AI bookkeeping does the coding. It cannot do the judgement.

This is a guide, not tax advice; check anything that matters to your business with your accountant.

What AI bookkeeping is genuinely good at

The vendors are not lying about the first half. Reading a receipt, matching it to a bank line and coding it the way you coded the last forty like it are pattern-matching problems, and pattern-matching is what this technology does well. A modern bank feed with receipt capture will handle the large majority of a small business ledger without anyone touching it.

That covers the part of bookkeeping that is really data entry: the recurring supplier, the monthly subscription, the fuel card, the card machine settlement. None of it needs a person, and if someone in your business is still typing those in, that is the money these tools actually save.

It is worth being precise about why it works, because the reason is also the limit. The software is not reading tax law. It is reading your history. Ask it to code something you have coded a hundred times and it will be right nearly every time. Ask it to code something it has not seen, and it will guess from the nearest thing it has.

Can AI do my bookkeeping?

It can do most of the recording and almost none of the deciding. AI bookkeeping codes transactions by matching the document and the bank line against your own history, which covers the bulk of a typical ledger. It cannot tell you whether a cost was business entertainment or staff welfare, and in the UK that distinction changes what you can reclaim.

That sounds like a small carve-out. It is not, because the cases it cannot decide are the same cases that get businesses into trouble.

The receipt is the same. The tax is not.

Take two restaurant bills for £180. One is lunch with a prospective client. One is a team lunch for your own staff. To optical character recognition, the software that reads text off an image, they are the same document: same merchant, same amount, same VAT line.

To HMRC they are not. VAT incurred on business entertainment, which HMRC defines as hospitality provided free to people who are not employees, cannot be recovered. VAT on entertaining your own employees is input tax and is not blocked under those rules. Same receipt, different answer, and nothing on the paper tells you which.

That is the shape of every hard case in bookkeeping. The fact that decides the treatment is not on the document. It is in somebody's head: who was there, what it was for, whether the journey was really for business. Software that reads documents cannot reach it, and a system confident enough to post without asking will post the wrong one.

Partial cases are worse. Where entertainment covers both employees and non-employees, HMRC expects the cost to be apportioned and only the employee share recovered. An apportionment is a judgement with a number attached. It is not extraction, and no amount of reading the receipt produces it.

The failure people miss is the confident one

A miscoding does not stay a single error. These systems learn from what you accept, so the first time a private cost goes through as a business expense and nobody changes it, that becomes the pattern. Next month it is coded the same way with more confidence. By the year end there are eleven of them and no flag anywhere, because the software is not uncertain. It is consistently wrong.

This is why every agent we build runs in shadow mode first, drafting without posting, so somebody can see what it would have done before it does it. The point of that stretch is not to check the easy transactions. It is to find the ones where the agent is confident and wrong, because those are the only ones that survive a review.

Does Making Tax Digital change this?

Yes, mainly by shortening the gap between a mistake and HMRC seeing it. From 6 April 2026, sole traders and landlords with qualifying income over £50,000 must keep digital records and send quarterly updates rather than one return a year. The threshold falls to £30,000 from April 2027 and £20,000 from April 2028.

Under the old rhythm, a coding error made in May was found the following January by whoever prepared the return. It was wrong for eight months and nobody outside the business ever saw the wrong version. Under quarterly filing it is submitted within weeks.

That is an argument for automation rather than against it. Four submissions a year is four times the work if a person is doing the coding, which is not a serious plan for a business that was already behind. It is roughly the same work if the software codes and a person reviews. What Making Tax Digital changes is not whether to automate. It is that the review now has to happen every quarter instead of every January.

What the review actually looks like

Not line by line. If you are checking every transaction, you have bought nothing.

Review by exception, and the exceptions are predictable. Anything the software flagged as low confidence. Anything in entertainment, subsistence, travel or motor, which is where purpose decides treatment rather than the document. Anything new: a supplier you have not used before, or a category with one transaction in it. Anything large enough that being wrong matters. On a typical small business ledger that is a small fraction of the lines, and it is where nearly all the errors live.

In the accountancy practice we built for, the reconciliation agent handles receipts and flags what it cannot settle, and the partners look at the flags rather than the ledger. That is the division of labour that works: the agent does the volume, the person does the judgement, and the agent is explicit about which is which. Our accountancy practice page sets out where else the same split applies, and we have written separately on what an agent should never do without asking.

What to do next

Take last month's ledger and mark every line the software coded on its own. Then go through entertainment, subsistence and travel properly, and count how many you had to change.

That ratio is your answer. If it is near zero, your business is simple enough that the software can run with a light quarterly check. If it is not, you have just found which categories need a person, and you can stop reviewing the rest. Either way you have measured it rather than guessed, which is the whole argument of working out the job before buying the tool.