In any operation that transforms raw material there is always a gap between what should have come out and what did. It appears at month end, on a single line. And precisely because it appears on a single line, it almost never gets solved.
A total has no cause
Intuition says there is one place where the loss happens. You look for that place, you do not find it, and the explanation left over is "human error" — which is not an explanation, it is the absence of one.
In practice the loss is almost always distributed: three or four small origins that nobody notices in isolation. Added up at month end they make a big number. But the total has no cause; only the parts do. As long as you are looking at the total, there is nothing to fix.
What you need is not a better inventory
The usual answer to this problem is to count more often. A weekly inventory instead of a monthly one gives the same kind of number, just sooner: it is still a total.
What changes the game is different — the same material counted at every transition it passes through, inside the same system. On arrival. Entering production. At the cut. On the way out. That is not four inventories; it is four measurements of the same unit, which lets you subtract one from another and locate where the gap is born.
The distinction sounds technical and is economic: a total tells you that you lost; a sequence tells you where.
The first cause usually comes from outside
There is a pattern we have seen repeat: once you start measuring in and out, the first thing that surfaces is often nobody’s fault inside the building.
The material coming in may carry less than the label declares. The spec sheet may be out of date, specifying a consumption that no longer matches what the machine does. Neither of these is found by training the team better, and neither shows up in an inventory — because an inventory compares what you have against what you should have, not what came in against what was declared.
Where AI belongs — and where it does not
The hard part of this circuit is not the arithmetic, it is the format. Delivery notes as PDFs, photographed labels, spec sheets in spreadsheets whose columns change from supplier to supplier. That is where the models help: they read what arrives, normalise it to the same unit and flag when a number drifts from expected.
What AI should not do is decide. Switching supplier, changing a spec sheet or stopping a cut are decisions with commercial consequences and they stay with the people on the floor. The system brings the number and the context; the decision stays human — by design, not by caution.
When this is not worth it
If the material is cheap, if the volume is low, or if the circuit has few steps, the counting costs more than the loss it reveals. Measuring has a price too: someone records, someone verifies, someone maintains.
The arithmetic we do before building is simple: value of the material crossing the circuit per month, multiplied by a conservative estimate of the loss. If the result does not comfortably pay for the build and the upkeep, we say so.
The essentials
| Material loss is distributed — the month-end total is a sum, and a sum has no cause | 01 | |
|---|---|---|
| Counting more often is not enough; you need the same unit counted at every transition, in one system | 02 | |
| The first cause found usually comes from outside: material arriving short, or out-of-date spec sheets | 03 | |
| AI reads and normalises what arrives; the decision to change supplier or spec stays with people | 04 | |
| With cheap material or low volume, measuring costs more than the loss — so it does not get built | 05 |
Written by Pedro · Founder
Other notes

What happens to an email order when AI enters the loop
Almost every company we talk to has a version of this process: an email arrives with an order — sometimes a PDF, someti…

Accounting without copy/paste: invoices that book themselves — with approval
The pile of supplier invoices is one of the most universal processes there is: they arrive by email, as PDFs, each supp…

AI in hospitality: the processes that no longer need to be manual
Hospitality runs on attention to the guest — which is exactly why it hurts to watch front-desk teams copy bookings from…

