AI Invoice Processing

Traditional OCR reads characters off a page and tries to find fields by where they sit on the layout — which works until a supplier moves their totals block or sends a two-page invoice with the summary on page two. AI extraction reads the document the way a person does: it understands that "Inv #", "Invoice No." and "Tax Invoice Number" are the same concept, that the number in the bottom-right is a total because of what surrounds it, and that a line item is a line item even in an unfamiliar table layout. The larger difference is what happens next. OCR hands you fields to check. This reads the document, validates it against itself (do the line items sum to the subtotal, does the GST calculation hold), checks it against the client file (is this a duplicate, does this supplier’s coding history make sense here), and only then presents it — with a confidence position and, where it is unsure, the reason.

Native PDFs, scanned PDFs, images, and photos taken on a phone in a ute at dusk — which is the realistic test. It handles multi-page invoices, invoices with the tax summary on a later page, statements containing multiple invoices, credit notes, and bills that arrive as text in the body of an email rather than as an attachment. Quality does matter: a sharp native PDF extracts more reliably than a creased fax of a photocopy, and where a document is genuinely too degraded to read with confidence, it says so and routes it to a person rather than producing plausible-looking numbers. That last behaviour is the one worth caring about — a bad read that announces itself costs you thirty seconds, and a bad read that does not can cost a client a lot more.

It runs the arithmetic and consistency checks a careful person would run, and surfaces anything that fails. That includes whether line items sum to the stated subtotal, whether the GST amount is consistent with the taxable component, whether the document carries the features you would expect of a tax invoice, and whether the supplier ABN on the document matches what the file has recorded for that supplier. Where something does not reconcile, the invoice is flagged with the specific discrepancy rather than posted. What it does not do is make the professional determination — whether a supply is taxable, GST-free or input-taxed in a client’s particular circumstances is a judgement for your team, and the AI’s job is to give them a clean document and a clear flag, not an opinion.

We would rather show you on your own documents than quote you a number, and here is why that is not evasion. "99% accurate" is close to meaningless without knowing what was counted: field-level accuracy on clean native PDFs from a fixed set of suppliers is a very different measurement from document-level accuracy across whatever actually turns up in a firm’s intake. A system can score brilliantly on the first and be unusable on the second. The metric that matters operationally is how much of your volume goes straight through without a human touching it versus how much lands in the exception queue — and that varies enormously by client mix and document quality. Bring a representative sample of real invoices to the consultation, including the ugly ones, and we will run them.

Those are the ones that pay for the system. A supplier bills for eleven units when ten were delivered, applies GST to a GST-free item, invoices at a price that does not match the agreed rate, or sends the same invoice twice with a different number. Because extraction is line-level rather than just grabbing the total, these become visible: the AI checks the maths within the document, compares against the supplier’s historical pattern in that file, and checks for near-duplicates that a person scanning a pile would miss. It flags; it does not adjudicate. Whether to query the supplier, accept it, or hold the payment is your team’s call — but they get to make it before the bill is paid instead of finding it in a review six months later.

Attached to the transaction it created, in the client’s accounting file. This is not a filing nicety — under ATO record-keeping rules the source document generally needs to be retained for five years, and the practical failure mode in most firms is not that documents are lost but that they are separated from the transactions they support, so nobody can find the right one under time pressure. Keeping the document attached to the entry it produced also means the audit trail is complete: the invoice, the extracted data, what the AI proposed, who reviewed it, and what they changed all sit together.