Automate the data entry, the categorization, the reconciliation matching, and the anomaly flagging. Do not automate judgment calls, period-end estimates, or anything a state or the IRS would ask you to defend. That line is stable across almost every ecommerce business, and it has held up better than the more aggressive predictions made about this category two years ago.
What makes the boundary easier to defend now is the 2026 adoption data. Firms are buying these tools fast and integrating them barely at all, and the survey figures below show how wide that gap has gotten.
What the numbers actually show
The Federal Reserve Banks’ 2026 Report on Employer Firms, drawn from the 2025 Small Business Credit Survey of 6,525 firms with one to 499 employees, found that 46 percent of firms reported that their business or its employees currently use AI, with another 15 percent planning to start within twelve months.
Then the same survey asked how deeply. About half of those users described themselves as experimenting. Forty four percent said partially integrated. Just 7 percent said fully integrated. The most cited challenge among users was accuracy, at 46 percent.
Thomson Reuters Institute’s 2026 AI in Professional Services Report found organization-wide usage nearly doubling to 40 percent from 22 percent the year before, while only 18 percent of respondents said their organization tracks the return on investment of its AI tools. Adoption is running well ahead of measurement.
The readiness picture is worse at the small end. A Fall 2025 survey of 1,735 executives conducted by AICPA & CIMA with the NC State ERM Initiative found that fewer than one in five smaller organizations had the talent or systems they considered necessary, against roughly a quarter of organizations overall reporting adequate AI-skilled staff or IT readiness.
Read together: lots of firms are touching this, very few have rebuilt a process around it, and the ones least equipped to do it safely are the small ones.
What to automate, with confidence
Transaction capture and coding. Pulling marketplace and bank data in and coding recurring transactions to the right accounts is pattern matching on a stable ruleset. Software has been better than a person at this for years, and the accuracy question is largely settled for anything that recurs.
Settlement reconciliation. A marketplace payout is a net figure after fees, refunds, reserves, and adjustments. Breaking it into components and tying it to the deposit is deterministic arithmetic against a defined file format. A human doing this by hand is doing a job that has a correct answer, slowly.
Cost of goods sold on a defined costing method. Once the method is chosen and landed costs are loaded, applying it per unit is mechanical. Note the order of operations there. The method is a judgment call. The application is not.
Anomaly detection. Flagging that fee load moved two points as a share of revenue, or that a SKU’s margin inverted, is exactly the kind of continuous monitoring a person will not do reliably every month. This is where the newer tooling earns its keep, and it is the one area where the automated version is not merely faster but genuinely better than the human alternative, because the human alternative is that nobody checks.
What not to automate
Anything involving a nexus determination. Whether inventory sitting in a state creates an obligation is a legal question with unsettled answers in places. Software can tell you where your inventory has been. It should not tell you what that means.
Inventory costing method selection and reserves. Which costing method to use, and what obsolescence reserve is defensible, are positions you take and may have to support. The IRS Publication 538 on accounting periods and methods covers the small business taxpayer exceptions and the inventory rules under Regulations section 1.471-1(b), and the reason to read it is to understand that these are elections with consequences, not settings.
Period-end accruals and cutoff judgments. Whether a cost belongs in this period or the next is frequently a matter of professional judgment on incomplete information.
The final review. Something should look at the finished statements and ask whether they make sense. Automated systems are confidently wrong in ways that pattern-match to correct, which is the failure mode the 46 percent accuracy concern in the Fed survey is describing.
The tooling, as it actually stands
Nobody in this category is selling a bookkeeper replacement, whatever the marketing suggests. What they sell is elimination of the mechanical half.
Sellerboard does profit analytics for Amazon FBA sellers with published pricing starting at nineteen dollars a month, and for a single-channel seller who wants per-product profit numbers without touching the general ledger, it is faster to set up and cheaper than anything accounting-integrated. That is a real edge and it matters for smaller operations.
A2X is the conventional settlement-journal choice, with open pricing from twenty nine dollars a month at low Amazon order volume, though cost of goods sold starts on the tier above. Link My Books covers similar ground into Xero and QuickBooks with cost of goods sold tracking and channel-level profit and loss.
ConnectBooks works the multi-marketplace accounting side, syncing Amazon, Shopify, Walmart, TikTok Shop, and eBay into QuickBooks Online, QuickBooks Desktop Enterprise, or Xero with automated cost of goods sold, real-time inventory tracking, SKU-level profit and loss, and settlement reconciliation. Its AI layer, Crunch, is in active beta.
The honest caveat across all of them: setup quality determines output quality, and the accounting-integrated tools ask more of you up front than the analytics-only tools do. A seller who loads incomplete landed costs gets confident, precise, wrong margins.
A division of labor that works
The arrangement most multi-marketplace sellers land on is software handling capture, coding, reconciliation, costing application, and monitoring, running continuously. A bookkeeper handling exceptions, the things the system flagged, and the monthly close. An accountant or CPA handling method elections, nexus questions, tax positions, and the annual review.
What changes when the automation works is not headcount, usually. It is that the human hours move from data entry to the questions that actually have money attached: which products to kill, whether the supplier increase can be absorbed, what the last quarter’s fee drift means.
The 7 percent full-integration figure in the Fed survey suggests most firms have not gotten there. The ones that have did not start by automating judgment. They started by automating the parts with a correct answer, and then hired the judgment back in at a higher level.
