Batch startup at a food products plant — the first label is where the shift is won or lost.

The problem

Batch startup is where the chain breaks — not steady-state production.

In a multi-SKU food products plant, the line at steady state is stable: once it starts, pack after pack comes out with few incidents. The problem is the SKU changeover, which happens several times per shift and concentrates three changes at once:

• Label reel — the ingredient declaration, allergens and format change. The previous batch's label should never appear on the new one.
• Date code — expiry date recalculated for that SKU, following whichever retailer rule applies (different weeks for private label vs. own brand).
• Seasoning recipe — sesame, cumin, whole wheat, no salt, tomato. The silo and the doser change; the intermediate cleaning should be recorded before starting.

The production order arrives on paper from the office (sometimes late, sometimes with a change scribbled on top in pen). The operator reads it, cross-checks it mentally against the first label out of the printer, and decides whether to start or call the shift lead. It is the natural bottleneck of the shift: decide right and nothing happens; decide wrong and the brand can end up in a recall for mislabeling. And they almost always do it alone, by the light of the control panel, with the line stopped and waiting.

How it fits the IRIS system

Connect in manual mode — the operator's phone digitizes the paper and the pack, without touching the line.

Batch startup does not need a new screen on the line or a modified PLC. It needs the information that already exists on paper (the office production order) and the information that already comes out printed (the first label) to reach the system at second zero, without retyping it and without depending on the operator's memory. That is what Connect in manual mode is for.

The operator photographs the paper production order. Connect digitizes it by OCR. They photograph the first label. Connect reads it and cross-checks it against the expected recipe. If something is off, it alerts at that very moment. The person signs — the line does not restart on its own.

How manual Connect operates at food product batch startup:

• Photo of the paper production order — the operator or shift lead photographs the production order that came down from the office with their phone or tablet. Connect digitizes the SKU, the expected recipe, the quantity, the customer and the expiry date to apply. In seconds, the order is in the system without anyone typing it in.
• Photo of the first printed label — before releasing the first pack, the operator photographs the first label out of the printer. Connect reads the ingredients, the declared allergens, the batch code and the printed expiry date by OCR.
• Cross-check against the plant recipe wherever it lives — Connect cross-references the digitized production order and the label with the silo recipe, wherever it lives: a modern ERP, a vertical bakery MES or the Excel file the R&D manager keeps in a shared folder. Nothing needs migrating; it connects where the data is.
• Alert in the earpiece — if the current batch's label declares “no tree nuts” but the previous batch's seasoning contained almond and the intermediate cleaning is not recorded as complete, Connect alerts the operator instantly with a short phrase: “careful — this label declares no tree nuts, but the cleaning after the almond batch is not recorded as closed.” The person validates or calls the shift lead; the line does not start on its own.
• Startup file saved — the photo of the production order, the photo of the first label and the recipe cross-check are archived per batch. When the IFS/BRC auditor asks for the startup file from February 14, there is nothing to reconstruct: it is right there.

See the full IRIS architecture →

Before and after

Eyeball manual startup vs. digitized startup with Connect

Aspect	Classic startup by the light of the panel	With iLEAN Connect in manual mode
Production order	On paper, read by hand, sometimes with pen edits on top	Phone photo — digitized by OCR in seconds
First label of the batch	Cross-checked from memory against the order, operator's call	Phone photo — cross-checked against recipe and order at second zero
SKU changeover on the line	20-25 min with three paper checklists	4-6 min with two photos and an alert if something deviates
Seasoning / silo recipe	R&D Excel that does not always reach packing	Connect reads it wherever it lives — ERP, MES or the Excel itself
Incident alert	The operator goes off to find the shift lead	Instant alert in the earpiece or phone — one short phrase
File for the IFS/BRC auditor	Rebuild the startup by hand, weeks later	Photo + cross-check archived per batch, automatically

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.

• Food products plant: 1-2 packing lines, 6-10 SKUs per shift, reel and recipe changes several times a day.
• Manual Connect pilot on batch startup — just the operator's phone/tablet, no construction work on the line. First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on the frequency of documented startup incidents (wrong label, misapplied expiry date, crossed seasoning) and the average cost of a recall or rework at your retailer.
• Expected reduction in startup incidents of ≥30% in the first three months, from simple automatic cross-checking at second zero. Estimate to be validated with your data.
• The hard lever is a single avoided recall: product pulled from the shelf, reverse logistics, destruction, brand damage. One recall pays for the pilot.

And the production manager's reasonable doubt

“What if the OCR misreads the label and lets a batch through with the wrong declaration?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recontextualizes a specific piece of data from one medium to another (reading the printed label and comparing it with the expected recipe), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: Connect holds the startup 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.