The Software Posted It, Not You: A Human Review Checklist for AI Bookkeeping

Automation has quietly become the default in small business bookkeeping. Bank feeds populate themselves, rules and machine learning suggest categories, and some tools will draft a VAT return with little more than a click. The appeal is obvious: fewer evenings spent in a spreadsheet and more confidence that nothing has been missed.

The difficulty is that speed of entry is not the same as accuracy of entry. HMRC does not treat a software-generated figure as inherently correct, and it does not accept that responsibility shifts to the tool when a category is wrong. The obligation to take reasonable care, and to keep records that support the figures on a return, remains with the business and its directors or proprietors. What changes with automation is the nature of the review you need to perform, not the need for one.

Accountability stays with the business

HMRC's guidance on reasonable care is framed around what a prudent person would do in the circumstances. Using software is not, in itself, a failure of care; using it without checking is. If an AI tool miscategorises a payment and the error flows through to a VAT return or a set of accounts, the starting point for any enquiry will be the quality of the process behind that entry, not the sophistication of the software.

It is worth being precise about roles. The software produces suggestions and, in some cases, automatic postings. The agent or adviser reviews and advises. The business owner approves and submits. Only the last of those carries the legal responsibility for the return's accuracy, and that responsibility cannot be delegated by subscription.

What a defensible automated process looks like

Three things make an automated bookkeeping process stand up to scrutiny: a documented process, evidence that a human reviewed the output, and the ability to reconstruct how a figure was reached. If any of those is missing, the process is relying on luck rather than on control.

A documented process can be short. A single page setting out which tools are used, who reviews what and how often, where approvals are recorded and how corrections are logged is usually sufficient. Evidence of review is the part most often skipped, and it is the part an inspector will ask for first.

A monthly review checklist

The following steps are a sensible minimum for a business using automated categorisation. They need not take long, but they should happen on a predictable cycle and leave a trace.

1. Reconcile to the bank statement, not to the software's own balance. Confirm that the closing balance in the tool matches the actual statement. A feed that has silently dropped a transaction will often reconcile to itself while diverging from the bank.

2. Sample the high-risk transactions. New suppliers, round-number payments, overseas transfers, cash withdrawals, director-related payments and anything posted to a suspense or default account deserve a second look. Sampling a fixed number each month, for example twenty entries, creates a defensible routine.

3. Check VAT-sensitive categories. These are the areas where automated suggestion is weakest and the cost of error is highest. They include entertainment, staff welfare, subsistence, motoring costs, property repairs versus improvements, imports and reverse-charge services, and zero-rated or exempt sales. A miscoded VAT treatment can distort a return even when the underlying transaction is genuine.

4. Look for duplicates and split entries. Feeds occasionally double-post, and rules occasionally split a single payment across two categories. A quick scan of repeating amounts or unusual pairings will usually catch these.

5. Inspect the audit trail. Most modern tools record who or what changed an entry and when. Confirm that manual overrides are identifiable and that the original suggestion, where relevant, is retained rather than overwritten.

6. Record a sign-off. A dated note confirming that the review took place, who performed it and what exceptions were found is worth more than it appears. It converts an informal habit into evidence.

Digital links matter more, not less

Making Tax Digital for VAT requires digital links between the source records and the return, and the same discipline is sensible well beyond VAT. A bank feed connected by API is a digital link. A spreadsheet that pulls figures by formula is a digital link. A figure retyped from one system into another is not, and neither is a copy-and-paste that breaks the chain.

Automation tends to improve digital links rather than harm them, provided the chain is understood. Where a bridging tool is used, or where a CSV is imported into accounting software, note the route in the process document so that it can be explained if asked.

When the tool gets it wrong

Errors will happen, and the response matters more than the error. Correct the entry promptly, record what was changed and why, and retain the original if the software allows. A short error log, kept alongside the bookkeeping, is a pragmatic control.

If an incorrect figure has already reached a submitted return, do not quietly adjust the next one without thought. The appropriate correction route depends on the tax and the size of the error: VAT adjustments, amendments to income tax returns, and disclosure to HMRC all have their own rules and time limits. Take advice where the amount is material or the pattern is repeated.

Records, backups and continuity

Keep the underlying documents, not just the software's summary: invoices, receipts, bank statements, contracts and credit notes. Export a full backup of the accounting data at least quarterly and store it somewhere independent of the tool. Confirm who holds administrator access, enable two-factor authentication and remove leavers promptly.

Retention periods vary by tax and by entity, but six years from the end of the relevant period is a safe baseline for most small businesses, and longer where property or capital transactions are involved. The point of retention is that records remain readable and complete, not merely present.

A workable habit

AI bookkeeping is not a threat to compliance; unreviewed AI bookkeeping is. A monthly routine of reconciliation, targeted sampling, VAT checks, audit trail review and a dated sign-off takes perhaps an hour for a typical small business and converts an automated process into a defensible one. The software can do the posting. The judgement, and the responsibility, remain human.