WhatsApp HACCP in an artisan PDO cheese dairy — the master reports as always; the system records underneath.

The problem

The master's island of knowledge — and a notebook the Regulatory Council would rather not see this empty.

In an artisan PDO cheese dairy two uncomfortable truths live side by side. The first: the master knows. He knows at what hour the whey asks for a lower temperature, in which vat a cold morning alters the curd, which batch will yield less because of the milk's initial pH. The second: that knowledge almost never reaches the HACCP notebook in time. The master posts to the workshop WhatsApp group — "vat 2 has drifted half a degree, I'm giving it three more minutes" — and carries on working.

The day the Regulatory Council auditor arrives:

1. The HACCP notebook is incomplete — events that did happen and were handled well are missing, simply because nobody wrote them down.
2. Rebuilding the traceability of the PDO batch means diving into the chat, screenshots and the shift's memory.
3. And the underlying question the auditor does not ask out loud but is thinking: "if the master retires tomorrow, what is left of all this?"

This is not clumsiness on the master's part: asking for paperwork from someone with his hands in the whey is asking him to do his craft worse. The answer is not to change the master — it is to change the system.

How it fits into the IRIS system

iLEAN does not add another digital notebook — it seals the crack between the master's WhatsApp and the PDO specification.

Connect is the ear and voice of the plant: it captures whatever arrives through the operator's natural channel and puts it into the system at second zero. It is the filler between the master's know-how and the file the Regulatory Council asks for. It demolishes nothing that already works — it seals the cracks.

The master reports by WhatsApp as always. Connect links it to the PDO batch. The evidence builds itself. The person signs — the batch is not released on its own.

The iLEAN pieces applied to HACCP in a PDO cheese dairy:

• Connect — listens to the workshop WhatsApp group (text and voice), transcribes the alert, keeps the photo if there was one. The capture grade is decided dairy by dairy: the master's phone, a full-duplex earpiece in the room, or an Edge terminal if the area has no coverage. It demands no change of habit.
• Agents — cross-check the time of the message, the vat or room mentioned and the production schedule to link the alert to the right PDO batch. When in doubt, they do not invent: they ask the master through that same WhatsApp with two one-tap options.
• Agents + Central (batch file) — prepare the batch file from second zero: literal transcript, time, photo, PDO batch, decision taken by the master. The Regulatory Council auditor receives a coherent dossier; the person validates and signs its closure. The master's knowledge stops evaporating — it becomes permanent capability of the dairy.

See the full IRIS architecture →

Before and after

Loose notebook vs. a linked PDO batch file with iLEAN

Aspect	Notebook + loose WhatsApp	With iLEAN Connect + Agents
The master's channel	Workshop WhatsApp group	Workshop WhatsApp group (the usual one)
Record of the HACCP deviation	Written down later, if remembered	Automatic, with literal transcript
Link to the PDO batch	By hand, at end of day	Automatic, confirmation if uncertain
Photo/voice of the moment	Lost in the chat	Attached to the batch file
The master's knowledge	Lives only in his head	Captured as a reusable pattern
Regulatory Council audit	Rebuilt by hand, weeks	Dossier per PDO batch, 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 concrete data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• PDO cheese dairy with 2-3 masters, several vats, HACCP records today in a notebook and loose workshop WhatsApp messages, annual Regulatory Council audit.
• Connect pilot (reading the WhatsApp group + an Agent linking to the PDO batch + batch file generation). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on hours per shift spent on records and the cost of an incident that reaches the specification.
• Reduction of batch file closing time ≥ 30%. The underlying lever is not efficiency: it is protecting the ability to pass the audit without surprises.

And the quality manager's fair objection

"What if the AI transcribes it wrong and leaves faulty evidence in the file?" — hallucination is a problem of free generation, not of anchored tasks. Transcribing an audio note or linking a message to a PDO batch in progress is an anchored task; the best models brought the error rate below 1.5%^[1] in this kind of task. And even so, nothing critical is decided alone: the agent proposes the uncertain match to the master and the person signs the batch closure.

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