlog and the backlog becomes a restatement.
A quarterly costing review. Specifically: does inventory on the balance sheet tie to the physical count at landed cost, and is freight where it belongs. The example above would have surfaced in one afternoon.
ConnectBooks has announced Crunch, an AI CFO built to answer questions using a seller's reconciled settlement, inventory, and channel data, and there is a waitlist open ahead of its release. It is not available to use yet, and the page describes what it will do when it opens up.
The message to give clients in the meantime is unglamorous and correct. The analysis tools that are coming will be as good as the books underneath them. A seller whose settlements post as lump-sum deposits and whose freight sits below the gross margin line will get fluent, confident, wrong answers from every product in this category.
Practices that want inbound ecommerce work can list with the ConnectBooks ecommerce accountant directory. The sellers who find you there will already have the SKU and channel profit reporting that makes the advisory conversation possible, which is a better starting point than a shoebox.
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
