AI against premium perfume counterfeiting — the bottle that leaves the plant twice should never leave at all.

The problem

Counterfeiting is visible; the gray market is not. And both do the same damage.

Counterfeiting in premium perfumery is the visible version of the problem: a cheap copy that cannibalizes sales, damages the brand image and, when the customer realizes what happened, pins the bad experience (a bad scent, a skin reaction) on the real brand. The gray market is the invisible version: the very same genuine product — the legitimate unit that left your plant — diverted to an unauthorized channel in another geography, sold at a discount and undermining the boutique network that sustains premium positioning.

Both sides are fought with the same tool: uniqueness per unit. Not per SKU. Not per batch. Per unit.

1. Without individual serialization — the bottle is indistinguishable from its copy, and the genuine bottle that was diverted is indistinguishable from the one sold through the right channel.
2. With individual GS1 DataMatrix serialization — every bottle carries its own license plate. The copy fails verification. The diverted bottle is traced back to the order and the distributor.
3. The real bottleneck — it is not deciding to serialize; it is verifying that the print came out right on every unit, on every shift, at every SKU changeover, without stopping the line.

The classic setup (a standard DataMatrix reader + a checklist) works 99% of the time. That 1% — a bottle that ships with an illegible code, two bottles carrying the same code because of a printing fault — is what opens the door to the gray market and, worse, leaves the team without the trail to close it.

How it fits the IRIS system

iLEAN Tracer is not one more reader — it is Edge with vision + Agents that close the loop.

The problem with serialization is not a lack of printing technology: that has been around for decades. The problem is verifying and recording every reading in real time, cross-referencing it with the order and the recipient, and reacting before the case is closed. There is an island at each of those steps. iLEAN acts as the putty that fills the gaps between the printing system, the ERP, the WMS and the distribution network.

Edge verifies the DataMatrix in milliseconds. Connect cross-references it with the order and the destination. The agent closes the file unit by unit — and the person signs every ejection.

The iLEAN pieces applied to premium perfume anti-counterfeiting:

• iLEAN Tracer (Edge + vision) — a terminal with convolutional neural networks (CNN) over the labeling line. It reads every DataMatrix right after printing and checks legibility, contrast, distortion and uniqueness in the database. If the code is not legible or already exists, it fires the actuator (ejector) in milliseconds. It works without a network — if the plant loses WiFi, Edge keeps verifying locally.
• Connect — captures the serialized unit and ties it to the order, the recipient and the authorized channel. If the bottle later turns up in an unauthorized channel, the data is there. It also captures what comes in from outside (the distributor's WhatsApp with an order change, the email from the anti-gray-market lead with an alert) at second zero.
• Agents — cross-reference the readings with the order, with the WMS and with the alerts from the anti-counterfeiting team. If a suspicious pattern appears (a code read twice, a unit expected in a Paris boutique that is reported in Dubai), the agent prepares the file so the legal team can act. It never decides alone.
• Three safety rings — the database of serialized units lives in ring 1, isolated. Whatever comes from outside (readings from points of sale, external alerts) is validated in ring 2 before it touches ring 1. For IT/CAIO: it fits the AI Act and regulatory traceability without giving up the power of AI on the outside.

See the full IRIS architecture →

Before and after

Classic DataMatrix reader vs. serialization with iLEAN Tracer

Aspect	Standard reader + checklist	With iLEAN Tracer (Edge + Connect + Agents)
Reading on black / gold / metallized glass	Failures from reflections and bottle shift	CNN trained with real samples from your line
Uniqueness checked on the spot	Later audit, batch by batch	At second zero, per unit, before the case is closed
Cross-reference with order and recipient	Manual / after the fact, in another system	Unified at second zero
Gray market detection	Through complaints or external audit	Pattern cross-referenced by an agent — trail back to the order
Operation without network	n/a	Edge keeps verifying on panel power
File for the legal team	Rebuilt by hand, weeks	Dossier per suspect unit, 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 and your distribution network. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Premium perfume packing line, 1-2 SKUs per shift, complex glass (black/gold/metallized), international distribution network with a presence on marketplaces.
• iLEAN Tracer pilot on one line (camera over the DataMatrix + actuator + integration with the printing system and the ERP). First value expected within a few weeks: the first run with 100% per-unit verification instead of sampling.
• Indicative payback between 4 and 9 months, depending on the volume of gray market estimated in your network and the cost per counterfeit unit detected in e-commerce.
• Hard lever: closing a single gray-market channel pays for the pilot. The drop in copies on marketplaces starts to show as soon as the network begins verifying the DataMatrix at destination.
• Expected reduction of units with an illegible code of ≥ 30% against your current baseline, and per-unit traceability from the very first shift.

And the operations director's reasonable doubt

“What if the AI mistakes a good code for a bad one and we eject genuine product?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely reads a matrix of dots and compares it against the database of valid codes, the best models brought the error below 1.5% ^[1]. And even so, what is critical is never decided alone: units with a doubtful reading go to a human validation screen before being ejected. The three safety rings exist precisely for this.

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