AI Bank Reconciliation Prep

It prepares the reconciliation; a person completes it. That distinction is not lawyerly hedging — it is the whole design. A reconciliation is a control: its purpose is to independently confirm that the ledger agrees with an external record, and a control that a machine both performs and signs off on has stopped controlling anything. So the AI does the labour — reading the statement, matching what genuinely matches, grouping the routine, isolating the ambiguous, and explaining each proposal — and your team confirms it. In practice a reviewer moves through the high-confidence bulk quickly and spends their attention on the short list of genuine questions, which is exactly where it should have been all along.

Force-matching is the failure mode we design against hardest, because it is worse than doing nothing. A statement line and a ledger entry that share an amount and a date are not necessarily the same transaction — two identical invoices to the same supplier, a payment and a refund of equal value, a transfer that appears on both sides. Software that matches on coincidence produces a rec that balances and is wrong, which is the most expensive outcome available because it has consumed the review that would have caught it. The AI matches on the strength of the evidence: reference numbers, payee, amount, timing and the file’s own history together. Where the evidence supports one answer, it proposes it with the reasoning shown. Where two candidates are plausible, it presents both and asks. Where nothing fits, it says nothing fits.

These are the bulk of the manual time, so they get specific handling. A single deposit covering fourteen invoices is decomposed and matched against the open items that reconcile to it. A part payment is matched to its invoice with the balance left open rather than the invoice closed. Bank fees, merchant fees and interest that arrive as separate small lines are recognised and grouped rather than presented as fourteen mysteries. Transfers between a client’s own accounts are matched as a pair so they do not get counted as income and expense. Foreign currency receipts, where the bank credits an amount that never equals the invoiced figure, are matched with the difference isolated as an exchange item for your team to treat. None of this is exotic — it is just tedious, which is precisely the argument for automating it.

They go on a list, with a chase attached. An unidentified statement line is a question for the client, not a coding decision for your bookkeeper to invent an answer to. The AI identifies the line, works out what it needs, and asks the client directly with the date, amount and payee — the details that make a client actually able to answer. Persistent unknowns are escalated to your team rather than parked in a suspense account. The suspense account is where reconciliations go to die: everything that could not be resolved on the day accumulates there until someone inherits a five-figure balance and a mystery. Making unknowns visible and chased, continuously, is the fix.

Either. Where the client’s file has a working bank feed into Xero, MYOB or QuickBooks, the AI works with the transactions as they arrive, which is the tidiest arrangement. Where there is no feed — an older facility, a merchant platform, a loan account, or a client who simply will not authorise one — statements can be imported and read, including PDF statements, which the same document engine handles. What matters more than the channel is continuity: reconciliation done weekly against current data is a modest task, and reconciliation done in one sitting three months later is an archaeology project.

It changes the shape of month-end more than it changes a single number, and we would rather describe the mechanism than quote you an hours-saved figure we cannot verify for your firm. The manual rec is dominated by volume — hundreds of lines that were always going to match, worked through by hand to find the dozen that do not. Automating the matching means the dozen are identified continuously, as they arise, rather than excavated at close. The reviewer’s time goes almost entirely to genuine questions. The secondary effect is usually the bigger one: problems surface within days instead of at close, when they are still cheap to fix and the client can still remember what the payment was for.