Accounts Payable Automation

It learns from the client file it is working in. Before it codes anything, it reads the client’s existing coding history — how the last twelve months of bills from that supplier were treated, which accounts in that specific chart of accounts were used, and which GST treatment was applied. A bill from a fuel supplier that has always been coded to Motor Vehicle Expenses with GST on Expenses gets the same treatment, and the AI shows you that history as its reasoning. Where a supplier is new, or the bill looks different to the established pattern (a capital purchase from a supplier who normally sends consumables, for example), it does not guess quietly — it flags the bill as an exception for a person to code. The coding rules stay yours: you can lock a supplier to an account, split by percentage, or require review on every bill over a threshold.

No. It prepares the payment run; it never releases money. The AI captures bills, codes them, matches them to purchase orders or delivery dockets where those exist, checks for duplicates, and assembles the batch — due dates, discount windows, remittance details. A human then reviews and approves inside your accounting software or banking platform, using your existing authorisation controls and bank security. We deliberately do not sit in the payment authorisation chain: separation of duties is one of the few controls that genuinely protects a firm and its clients against fraud and error, and automating it away would weaken exactly the thing that makes the process safe.

They go into an exception queue with the reason attached, rather than being coded on a guess. Typical exceptions: a supplier the file has never seen, an amount well outside the historical range, a bill that appears to duplicate one already entered, a tax treatment that does not match the supplier’s pattern, an ABN that does not check out, or a document too poor to read reliably. Your team clears the queue, and each correction feeds back into the pattern for that client. The point of the confidence threshold is that you should be able to trust the coded pile without re-checking it — which only works if the uncertain items are pulled out honestly instead of buried in the volume.

Most firms use a dedicated intake address — clients and their suppliers email bills to it, or the client forwards them from their own inbox. From there the AI reads PDFs, scanned images and photos of paper invoices, and handles bills that arrive as email attachments or in the body of an email. Bills that suppliers deliver through e-invoicing or supplier portals can be forwarded in the same way. What matters more than the channel is that intake is consistent: the biggest single cause of month-end chaos in AP is bills arriving in five different places, half of them in someone’s personal inbox.

It is a replacement for the data-entry component of their day, which is usually the part they like least and the part clients value least. The work that remains is the work that requires judgement: resolving disputed bills, chasing suppliers about missing credits, deciding whether an expense is capital or revenue, managing cash flow across a payment run, and talking to clients about what their creditors position actually means. Firms typically use the recovered hours to move junior staff onto client work rather than to reduce headcount — partly because that work bills at a better rate, and partly because the people who understand the file are exactly the people you want reviewing exceptions.

Yes — each client file is a separate, isolated context. The AI codes against that client’s chart of accounts, that client’s supplier history and that client’s rules, and a bill captured for one entity cannot be posted to another. This matters for firms in a way it does not for a single business: you are handling other people’s financial records, and under the Privacy Act 1988 you are accountable for keeping them separate and secure. Access is controlled per staff member, and every action against a file is logged — who touched what, when, and what the AI proposed before a human changed it.