The start of production stops living on three sheets of paper

The problem

The batch is defined by three sheets of paper that nobody digitizes until the evening.

The work order is printed from the ERP in the morning, but what actually happens on the floor gets written on top of it by hand: a supplier swapped at the last minute, the milk or butter lot that was really opened, the note that the béchamel came out thicker than usual. That handwriting is the only faithful record of what happened, and also the most fragile one. It sits in a folder until somebody collects it at the end of the day. When an audit or a distributor complaint arrives, the shift has to be reconstructed one sheet at a time. The cost is not only the production manager's transcription time. It is that for hours the central system and the plant floor tell two different stories, and the true one is written in ballpoint pen.

• The work order prints from the ERP at seven in the morning. By eight it already carries handwriting: the butter lot that was really opened, the fish supplier swapped at the last minute, the note that this béchamel came out thicker than the recipe says.
• The recipe sheet and the start-up log are where the supervisor records what the kitchen actually did: which allergens went into the kettle, which forming machine was used, at what time the first tray reached the tunnel.
• Those three papers sit in a folder for four to eight hours, and over a weekend for two days. During that gap the batch exists physically and exists in no system anyone can query.
• When a retail complaint or an auditor asks what went into a batch, the shift is reconstructed one sheet at a time, and the only faithful version is in ballpoint pen.

How it fits the IRIS system

Connect in photo mode — the kitchen keeps its papers and adds a photo.

Connect in "photo of the analog" mode.

It is the case with the least friction in the whole matrix, because it asks the kitchen to change nothing: the supervisor photographs what they already filled in. That is why it is usually the first one approved, and why it becomes the base that the allergen and labeling rings later rest on.

• The supervisor photographs the work order, the recipe sheet and the start-up log with a phone, exactly as they already do with everything else.
• A grounded language model extracts structured fields: reference, batch, line, recipe, raw material lots consumed, declared allergens and observations.
• The result passes through the early human validation screen (case 5) and only then enters the master.
• The batch sits in central memory with the photo of the original paper attached as evidence. No change of habit and no training.

See the full IRIS architecture →

Before and after

Today's batch start versus the captured batch start

Aspect	Today, on paper	With iLEAN Connect
When the batch exists in the system	4-8 h later, more over a weekend	Second zero
Raw material lots actually opened	Handwritten on the order	Extracted and tied to the batch
Allergens declared for the recipe	On the recipe sheet, in a folder	A structured field per batch
Answering a distributor complaint	Reconstruct the shift sheet by sheet	One query by batch
Evidence behind the record	The paper, if it is found	The photo of the original paper
The supervisor's habit	—	Unchanged: two or three photos

Impact estimate

Impact estimate — to be validated with your numbers.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 4-9 months, depending on batches and references per day.
• The production manager's transcription time recovered, which today goes into copying the day's orders into the ERP after the kitchen has closed.
• From four to eight hours of documentary lag down to second zero, batch after batch, including weekends.
• And the real milk, butter, egg and fish lots stop being a handwritten note and become the field that every later traceability query starts from.

estimated payback of 4 to 9 months depending on batches and references per day, plus the production manager's recovered transcription time. *Estimate to be validated*.

And the fair question from the production manager

“What if it misreads the lot number the supervisor wrote by hand?” — reading fields from a document with a known structure, even with handwriting on top, is an anchored task, where the best models drop below 1.5% error [1]. And there is a second net: the extraction goes to the line-side tablet and the supervisor confirms or corrects it before anything crosses into the master. A low-confidence lot is flagged, never assumed.

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