Zero labeling recalls — iLEAN's agentic system for breadstick and crispbread snacks.

The problem

The most silent and most expensive failure of any food plant.

It is the most silent and most expensive failure of any food plant, and it can happen to all of them: the product was made perfectly, but goes out to the shelf with the wrong label, a misprinted expiry date, an incorrect batch or an allergen declaration that does not match that formulation. It is not a manufacturing failure: it is an information failure, and it is today the sector's leading cause of product recalls.

It happens, above all, at SKU changes, when the previous format crosses with the next one: a label reel not changed in time, a date code left over from the previous shift, a correct product inside a package announcing something else.

The classic system (an attentive operator + a paper checklist + the shift lead's glance) works 99% of the time. And that 1% is what ends in a RASFF alert, a shelf recall, a call from the retailer at 10 pm. The cost is not the sanction: the cost is the recall, the rework, the reverse transport, the destruction of well-made product and the brand damage that shows on no invoice but weighs for years.

1. Batch startup — Connect capture. When the batch starts, the operator photographs the production order document about to run on the line and the first product label to be printed for the whole batch. The AI digitizes both documents and stores them in central memory.
2. Coherence verification — central AI + a personal assistant in the earphone. The central AI verifies that the production order and the printed label correspond, and guides the operator step by step, through a personal AI assistant in the earphone, to review and confirm one by one that specific label's critical points.
3. 100% in-line inspection — Edge + JIDOKA AI. The AI terminals installed on the line review every package by vision (dozens of times per second) and verify that the label, the expiry date, the batch and the allergen information are correct, automatically extracting any non-conforming unit (JIDOKA AI) without stopping the line, or even stopping it if the error is repetitive or a manager requests it through their personal assistant in the earphone.
4. Traceability to the ERP — legacy integration. The AI integration system pushes each batch's traceability directly to the ERP (what was printed, on which unit and at what time), leaving a complete and auditable record, without paper and without manual transcriptions.
5. Assisted SKU change — SMED AI. The SMED AI system oversees the SKU change and verifies that it was executed per the set standard, on time and in form, detecting any deviation the instant it happens, not days later.
6. Legacy machine capture — graduated Connect. All this process information is captured by AI even from old machines with no interface, and is automatically integrated with the company's legacy systems (ERP, MES, labelers, coders).
7. Batch release — only if everything adds up. The line does not release product for shipment until the system confirms that every package in the batch is conforming. The gap through which the vast majority of recalls slip simply ceases to exist.

The real result: this system eliminates the rework, the relabeling, the stops at format changes and the customer rejections at receiving — recovering line hours every shift — and, above all, closes at the root the exact point where most incidents happen: not a failure in making, but an information failure on a well-made product. It is not an insurance you might use once a year: it is measurable saving every day and, at the same time, the best possible prevention of the incident no plant director wants to live through.

See the full IRIS architecture →

Before and after

Classic labeling vs. the iLEAN agentic system

Aspect	A line with classic control	A line with the iLEAN agentic system
Batch startup	The operator reads the order, checks the first label by eye	A photo of the order + first label; the AI crosses and guides the operator by earphone
The SKU change	A paper checklist, the shift lead's signature	SMED AI oversees the sequence and detects the deviation on the instant
Per-package verification	Manual sampling every N units	100% vision of the batch, dozens of inspections per second
A non-conforming package	Reaches the end of the line, detected late or not at all	JIDOKA AI: automatic ejection without stopping the line
Traceability to the ERP	Manual transcription at the end of the shift, on paper or Excel	A direct push — what printed, on which unit and at what time
Old coders and labelers	"They do not integrate, it costs more than it is worth"	Graduated Connect captures them (a panel photo, a serial read, direct integration)
The batch's release	On trust in the process	Only when the system confirms every package in the batch is conforming
The frequency of label recalls	1 every N batches — the 1% that slips to RASFF	Zero, by the system's design
The dossier for IFS/BRC	Rebuilt by hand, weeks	A per-batch file, automatic, with an image of every package

Impact estimate

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

• Breadstick/crispbread plant with 2-3 multi-SKU packing lines (salted, whole wheat, with sesame, allergen-free), coders and labelers of several generations.
• A pilot on one line with the complete combination — Connect capture at batch startup, Edge with JIDOKA AI on every package, SMED AI on the SKU change, ERP integration. First value within a few weeks.
• Indicative payback between 3 and 9 months. The hard lever is triple: (1) daily savings on rework, relabeling and hours lost at format changes, (2) elimination of customer rejections at receiving, (3) a single recall avoided pays for the whole pilot — product recovered from the shelf, reverse transport, destruction and brand damage.
• The pilot line operational in 2-3 months; replication to the remaining lines executed by your own people trained during the pilot, over the following months.

And the fair question: "what if the system itself errs?"

The quality manager's right question. Two answers that hold each other up. The technical one: hallucination is a problem of free generation, not of anchored tasks — reading a printed label and comparing it with the production order is as anchored a task as they come. In tasks of this kind, the best models brought the error below 1.5% ^[1]. The architectural one: the iLEAN system is built on the three IRIS safety rings — Connect transports, the Agents decide, the person signs. The critical verification never executes alone. It is not that everything passes through a person; it is that the system allows it where it matters, and here it matters.

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