Cocoa and chocolate traceability with AI — from the sack to the bar, without fighting for it by hand.

The problem

Cocoa comes in through twenty doors. Chocolate goes out through one.

A mid-sized chocolate factory can take in cocoa from eight to fifteen different origins over the course of a year, with different harvests, different certifications (UTZ, Rainforest, organic, fair trade) and different delivery-note formats. Each one comes in through a different door:

1. The physical sack — with the exporter's label, sometimes illegible, sometimes carrying nothing but a code only the buyer understands.
2. The delivery note — on paper from the carrier, as a PDF emailed by the importer, or by EDI if the supplier is a large one. Three routes for the same piece of data.
3. The certificate — UTZ/Rainforest/organic in another email, on another day, sometimes weeks after the sack itself.
4. EUDR — plot geolocation, supplier due diligence, signed declaration. One more folder.

After that, the cocoa is blended in hoppers, roasted, winnowed, conched with butter and sugar from other batches, molded. Bean-to-bar traceability has to survive that chain, and it is nearly always held together with spreadsheets, the master chocolatier's notebook and the warehouse lead's memory. It works 95% of the time. The other 5% is an EUDR complaint, a retailer audit or a batch rejected at customs.

How it fits the IRIS system

iLEAN does not add another spreadsheet — it seals the cracks between the islands you already have.

The problem with cocoa traceability is not a lack of information: it is heterogeneous information arriving through different channels, which nobody has time to put into a common shape before the physical batch moves on. iLEAN acts as the putty that fills the gaps between supplier, warehouse, ERP and plant, without asking you to change anything.

Connect captures the delivery note however it arrives. Edge reads the physical sack. The agent maintains the batch genealogy right through to the bar. The person signs — the person always signs.

The three iLEAN pieces applied to bean-to-bar traceability:

• Connect — outward listening: the importer's email with the certificate, the supplier's WhatsApp with the plot geolocation, EDI where it exists. And inward listening: the receiving operator dictates the weight through the earpiece, the warehouse operator photographs the paper delivery note, the R&D manager says out loud “batch 2401 is 60% Ghana + 40% Ecuador”. It all enters at second zero.
• Edge — a vision + OCR terminal at receiving that reads the sack label and the printed delivery note. And on the line, above the roaster and the molding machine, it maintains the physical batch-to-product assignment at every step. It works with no network.
• Agent — cross-checks delivery note + sack + certificate + EUDR + work order and keeps the per-batch traceability dossier live and queryable. When the complaint or the auditor arrives, there is nothing to rebuild: the file already exists, with a photo of the delivery note, geolocation and the signature of whoever gave the OK.

See the full IRIS architecture →

Before and after

Manual traceability vs. cross-checked traceability with iLEAN

Aspect	Spreadsheet + notebook + email	With iLEAN Connect + Edge + Agent
Delivery-note capture	Paper into a folder, PDF into an inbox	Edge OCR + Connect email/EDI, at second zero
Sack identification	Label read by eye by the warehouse lead	Vision + reading, automatic assignment to the batch
Blend → bar genealogy	Rebuilt after the fact if a complaint arrives	Maintained live at every step of the process
Per-batch EUDR file	A digital folder full of mismatched files	Structured dossier: photo + GPS + signatures
Customer complaint	Weeks to answer, risk of finding nothing	Minutes, with visual evidence of the batch
IFS/BRC/retailer audit	One person tied up for two weeks	Export the file straight from the system

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.

• Factory with 6-12 active origins, 2-3 lines (roasting, conching, molding), annual volume between 500 and 5,000 t of cocoa processed.
• Connect + agent pilot at receiving and roasting: delivery-note and certificate capture, per-batch dossier, integration with the scale and the ERP. First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on how often EUDR or retailer complaints come in and on the time spent today on audits and manual reconstruction.
• The hard lever is one single consignment not rejected at customs or returned by a retailer, avoided because the file was ready. A single batch pays for the pilot.

And the quality manager's reasonable doubt

“What if the agent pulls a value off the delivery note wrong?” — hallucination is a problem of free generation, not of anchored tasks. When the AI simply recontextualizes a piece of data from a document into a system (it reads a weight, it deposits it in the right batch), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: the dossier proposes, 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.