AI allergen labeling in cheese dairies — a mistake is not a fine, it is a recall.

The problem

Three realities that are almost never in the same place.

A correct cheese label is the cross-reference of three data points that, in most plants, each live on their own island:

1. What is in the batch — the recipe. Usually in the ERP, in an MES, or in a spreadsheet whose last editor never told packing about it.
2. What ran before on the line — the production sequence and the intermediate cleaning. Often in another system, sometimes only in the shift lead's head.
3. What the regulation requires — EU Regulation 1169/2011 and its 14 mandatory allergens, plus the retailer's internal rules if you export to the UK or the US.

The quality manager knows this, but cannot be in three places at once. The operator has three screens and two checklists, and the SKU changeover is precisely the most vulnerable moment. The classic system works 99% of the time. That 1% is the one that ends up in RASFF — and every time it is a recall, not an incident.

How it fits the IRIS system

iLEAN does not add a fourth system — it seals the cracks between the three you already have.

The labeling problem is not a lack of information: it is information that lives on islands and that, at the critical moment (the SKU changeover, the upstream line that changes recipe without warning), does not reach the decision-maker in time. iLEAN acts as the putty that fills those gaps, without asking you to change your ERP, your MES or your packing line.

Edge sees the label before sealing. Connect reads the recipe wherever it lives — ERP, MES, a forgotten spreadsheet. The agent cross-references it against EU 1169/2011 and, if something does not add up, holds the batch. The person signs — never the other way round.

The three iLEAN pieces applied to allergen labeling in a cheese dairy:

• Edge — a terminal with machine vision (CNN) over the packing line. It reads the printed label in milliseconds, extracts the list of declared allergens and triggers the actuator (rejecter, stack light) if it does not match what was expected. It works without a network. If the plant loses WiFi, Edge keeps reading and holding.
• Connect — captures the batch recipe whether it comes from a modern ERP, a vertical MES, or a spreadsheet the R&D manager updates every morning in a shared folder. And it also captures what arrives from outside (an email from the importer with the new US retailer rule, a WhatsApp message from the supplier about a change in the rennet) at second zero, without anyone forwarding anything.
• Agent — cross-references the recipe, the line's history (what ran before and with what cleaning), the image of the label and EU Regulation 1169/2011. If there is a deviation, it does not send an email at 10 p.m.: it holds the batch and alerts the quality manager through whichever channel they use. The person validates and signs; the line does not restart on its own.

See the full IRIS architecture →

Before and after

Manual labeling vs. cross-referenced labeling with iLEAN

Aspect	Manual labeling + checklist	With iLEAN Edge + Connect + Agent
Source of the recipe	ERP, MES or spreadsheet — the operator checks one	All three unified at second zero
SKU changeover	Paper checklist, shift lead's signature	Edge verifies the label + the agent verifies the intermediate cleaning
Cross-contact	Assumption: "if we cleaned, we're fine"	Cleaning traceability cross-referenced with the production sequence
Error detection	Later audit or customer complaint	Before sealing, on the line, in milliseconds
Operation without a network	n/a	Edge keeps running on its own power
File for the IFS/BRC auditor	Rebuilt by hand, weeks of work	Per-batch dossier, automatic, with a photo of the label

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• A mid-sized aged-cheese plant, 2 packing lines, multi-SKU with a variable recipe (with and without tree nuts, with and without egg lysozyme in long-ripened cheeses).
• Edge pilot on one line (camera over the label + actuator + integration with the batch recipe in the ERP/MES). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on how often labeling incidents have been documented over the last few years and the average cost of a recall or rework with your retailer.
• The hard lever is a single recall avoided: product pulled from the shelf, reverse logistics, destruction, brand damage. One recall pays for the pilot.

And the quality manager's reasonable doubt

"What if the AI gets it wrong and lets a mislabeled batch through?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recontextualizes a piece of data from one system into another (reading the label and comparing it with the recipe), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds the batch and the person signs. The three safety rings exist precisely for this.

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