AI bookkeeping

Most of the work falls into four tasks. Bank feeds bring statement lines into the ledger automatically. Categorisation suggests an account, tax rate and contact for each line. Document capture reads receipts and bills. Reconciliation proposes matches between statement lines and existing invoices or bills.

Vendors describe how their own automation decides. Xero, for example, says its JAX bank reconciliation works from bank rules, matches to existing documents, the user's own reconciliation history, and predictions based on how other Xero users reconciled similar transactions, and only reconciles automatically when it is highly confident; otherwise it suggests a match and leaves the line for a person.

That pattern, automatic when confident and suggested otherwise, is the useful mental model for AI bookkeeping in general. The question is never whether the software can code a line, but what confidence threshold applies and who sees the lines below it.

Generative AI adds a different layer: drafting client queries about unexplained transactions, summarising a month's exceptions, or explaining a variance in plain English. Those outputs need checking in the same way as any other draft.

Start with explicit bank rules for recurring, low-risk transactions. A rule is deterministic and reviewable, which makes it easier to defend than a learned prediction.

Decide which accounts are never auto-posted: VAT-sensitive codes, directors' loan accounts, capital items, payroll and anything that crosses into another tax. Those should always land in a review queue.

Keep the audit trail. Most platforms record whether a line was reconciled by a rule, by automation or by a named user. Make sure that history is retained and that staff do not bulk-accept suggestions to clear a queue.

Data protection still applies. Client bank data is personal data for sole traders and landlords; read the ICO guidance on AI and data protection and confirm what your software provider does with the data, including whether it is used to train shared models.

A bank feed shows an amount, a date and a description. It does not show whether a purchase was personal, whether an asset should be capitalised, or whether a payment from a director is a loan or income. AI bookkeeping can only guess at those from patterns.

ICAEW's generative AI guide for accountants covers opportunities alongside risks and limitations, ethics, legal considerations and cyber security. It is worth reading before extending generative tools from drafting into anything that changes the ledger.

Receipt capture reduces typing but still needs a person to confirm the supplier, VAT amount and whether the receipt is a valid VAT invoice. A clear photo of the wrong document is still the wrong document.

None of this argues against automation. It argues for knowing where the boundary sits, so the time saved on routine lines is spent on the judgement calls.