Can AI do your bookkeeping?

Partly. An AI employee can run the administration around your books. It collects documents, chases the client who has not sent them, answers the routine question about where a receipt should go, and keeps a readable record of all of it. It cannot be your bookkeeper. The ledger, the reconciliation and the VAT return stay with a qualified person, and it is better to say that here than to let you discover it in month three.

So this post is a fence. Working from the AI employees role as it already stands, here is what one is genuinely good at inside a finance function, what it is not scoped to do, and why the second list is longer than a sales page would like.

Partly, and the split is the useful part of the answer. An AI employee is built for the work that surrounds bookkeeping. It collects the documents, chases the client who has not sent them, reminds people before a deadline, and keeps a readable record of what went out and when. The bookkeeping itself stays with a qualified person, because a ledger, a reconciliation and a filing each need somebody who is accountable for the number.

The split is deliberate, and it is the same line we hold everywhere. An AI employee works inside scripts you approved, and anything carrying a commitment or a judgement goes to a named person on your team. Proposals, contracts and invoices already wait for a person across our email automation work. A ledger entry is the same kind of object.

The chasing and the record keeping. Requesting receipts, statements and invoices, following up when they do not arrive, confirming what has landed and what is still outstanding, answering the routine question about where a document should go, and prompting a deadline while there is still time to act on it. All of that is administrative work with a paper trail, and none of it asks the employee to decide what a transaction was.

• Requesting receipts, statements and invoices from a client, then following up when they do not arrive.
• Confirming what has been received and what is still outstanding, so nobody reconstructs it from memory at the deadline.
• Answering the routine questions about where a document goes and which period it belongs to.
• Prompting a deadline while there is still time to do something about it.
• Keeping an audit trail of every request, reply and handover, readable in full.

Because a reconciliation is a judgement about what a transaction actually was, made by somebody who carries the consequence of getting it wrong. A payment that looks like a supplier invoice may be a refund, a personal draw, or a bill that was already recorded last month. An employee working inside an approved script can flag the lines that do not match the pattern and put them in front of your bookkeeper. Deciding what they are is the work you pay a bookkeeper for.

Try it on a single line of a bank statement. Four hundred pounds paid to a builder could be a repair, an improvement that belongs on the balance sheet, a deposit against work not yet done, or the second half of a bill already recorded. Each answer changes the accounts and one of them changes the tax. Choosing between them is professional judgement, and the person choosing carries the consequence.

No. Under Making Tax Digital, a VAT-registered business has to keep its VAT records in what HMRC calls functional compatible software, and the return reaches HMRC through that software's own connection. HMRC does not accept data moved by hand between programs as a digital link, so anything that works by reading a mailbox and retyping figures is the wrong shape for the job. Filing stays with your accountant and the software they file from.

HMRC is specific about the mechanism, which is what settles this. VAT records have to be kept and preserved in functional compatible software that can send returns to HMRC and receive information back through its interface, and HMRC rules out moving data by hand between software programs as a digital link (HMRC VAT Notice 700/22, Making Tax Digital for VAT, last updated 1 April 2022). An assistant that reads a mailbox and retypes a figure into a form sits outside that by design.

Often, because the costly part of finance admin is the waiting. A fortnight passes while a client means to send a statement. An invoice query sits for four days because everybody assumed somebody else had it. A deadline arrives before the paperwork does. An employee that owns the chasing hands your bookkeeper a complete file on the day they sit down to the work, and the hours it recovers were never billable in the first place.

There is a second reason worth naming. Finance admin is where a small firm's goodwill quietly leaks, because the chasing gets done by whoever has a spare ten minutes and lands as a stiff email from a different person each time. One worker doing it in a single approved voice, on a schedule you set, reads better to the client than a rota of reluctant volunteers.

This is the shape we take with regulated work generally. Administrative handling is automatable, and regulated judgement belongs to a person, which is the boundary we set out across four markets in the rules on AI and client intake. Bookkeeping sits in the same family. The paperwork around it can be automated. The professional act at the centre of it stays where it is.

Ask the supplier to write down which of the two columns above their product sits in, and get it into the scope document. Anyone selling an AI bookkeeper should be able to say who is accountable for a wrong figure and what happens on a filing deadline. Where we run a system on your behalf, that boundary is settled in writing before launch, along with the three other things set out in working with AI employees, and what we do and do not hold is published on security, including the certifications we have not got.

Somebody has to own the chasing, and in most small practices nobody does. That is the role we would scope, quoted in writing before any work begins, with a scope document saying in plain words that your books stay with your bookkeeper. The contact page reaches us.