Automated nutrition labeling with AI — changing one ingredient shouldn't mean reviewing hundreds of labels.

The problem

A supplier change turns the labeling team's morning into a Sunday at the office.

The quality manager gets the notice at 9:00: the oil supplier has changed origin and the technical data sheet is slightly different. What for him is “one more data point” means four days for the labeling manager, going through Excel sheets one by one to identify:

1. Which SKUs use that ingredient — and here there is already room for one to slip through, because the recipe lives in the ERP, the percentages in an R&D sheet and the list of active SKUs in the retailer's ECM.
2. What changes in the nutrition panel — recalculating calories, fat, saturated fat, carbohydrates, sugars, protein, salt. Multiply by the new data sheet, apply rounding, apply tolerance.
3. Which claims fall — if the new fat pushes above the threshold of Regulation EU 1924/2006, the “low fat” claim is no longer legal and a reprint is due.
4. Which Nutri-Score changes — the Santé Publique France algorithm is public but has nuances (fiber, fruit/vegetables/legumes, sodium); recalculating it by hand is slow and error-prone.
5. Which retailer to notify — and with what lead time, because each one has its own in its label management system.

Four days later, the label ships. And sometimes an SKU slips through that nobody had flagged, because the recipe was duplicated in another folder. The classic system works 99% of the time. That 1% is what triggers the incident with the importer — and, in the worst case, the recall.

How it fits the IRIS system

iLEAN does not add one more Excel sheet — it seals the gap between recipe, regulation and print shop.

The bottleneck is not the nutrition calculation itself (that is arithmetic); it is that the recipe lives on one island, the regulation lives on another (EU 1169/2011, FDA 21 CFR 101, EU 1924/2006, Nutri-Score) and the print shop lives on a third. None of them talks to the next unless a person plays messenger. iLEAN is the putty that joins the three — without asking you to change your ERP, your MES or the label management system you already have.

Connect captures the recipe wherever it lives. The Writer agent recalculates the panel and the Nutri-Score and verifies claims. The Tracer agent detects which SKUs change and prepares the resend to the print shop. The person signs the list — nothing gets printed without a signature.

The iLEAN pieces applied to automated nutrition labeling:

• Connect — captures the recipe whether it comes from a modern ERP, a vertical MES, an R&D sheet, or the technical data sheet PDF the supplier sent by email. Critical information almost never arrives through an API; it arrives by email, WhatsApp or phone call. Connect integrates it at second zero, without anyone forwarding anything.
• Writer agent — recalculates the nutrition panel applying the rounding and tolerance rules of EU 1169/2011 or FDA 21 CFR 101 depending on the destination market. It generates the Nutri-Score with the official algorithm and verifies functional claims one by one against Regulation EU 1924/2006. If a claim falls, the agent alerts marketing before printing.
• Tracer agent — cross-references the change with the inventory of active SKUs, identifies every affected one (including those living in duplicated folders), flags which need reprinting, regenerates the artwork and prepares the package for your print shop or your label management system (Loftware, NiceLabel, BarTender). The quality manager signs the list. Without a signature, nothing goes out.

The brain (Brain) that orchestrates the agents lives in Central; critical data stays in the plant's ring 1, not in just any cloud (see the IRIS architecture and the three safety rings).

See the full IRIS architecture →

Before and after

Manual recalculation vs. automatic recalculation with iLEAN

Aspect	Manual recalculation + Excel	With iLEAN Connect + Writer + Tracer
Time from ingredient change to label ready	Days, sometimes weeks	Hours — the person only validates
Affected SKUs detected	The ones the team remembers	100% — including duplicates in other folders
EU 1169 vs FDA template	Two parallel calculations by hand	One recipe, two automatic outputs
Nutri-Score	Excel sheet with a homegrown formula	Official Santé Publique France algorithm built in
Claim verification (EU 1924/2006)	Manual check at the end, if there is time	Automatic — a claim that falls is flagged before printing
Delivery to print shop / label management	Email + attachment, package by package	Export to the right format (Loftware, NiceLabel, BarTender, PDF/X)

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 catalog. It gives the committee an order of magnitude; we refine it during the diagnostic.

• Company with 80-300 active SKUs in one or several markets (EU + UK + US, for example). Supplier or ingredient changes 2 to 6 times a year, multiplied by the number of retailers that demand their own artwork.
• Deployment of Connect + Writer/Tracer agents integrated with your ERP/MES and your label management system. First value expected within a few weeks on a pilot product family (sauces or cookies, for example).
• Expected reduction in mass regeneration time ≥30% (often much more, but that is the defensible floor). Fewer labeling errors from undetected SKUs: the hard lever is a single importer or retailer incident avoided.
• Indicative payback between 4 and 9 months, depending on SKU volume, frequency of changes and the average cost of reprints/claims in your chain.

And R&D's reasonable doubt

“What if the AI gets the nutrition calculation wrong?” — hallucination is a problem of free generation, not of anchored tasks. Recalculating a nutrition panel from a quantitative recipe is exactly an anchored task: the AI recontextualizes one piece of data (grams per ingredient) using another (unit nutritional value) and applies deterministic rules (rounding, tolerance). In this type of task, the best models brought error below 1.5% ^[1]. And even so, the person validates the list — without a signature, nothing goes to print. The three safety rings are there for this.

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