Plant-based beverage control with AI — a cross-contact allergen gives no warning; it shows up in RASFF.

The problem

A line shared across five allergens does not forgive a half-done CIP.

The typical plant-based beverage plant has grown by stacking SKUs on the same infrastructure: one UHT, one aseptic filler, one bottling line. That is good lean — but it accumulates risk:

• Interim CIP between oat → almond — switching from a gluten-containing cereal (non-certified oat) or tree nuts, depending on the recipe; a short rinse leaves traces.
• Variable fortification — calcium, vitamin D and vitamin B12 at different dosing rates per SKU, on pumps that are sometimes shared.
• A small special-recipe batch — the operator loads the previous SKU's recipe out of habit; the label declares one thing, the tank carries another.
• An aseptic stage that forgives nothing — if the CIP is incomplete, the microbiological problem shows up late, in the market.

The quality manager knows it, which is why there are checklists, the CIP PLC and sampling. It works 99% of the time — and the remaining 1% is what ends up in AESAN, RASFF or with a retailer pulling the batch. Every allergen recall is hard cost (product off the shelf, reverse transport, destruction) plus brand cost, which weighs more in the medium term.

How it fits the IRIS system

iLEAN does not add a fourth UHT — it seals the cracks between the recipe, the CIP and the label.

The problem is not a lack of control: it is information in islands (recipe in the ERP, CIP in the PLC, fortification dosing in Excel, label in yet another system) that nobody cross-checks in the very second of the SKU changeover. iLEAN acts as putty between the systems you already have, without asking you to change the UHT, the aseptic unit or the bottling line.

Connect reads recipe, CIP and fortification. Edge sees the label before palletizing. The agent checks against the allergen regulation and, if something does not add up, holds the batch. The person signs — never the other way around.

The iLEAN pieces applied to plant-based beverage control:

• Connect — captures the CIP report from the PLC (cycle, temperature, conductivity, rinse duration), the actual fortification dosing, the recipes from the ERP/MES and the production orders. It also captures what arrives from outside (a calcium supplier change by email, a retailer audit alert by WhatsApp) at second zero.
• Agents — cross-check the data at every SKU changeover: did the CIP cover the allergens to be removed? Is fortification in range for the declared SKU? Does the label loaded in the printer match the actual recipe? If something does not add up, they hold the batch before the aseptic stage and alert the quality manager. They do not send an email at 10 p.m. — they block the startup and reach the manager on their live channel.
• Edge — a vision terminal (CNN) over the packing line. It reads the printed label (allergen list, declared fortification, batch date) and checks it against the batch's actual recipe and the CIP dossier. Works without a network: if the plant loses WiFi, Edge keeps reading and holding.

See the full IRIS architecture →

Before and after

Classic control vs. cross-checked control with iLEAN

Aspect	Checklist + PLC + sampling	With iLEAN Connect + Agents + Edge
CIP between two plant-based SKUs	Cycle launched, validated after the fact	Validated against the next SKU's recipe, before startup
Fortification	Dosing read on a sheet	Read in line, checked against the declared recipe
Label vs. recipe	Visual check at SKU changeover	Edge reads 100% before palletizing
Small special batch	Risk of loading the previous recipe	The agent confirms recipe vs. order and vs. label
Operation without network	n/a	Edge keeps operating with its own light
File for the IFS/BRC auditor	Manual reconstruction	Per-batch dossier, automatic

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.

• Plant with 1-2 UHT + aseptic lines shared across oat, almond and soy (at least), fortification with calcium + vitamins, exporting to several markets.
• Connect + Agents pilot on CIP/recipes + Edge over the packing line. First value expected within a few weeks: the first block of an SKU changeover with an insufficient CIP or a swapped label.
• Indicative payback between 4 and 9 months. Reduction of recalls/incidents from undeclared allergens ≥30% (sub-sector inferred estimate, to be validated).
• The hard lever is a single avoided allergen recall: product recovered from the shelf, reverse transport, destruction, brand damage. One recall pays for the pilot.

And the quality manager's reasonable doubt

“What if the AI gets it wrong reading a label or a CIP report?” — 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 to another (reading the label, reading the CIP report, comparing against 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 are there precisely for this.

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

Related cluster: allergen labeling in cheese dairies, yogurt fermentation, nutrition labeling.