One wrong keystroke can no longer stop the customer's line

The problem

A whole pallet of polypropylene labelled as polyethylene is not a misprint.

The bagging line and its print-and-apply labeller are relatively modern, five to fifteen years old, and they already have an API or a command interface. But they are isolated from the ERP. So at every batch change somebody keys the same data in a second time on the labeller panel. It was never integrated properly because it was never the priority of the quarter. And one wrong keystroke means a full pallet of polypropylene ships labelled as polyethylene. That is not a misprinted label: that is a line stoppage at the customer, a truck coming back and a non-conformity opened. The only barrier that exists today is the attention of one person keying in a value that was already written somewhere else.

• The bagging station and its print-and-apply unit are relatively modern and already have an API or a command interface. But they are isolated from the ERP.
• So at every batch change somebody types the same data a second time on the labeller panel.
• It was never integrated properly because it was never the quarter's priority.
• And one wrong keystroke means a whole pallet labelled with the wrong polymer — which is a line stoppage at the customer.

How it fits the IRIS system

Connect stitches both systems via API. Without replacing either machine.

Connect stitches the two systems together over API. When the ERP marks the bagging order as started, it automatically publishes polymer, grade, batch, target weight and certificate reference to the printer. A batch change becomes one tap in the ERP, not a keystroke on a panel. Neither machine gets replaced. And there is a second-order effect worth more than the first: if the batch has no validated lab result, the label is simply not issued. It becomes physically impossible to label, and therefore to ship, an unreleased batch. That is where this joins up with the laboratory example.

And there is a second-order effect worth more than the first: if the batch has no validated lab result, the label is simply not issued. Shipping without release becomes physically impossible.

See the full IRIS architecture →

Before and after

Labelling, before and after

Aspect	Today	With iLEAN Connect
Polymer, grade, batch and weight	Typed on the panel	Published from the active order
A batch change	Typing on the labeller	One tap in the ERP
Human labelling error	Possible at every change	Removed at source
Batch with no validated lab result	Can be labelled and shipped	The label is not issued
Do the machines have to change	—	No: their own interfaces are used
New middleware to maintain	—	None

Impact estimate

Impact estimate — to validate against your numbers.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 3-6 months — one of the fastest of the set.
• The integration is bounded and the risk it removes is high: zero human labelling error.
• No unreleased batch can leave labelled, which is a physical barrier rather than a procedure.
• And the batch change stops depending on who is standing at the panel.

Payback 3-6 months · zero human labelling error and no unreleased batch ever labelled. This is one of the fastest returns in the pool, three to six months, because the integration is narrow and the risk it removes is expensive. What drops sharply is the return caused by mislabelling and the non-conformity that comes with it. An estimate to validate against the plant's labelling incident history. Conservative ranges, an estimate to validate against the plant's real data.

And the fair question from the production manager

"Is this another system to maintain?" — no, and that is deliberate. No middleware is introduced: what gets used are the interfaces the machines already ship with and that are already paid for. The work is stitching them to the active order, not adding a layer between them.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.