Paperless batch startup

The problem

The two sheets that define the batch spend the shift on a clipboard.

the order is printed from ERP data and any last-minute adjustment gets handwritten. The formula sheet for customer-specific blends can carry variations not always logged. Both can sit undigitized for hours, recovered only when someone transcribes them at shift end.

• The production order prints from the ERP, and then reality corrects it by hand: a sack of sweet paprika swapped for another origin lot, a percentage adjusted because the incoming color came in low, a component substituted the morning it ran short.
• The formula sheet for customer-specific blends carries versions that are not always logged. Two blends with the same commercial name can differ in the allergen they declare, and the difference lives in pen on paper.
• Between the mixer starting and somebody transcribing the paperwork, hours pass — and over a weekend, days. Nobody designed it that way; it is simply that transcription is the last thing anyone does and the first thing that gets dropped when the plant is busy.
• During that gap the batch physically exists, is packed and may even ship, while existing in no system anyone can query. Which is exactly when a customer calls asking what went into it, or when quality needs to hold the batch before it leaves.

How it fits the IRIS system

Connect in photo mode — the operator keeps filling in the same paperwork.

Connect photo mode. A grounded LLM extracts batch, formula, percentages and allergens straight from the photo. The data enters central memory without the operator changing their routine.

It is the shortest of the twelve to deploy and the one that creates least friction at the mixer, because it asks for no change of habit: the operator adds a photo to what they already did. Nothing is taken away, no new screen is introduced and no step is added to the changeover. That is usually why it is the first one approved, and why it is the one that survives the third month.

See the full IRIS architecture →

Before and after

Batch startup on paper versus batch startup captured

Aspect	Today, on paper	With iLEAN Connect
Capture latency	Hours, and days over a weekend	Zero lag
Formula version used	Whatever was on the sheet	Recorded with the batch
Declared allergens	In pen, in the margin	Structured field, tied to the batch
Origin lots consumed	Reconstructed afterwards	Linked at startup
Answering a customer query	Hunt through the clipboard	One lookup by batch
Operator's routine	—	Unchanged: two photos

hours of latency until transcription → zero lag. Scattered paper → instant web lookup by batch and formula.

Impact estimate

Estimated impact — to validate with your own 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 how many batches you start per day.
• Transcription time recovered by production and by the quality team, which is where it is scarcest and where it is currently spent on typing rather than on control.
• From hours of latency — days over a weekend — down to zero lag, on every batch and not only on the ones somebody remembers.
• And the formula version actually blended stops being a memory exercise when a customer complaint lands weeks later.

estimated payback of 4 to 9 months depending on daily batch volume. *Estimate to validate*.

And the fair question from the production manager

"What if it misreads a percentage in the formula?" — reading fields from a document with a known structure is an anchored task, where the best models drop below 1.5% error [1]. And there is a second net: the summary goes to the tablet next to the mixer and the blending manager confirms or corrects it before anything crosses into the central system.

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