Invoices matched to purchase orders, with every exception escalated rather than absorbed.

A finance team reconciles supplier invoices against purchase orders and delivery records at the end of every month. The volume is high, the work is repetitive, and almost all of it is a comparison between documents that live in different systems.

The team is experienced and the process is well understood. What it lacks is time: the reconciliation window is fixed, the volume is not, and the same people are responsible for the exceptions and for the reporting that depends on them.

Most of the work is not judgment. It is retrieval and comparison — finding the purchase order, finding the delivery note, checking the numbers, checking the terms. Experienced people spend their month doing it because there is nobody else to do it, and because the small proportion that does require judgment is genuinely consequential.

That proportion is the problem with automating naively. A mismatch might be a typo, a partial delivery, a price change nobody recorded, or a supplier billing for something that was never ordered. Those look identical in the data and are completely different in consequence, and one of them involves paying money that should not be paid.

The other failure mode is an automation nobody trusts, where every match gets reviewed anyway. That is slower than the manual process it replaced, because now there are two systems to reconcile instead of one.

The workflow contains nine decisions. Eight of them are comparisons. One of them moves money. Do these documents agree? And if they do not, is this a clerical difference or a payment that should not be made?

Matching an invoice to a purchase order is low judgment and, done correctly, recoverable: a wrong match surfaces at the next check and can be undone. Approving a payment above the agreed threshold is the opposite — it may be mechanical to execute, but it is consequential and effectively irreversible once the payment leaves. Rating those two decisions on the same scale is how automations end up approving things they should not.

The coworker would match invoices to purchase orders and delivery records across the four systems involved, chase missing supplier documents on a defined schedule, and queue anything that does not reconcile with a note explaining precisely which field disagrees and by how much. Nothing above the agreed threshold is approved without a person.

The work stays in the finance team’s existing systems. Matched items are marked in the ledger they already use, exceptions land in the queue they already work, and chasing goes out from the address suppliers already recognise. Before clearing an exception, the reviewer reads a note explaining why it did not reconcile — which field disagrees, by how much, what the coworker checked, and which of the plausible explanations it could not rule out.

The comparison work would stop consuming the month. The team spends its time on the exceptions, which is the only part of the workflow that needed experienced people in the first place, and the reconciliation window stops depending on how much volume arrived.

Suppliers change their document formats without telling anyone, which is the most common cause of a matching failure and the least interesting to diagnose. We watch match rates by supplier so a format change shows up as a pattern rather than as a bad month.

Repeat exceptions of the same shape usually mean a rule is missing rather than a supplier is at fault, and those become changes to the decision map. Thresholds also move as the business grows, and moving one is a decision recorded with the client, not a setting someone adjusts.