AI bookkeeping for one person, and where it goes wrong

The pitch for automated bookkeeping is that the machine reads your bank feed, assigns every line to a category, and hands you a finished set of books. For a one person business this mostly works, and the way it fails is worth understanding, because the failure does not look like a failure. It looks like a tidy ledger that balances and is wrong.

Nothing here argues against using these tools. Matching a receipt photograph to a card transaction is genuinely tedious work that software does better than you do at eleven at night. The argument is narrower: know which four kinds of line you have to look at yourself, and know why the software cannot get them right no matter how good it gets.

Why does categorisation go wrong on the lines that matter?

Because the tool is predicting, not deciding. It sees an amount, a date and a vendor string, and it assigns the label most commonly attached to that pattern across everybody's books.

One accounting practice puts the mechanism plainly: AI is not making decisions, it is making predictions, and a vendor payment could be a loan repayment, an equipment purchase or a reimbursable expense while the model picks the most common classification rather than the correct one. That is fine when your business is typical. Your deductions, by definition, live where it is not.

The same source lists the specific errors that recur: expenses in the wrong category, transfers mislabelled as income, personal and business spending not separated, and adjustments overlooked. Read that list again with an eye on direction. Some of those errors raise your reported profit and some lower it, and a set of books can contain both at once while reconciling perfectly.

What does a real misclassification look like?

Mundane and expensive. Two worked examples from an accounting practice's audit checklist make the shape clear, and neither involves anything exotic.

In the first, a 4,500 dollar invoice for a one off patent filing was categorised as a software subscription because the vendor name looked like a software product. It should have been a professional fee, and the practice notes that piling professional fees into subscriptions can trigger a deduction ratio anomaly in an audit. The books balance. The pattern of your spending now tells a story about your business that is not true.

In the second, a recurring 1,200 dollar monthly payment was filed under dues and subscriptions when it was payment to a contractor. Same amount, same day of the month, every month: it looks exactly like a subscription because in every measurable respect it is shaped like one. The consequence is not a wrong total. It is a missed 1099-NEC filing by the January deadline and a penalty per missing form.

That second example is the one to hold onto. The error is invisible in every report you would look at. Revenue is right, expenses are right, profit is right. What is wrong is a categorisation that carries a legal obligation nobody triggered, and you find out about it from a letter.

Which transactions should you always check yourself?

Four kinds, and the list is short enough to do monthly in fifteen minutes.

Anything paid to a person. Contractors, freelancers, a friend who did your photography. These carry filing obligations that no amount of pattern matching can infer from a bank string, and they are the ones most likely to be shaped like a subscription.

Transfers between your own accounts. Moving money from your business account to your savings is not income and not an expense. Tools misread this often, and when they read a transfer as income they inflate your revenue and therefore your tax on money you already had.

Anything mixed. A phone bill, a car, a room in your house, a laptop used for both. The business share is a judgement about how you actually live, and the checklist above lists split verification as a specific review item for exactly this reason.

Anything unusually large. Big purchases can carry treatment choices that change your tax materially, and the automated path takes the default rather than the advantageous one. The same practice suggests a threshold review on transactions above 2,500 dollars rather than letting them categorise themselves.

What happens when the marketplace pays you net?

Your revenue disappears and reappears smaller, and nothing in the bank feed tells you the difference was fees. This is the single most common way a small seller's books end up understating both income and expenses by the same amount.

Picture a payout of 842 pounds landing in your account. The underlying sales were 1,000 pounds, the platform took commission, payment processing and an advertising charge, and it settled the net figure. If the feed records 842 pounds of income, your revenue is understated by 158 pounds and you have claimed no deduction for fees you genuinely paid. Profit comes out roughly right, which is exactly why nobody notices.

It matters for three reasons beyond tidiness. Any threshold measured on turnover rather than profit is now computed from a number that is too low, and registration thresholds work that way. Your margin analysis is wrong, because the fees are invisible rather than assigned to the channel that charged them. And if a platform later reports gross figures to a tax authority while your return shows net, the mismatch is yours to explain.

The fix is not clever. Take the settlement report the platform produces, not just the bank line, and record gross sales and each fee category separately. Most tools can import that report; the failure is usually that nobody switched it on, because the bank feed alone produces something that looks complete.

Do refunds and chargebacks need separate treatment?

Yes, and they are the second place net settlement quietly distorts the picture. A refund is not a business expense, it is a reduction in revenue, and a chargeback usually arrives as two separate events with a fee attached to one of them.

Categorised as an expense, a refund leaves your reported sales too high and your costs too high. The totals still work out, and once again the profit figure defends itself while the underlying numbers are wrong. If you are watching a dispute ratio for the reasons set out in our piece on what a blocked or disputed order actually costs, you need those events recorded distinctly anyway, so the bookkeeping fix and the payments fix are the same work done once.

Foreign currency adds one more layer worth naming. A sale in one currency, settled in another, on a date that may not match the sale date, produces a difference that is neither income nor an expense in the ordinary sense. Automated categorisation has no view on this at all. It is not a hard problem, but it is one nobody solves by accident, and it compounds quietly across a year of small international orders.

How long do the records have to survive?

Longer than most software subscriptions last. The retention rules are set by the tax authority rather than by your tool, and they run considerably further than the three years most people assume.

The IRS publishes a period of limitations ladder: three years in the ordinary case, three years from filing or two years from payment where you claim a credit or refund, seven years for a claim from worthless securities or a bad debt deduction, six years where unreported income exceeds 25% of gross income shown on the return, and indefinitely where no return was filed or a fraudulent one was. Employment tax records go four years from when the tax was due or paid. Records about property run until the limitations period expires for the year you dispose of it.

That last one is the sleeper. Buy equipment in 2026, sell it in 2032, and the purchase record is still live into the mid 2030s. Any tool holding your only copy of that receipt is now a decade long dependency, and the question of whether it exports cleanly stops being a preference and becomes the main thing about it.

What should you check before you commit to a tool?

The export, first and before anything else. Can you get every transaction, every category, and every attached receipt image out in a format you can read without the vendor, and can you do it today rather than on request?

A CSV of transactions is not sufficient on its own, because the receipts are the supporting documents and they are frequently stored separately as images the export does not include. Test it properly: run the export, open the archive, and check that a specific receipt from three months ago is actually in there and legible. Doing this in your first month costs an hour. Doing it during a wind down costs you the records.

The second thing to check is whether the categorisation history is exportable in a form that carries the reasoning. Knowing that a payment was classified as a professional fee is useful. Knowing when it was changed, and by whom, is what makes a set of books defensible later. The general principle of keeping a trail on anything automated is the same one we set out for deciding what a system may do without a person approving it.

Does quarterly reporting change the calculation?

Yes, and for a lot of sole traders it already has. Digital records stopped being a preference and became a requirement, on a schedule.

Under Making Tax Digital for Income Tax, UK sole traders and landlords registered for Self Assessment must use the regime once their qualifying income passes a threshold: 50,000 pounds based on the 2024 to 2025 tax year, from 6 April 2026, then 30,000 pounds from 6 April 2027 and 20,000 pounds from 6 April 2028. Those in scope have to use compatible software and keep going with Self Assessment until the switch applies to them.

The practical effect on a small seller is not the software requirement, which most already meet. It is the cadence. A categorisation mistake used to sit undisturbed until a year end tidy up, where you would notice it while looking at twelve months at once. Under quarterly reporting the same mistake goes out in an update three months after it happened, and the tidy up habit that used to catch it no longer sits between the error and the filing.

Which argues for the monthly fifteen minutes rather than against the software. The tool does the volume, you do the four categories above, and the quarterly figure is right when it leaves.

What the tool genuinely does better than you

Volume, consistency and matching. It is worth being clear about this, because the failure modes above are easy to read as an argument for doing it by hand, and doing it by hand is worse.

Reading a receipt photograph and pulling the amount, date and vendor out of it is a solved problem and a tedious one. Matching several hundred card transactions against several hundred receipts is work where human attention degrades and software does not. Flagging a duplicate charge, or a subscription that renewed after you thought you cancelled it, is something the machine notices and you do not.

The pattern across those examples is that the tool is reliable when the answer is contained in the document, and unreliable when the answer is contained in your head. That is the same line that runs through most of these tools, and it is the useful test to apply before automating anything: is the fact written down somewhere, or does it depend on knowing why. Extraction from documents specifically is a mature capability, and the mechanics of that pipeline are covered in our piece on turning documents into structured data.

Should a very small business bother with any of this?

The bar is lower than the marketing suggests. If you take a few hundred transactions a year through one business account, a spreadsheet and a folder of receipts satisfies every requirement in this article, and the IRS is explicit that you may use any system that clearly shows income and expenses.

What pushes you towards software is not size, it is mixture: several payment methods, a marketplace paying out net of fees, foreign currency, contractors, stock. Each of those adds a reconciliation problem where automation earns its money. If you have none of them, buying a tool to categorise sixty transactions a year is a subscription rather than a solution.

The overlooked cost either way is your own time spent moving data between systems that do not talk. Payment processor in one place, sales in another, expenses in a third, and a monthly ritual of exporting and pasting. That friction is a structural problem rather than a software one, and it is much easier to solve when the shop and its records are systems you control rather than services you rent access to. The same argument turns up whenever a format changes, which is the practical case for running a storefront you can wire into your own tools. Bookkeeping is mostly the tax on data living in places that do not speak to each other.