In almost every company where I look at the chain, the optimisation happens at the same end: the invoice. That is where the pain is visible, where the backlog piles up, where the people sit who have to chase. But the effort does not arise there. It arises weeks earlier, the moment somebody creates a purchase order that has nothing to do with the way the supplier will later bill.
An honest look therefore does not start with checking but with the order line — and it does not end with approval but with payment. Everything in between is a chain, and you do not judge a chain by its last link.
The one condition for automatic approval
An invoice passes without a human touch when three states of the same line agree: ordered, delivered or performed, invoiced. If all three match, the system can approve. That is the entire logic, and it is decades old.
What breaks it is the same list, equally decades old: different units of measure between order and invoice, partial deliveries without a matching partial order, price tiers not represented in the order line, ancillary costs such as freight or packaging without a line of their own, rounding differences of a few cents with no defined tolerance, variations nobody fed back into the order. None of these is an accounting error. All of them follow from an order built differently from the invoice that follows it.
The special case everybody has: the blanket order
Some purchases cannot be forced into lines with quantity and price, however hard you try. Maintenance, where nobody knows in advance what is broken. Services billed by effort. Consumables against a framework. That is what blanket or limit orders are for, and they are not sloppiness but a deliberate decision.
They do have one awkward property: they have no counterpart to match against. There is an amount and an account, but no line for an invoice row to run against. That is exactly why this kind of invoice ends up in the manual queue in almost every company — and because the volume is significant, it drags the automation rate down even when everything else runs cleanly.
Blanket orders are nevertheless manageable, under four conditions. First, they need value-based confirmation of what was performed, by the requisitioner — a service entry sheet or a value-based goods receipt, in the system rather than by email. Second, they need a tolerance that is defined rather than argued case by case. Third, they need a limit per period and per approver, so a framework does not quietly become a permanent state. And fourth, they need an automatic block on overrun that holds even when somebody promised otherwise on the phone.
With those four in place, a blanket order becomes a deliberately chosen control point rather than a hole in the chain. Without them, it is the reason the automation rate sticks at sixty per cent.
Why early-payment discount is the most interesting number here
The discount period runs from receipt of the invoice, not from approval. Every day a document sits in the approval loop comes off that period — and nobody notices, because the loss is never posted anywhere. It simply does not appear.
And it is large. Two per cent for paying in fourteen days instead of thirty net corresponds to an annual rate of roughly forty-six per cent: you forgo two per cent for sixteen days, and there are about twenty-two such periods in a year. There are few places in a company where money can be earned as reliably as by meeting a deadline that has already been agreed.
The consequence for process design is uncomfortable and simple: throughput time from receipt to approval is not an accounting metric but an earnings figure. Cutting it by four days moves more, in many companies, than the last round of price negotiations.
Large organisation, many people involved — is it even possible?
Yes, and not despite the size but because of it: in a company with many people involved, a clean rule pays off because it is applied often. What is needed is not less involvement but less discretion in the wrong places.
Concretely: an approval matrix held in the system rather than in a policy nobody reads. Deputy rules that take effect automatically, instead of an inbox filling up over a holiday. Escalation by waiting time rather than by someone chasing. And a decision about which deviations need a human at all — a ninety-cent difference on a twelve-thousand-euro invoice should not create a task, and in a surprising number of companies it does.
Where artificial intelligence helps, and where it does not
It helps where the chain is unstructured: invoices as PDF, as scan, as a photograph of a delivery note, in changing layouts. It extracts lines, matches them to an order, spots deviations, proposes an account assignment, finds duplicate payments and flags anomalies that a human processing three hundred documents a day will miss.
It does not help with master data, with a mess of units, or with an approval rule that does not exist. A model asked to guess which of four cost centres is meant will guess. Put the tool before the order and you get faster mistakes.
What an analysis actually delivers
Five results, all of them numbers rather than adjectives. Throughput time per stage, from requisition to payment. The share of invoices that pass without a human touch today. The share with deviations, broken down by cause and frequency. The discount forgone over the last twelve months, in euros. And a ranked list of measures by effect and effort that you can put in front of your management.
To get there I sit next to the people who do the work and time it with them. That is unspectacular and regularly the most revealing week of the whole engagement, because it usually turns out that the expensive step is a different one from the suspected one.
When it goes deeper into procurement
My strength is the chain and the question of what in it can be handled by machine. Where it turns to sourcing strategy, category management or supplier development, I bring in the procurement specialists at Metroplan. I would rather say that up front than pretend afterwards that I knew it all myself.