The marketplace now keeps the books. Someone still has to check them

For small online sellers the platform has become the ledger, and AI can sort it in minutes, but it cannot decide what the numbers mean or answer for them.

Ask a small online seller where their books are and the honest answer is increasingly “on Amazon”. Or eBay, Etsy or Shopify. The platform holds the orders, the refunds, the fees and the payouts. The spreadsheet the seller keeps, if they keep one, is a copy of somebody else’s ledger. 

That shift is one of the more useful things to happen to small-business bookkeeping in years, and AI makes it more useful still. It also creates a new kind of mistake: one that looks tidy. 

The ledger has moved to the platform 

Start with the money. A marketplace does not pay a seller sale by sale. It pays a net figure. eBay’s own help pages say that selling fees and other costs “are automatically deducted from your sales proceeds before they appear in your account”. 

The settlements we reconcile each quarter arrive the same way: one deposit, with sales, refunds, fees and often advertising already netted off. The bank statement shows one number. The VAT return needs every piece of it, pulled back apart. 

Then the VAT. For some sales the marketplace now accounts for the VAT itself. Since 1 January 2021, where an overseas business sells goods already in the UK through an online marketplace, HMRC’s guidance says the marketplace is “liable for the VAT”. 

The same applies to goods imported in consignments worth £135 or less. Both rules turn on where the seller is based or where the goods are when they are sold. A UK seller shipping from a UK warehouse still accounts for VAT on their own sales, and that is exactly the kind of distinction that gets lost when every line of data looks the same. 

And the platforms now report to HMRC on the income of the people who sell through them. The rules for digital platforms took effect on 1 January 2024, and the first reports were due in January 2025. Each year an operator must report its reportable sellers to HMRC by 31 January, and must give each seller the same information by the same date. 

The exemption for small sellers of goods is narrow: fewer than 30 sales and 2,000 euros (about £1,700) or less in the year, and both have to be true. Someone with 25 sales worth £3,000 is reported. So is someone with 40 sales worth £500. 

So HMRC now holds a version of the seller’s turnover that the seller never typed in. The books have moved, and the tax authority is reading from the same copy. 

What AI does well with it 

This is where AI earns its place. Platform data is high volume, structured and repetitive, which is exactly what models are good at. 

Categorising is the first job. A busy seller can have thousands of lines in a quarter, each needing a VAT treatment: a standard-rated sale, a zero-rated sale, a refund, a fee, postage. A model that has learned the seller’s own history does a first pass in minutes, and does it the same way every time. 

Matching is the second. Each payout has to be tied to the bank deposit it became, then pulled apart into the sales, refunds and charges that made it up. That is tedious, rule-bound work, and a machine does not get bored of it at eleven at night. 

The third, and the most underrated, is flagging. A good system says when it cannot confirm something: a refund with no matching sale, a fee that does not look like last quarter’s, a deposit it cannot place. A list of exceptions is worth more than a confident total, because it tells a person where to look. 

What it cannot do 

Three things remain stubbornly human. 

Deciding whether a line is in scope. Whether a sale belongs on this seller’s return or was accounted for by the marketplace. Whether a fee invoiced from abroad needs the reverse charge, under which, in HMRC’s words, you “must act as if you are both the supplier and the recipient of the services”. 

The data records the transaction. The treatment depends on facts the platform does not hold, such as where the seller is established and where the goods were at the moment of sale. 

Noticing what is not there. AI is good at checking what it can see. It is poor at paperwork that never arrived. HMRC’s VAT guide puts it plainly: to reclaim VAT you “must hold valid evidence that you have received a taxable supply”. 

The supplier invoice that was never sent does not show up in any feed as an error. It is simply absent, and the VAT claimed against it is unsupported until somebody goes and gets it. In our practice, chasing that missing document is still a job a person does every quarter. 

Putting a name to the return. The seller stays responsible for what is filed under their name. HMRC’s factsheet on inaccuracy penalties says that if you ask someone else, such as an adviser, to act for you, “you must do everything you reasonably can to prevent any inaccuracies.” The adviser, in turn, is held to HMRC’s standard for agents, which expects them to “avoid including figures in returns or claims which are unsubstantiated or speculative.” 

Both obligations sit with people. A model can propose a figure. It cannot answer for one. 

The new job is checking 

None of this is an argument against AI. Marketplace data is far better than a shoebox of receipts ever was, and the tools that read it are good and getting better. The risk is that tidy output feels like checked output. A clean, fully categorised quarter can still have the wrong sales on it. 

So the bookkeeping job has changed shape rather than disappeared. Less keying, more judgement. The person on the file now spends their time on the exceptions list, the scope questions and the documents that should exist but do not. 

For a seller, the practical point is short. Download your platform reports, keep your supplier invoices, and when you ask where a figure on your VAT return came from, make sure the answer is a person who checked it, not just a platform that produced it. 

The marketplace keeps the books now. Someone still has to check them.