Automation

Where automated bookkeeping still needs a person: three tax decisions the data cannot make

By Harvey Dhillon, Founder and CEO, Zmartly

Automated bookkeeping is very good at the middle of the job. It reads a bank feed, matches a payment to a receipt, and puts 4,000 lines into sensible categories before anyone has finished their coffee. The trouble sits at the edges, where the right answer is not in the data at all. Those edges are where tax mistakes are made. 

This is not an argument against automation. It is an argument for knowing exactly where to put the human. Here are three decisions in UK VAT that no categorisation model can make from the ledger alone. 

Decision one: whether you are about to cross a threshold 

A UK business must register for VAT if its taxable turnover for the last 12 months goes over £90,000. It must also register if it expects turnover to go over £90,000 in the next 30 days, according to the government’s guidance on registering for VAT. The first test looks backwards, so software handles it well. The second test is a forecast about the future, and it depends on orders, contracts and plans that are not yet in any ledger. 

Picture a seller who has been turning over £7,000 a month and has just been offered a single wholesale order worth £95,000, to be delivered this month. Every figure in the books says they are comfortably under the line. The legal position changed the day they accepted the order, because they now expect to go over £90,000 in the next 30 days. A system that only reads the past will stay calm and be wrong. 

The fix is not clever. Someone has to ask the seller, regularly, what is coming. A dashboard can prompt that question, but it cannot answer it. 

Decision two: who actually owes the VAT on a sale 

On a marketplace, the same sale can create very different tax duties depending on where the goods are and who the seller is. Under the government’s rules for VAT and overseas goods sold using online marketplaces, the marketplace is liable for the VAT on consignments of £135 or less sent from outside the UK to UK customers, unless the customer is a VAT-registered business that gives its VAT number. It is also liable for the VAT on goods of any value that are already in the UK at the point of sale and sold by an overseas business. 

For a seller using those platforms, this can change what the payout report means, depending on where the goods are and where the seller is established. A line that looks like a sale with no VAT may be one the platform has already accounted for. Treating it as an untaxed sale produces a return that is wrong in a way no arithmetic check will catch. 

A model can learn to recognise the pattern. What it cannot do is know that a seller changed fulfilment route in March, or that a product moved from a warehouse overseas to one in Britain. Those facts live in the seller’s operations, not in the transaction data. 

Decision three: whose name goes on the number 

There is a quieter shift under both of these. Since 1 January 2024, platforms have been required to collect details about their sellers and report them to HMRC, as the government’s guidance for sellers on digital platforms explains. Details are not reported for sellers who make fewer than 30 sales of goods in a calendar year and receive less than 2,000 euros (about £1,700) for them. 

So HMRC increasingly holds an independent record of what the marketplace says a seller earned. That raises the stakes for the seller’s own books, because the two records will be compared. A return built on a tidy but mistaken categorisation is now easier to spot than it was. 

This is why the last step should stay with a person. Someone qualified has to look at the number, understand why it is that number, and be willing to put their name beside it. An automated system can explain its output. It cannot be accountable for it. 

What this means in practice 

If you are building or buying automation for bookkeeping, the useful design question is not “what percentage can it do alone”. It is “where does it stop, and does it say so”. Three habits help. 

– Make the system flag what it is unsure about, rather than guessing. A flagged line costs a minute. A confident wrong one costs a penalty and a correction. 

– Ask forward-looking questions on a schedule, because the threshold risks are in the plans, not the ledger. 

– Keep a named human responsible for the filing, and give them time to read, not just to click approve. 

Used this way, automation does not replace the accountant. It gives them back the hours spent typing, so they can spend them on the few decisions in each quarter that really are decisions. 

Related Articles

Back to top button