Most businesses we work with have already made the first move. The bank feeds in, reconciliation is largely automatic, and the ledger is broadly current. That work is well covered in our guide to getting more from Xero, and it is the right place to start.
But it is only the sales and banking side. Walk into the same finance function and look at what happens when a supplier invoice arrives, and the picture is usually very different: a PDF in an inbox, forwarded to someone for approval, keyed in by hand, and paid by rekeying the details into online banking.
That is where the remaining cost sits. It is also where the risk sits.
Why the purchase side matters more
Three reasons, in ascending order of importance.
It is the larger volume. Most businesses issue fewer sales invoices than they receive purchase invoices, and the sales ones are generated by a system while the purchase ones arrive in whatever format the supplier chose.
It is where the control failures are. Sales invoicing errors cost you money slowly. A purchase ledger without enforced approval limits, a documented audit trail and segregation between approving an invoice and paying it is how businesses lose money quickly. Invoice redirection fraud — a genuine supplier’s bank details changed by someone who is not the supplier — remains one of the most common losses in owner-managed businesses, and it succeeds precisely where payment details are edited manually.
It is what the 2029 e-invoicing mandate actually changes. When invoices arrive as structured data rather than PDFs, the businesses that already process them systematically will absorb it. Those still keying them in will be rebuilding the process under a deadline.
What the purchase-to-pay chain looks like automated
Capture. Supplier invoices arrive in one place — a dedicated inbox or supplier portal — and are read automatically, with line detail, VAT and supplier matched against the ledger. The job becomes reviewing an extraction rather than typing an invoice.
Matching. Where you raise purchase orders, invoices are matched to the order and the goods receipt automatically, and only exceptions reach a person. If you do not raise purchase orders and spend is material, that gap is worth closing before automating anything else.
Approval. Invoices route by rule — by value, by cost centre, by supplier — with authority limits enforced by the system rather than remembered by the finance manager. Approvals happen on a phone, which is the difference between a two-day cycle and a two-week one.
Payment. Approved invoices flow into a payment file that goes to the bank without anyone retyping a sort code. Supplier bank details are locked and changes require verification, which closes the redirection fraud route.
Expenses and card spend. Card transactions matched to receipts captured at the point of spend, which ends the month-end scramble and materially improves VAT recovery.
Where does AI genuinely help?
Selectively, and not where it is most heavily marketed.
It is good at extracting data from unstructured documents, coding transactions from historic patterns, and — most valuably — flagging anomalies. A duplicate invoice. A supplier payment well outside its usual range. A VAT treatment inconsistent with the last twenty transactions of the same type. A change of bank details on an invoice from a long-standing supplier.
That last category is the real prize, because these are the things a person scanning a ledger would not catch. It is also a very different proposition from asking software to make judgements about the substance of a transaction, which it should not be doing unsupervised.
The right mental model is a fast, tireless assistant whose work gets checked, not a replacement for a bookkeeper.
What goes wrong
Automating a broken approval process. If nobody is clear who approves what, software will simply route invoices to the wrong person faster. Agree the authority matrix first; it takes an afternoon.
No exception handling. Every implementation produces invoices the system cannot match. If there is no defined route for those, they accumulate in someone’s inbox and the whole thing quietly reverts to manual.
Supplier data left untidy. Duplicate supplier records, inconsistent naming and missing VAT numbers will undermine matching however good the tool is.
No owner. Where automation is nobody’s job, exceptions build up until someone is running manual workarounds for a third of transactions — and you have the licence cost without the benefit.
What is the return?
For a business processing a few hundred purchase invoices a month, the direct time saving is usually measured in days per month. But the more durable gains are the ones that do not appear on a spreadsheet: knowing your true creditor position in real time rather than at month end, an approval trail that stands up to an auditor or an insurer, and one fewer route for a fraudulent payment to leave the business.
And when structured invoices become mandatory in April 2029, the change will be a format update rather than a project.
Frequently asked questions
We already use Xero. Isn’t that enough?
Cloud accounting handles the ledger, bank feeds and reconciliation well. It does not, by itself, capture supplier invoices, enforce approval limits or produce payment files. Those are usually separate capabilities, either add-ons within the same ecosystem or connected apps, and they are where most of the remaining manual effort sits.
What should we automate first on the purchase side?
Invoice capture, then approval workflow. Capture removes the keying; approval is where the control benefit sits. Payment file generation follows naturally once the first two are working. Do them in sequence rather than together.
How does this help with invoice fraud?
Invoice redirection fraud relies on someone manually changing payment details. Where supplier bank details are locked in the system, changes require verification, and payment files are generated from approved records rather than typed into online banking, that route is substantially closed off.
Is AI reliable enough to process supplier invoices?
For data extraction, pattern-based coding and anomaly detection, yes — with human review. It should not be making unsupervised decisions about the substance of a transaction. Treat it as an assistant whose output is checked rather than a replacement for a bookkeeper.
Does this prepare us for the 2029 e-invoicing mandate?
Substantially. The mandate replaces PDFs with structured data arriving directly into your system. A business that already handles inbound invoices through a defined capture, matching and approval process will experience that as a format change. A business still keying invoices in will be redesigning the process under a deadline.