Humans in command: the blend is validated before the ERP

The problem

A wrong blend in the ERP ends up on the customer's spec sheet.

If the blend or batch destination enter the ERP wrong, they contaminate the spec sheet sent to the bakery customer -a commercial risk, not just a quality one.

• The ERP feeds the specification sheet sent with every flour delivery: protein, moisture, W, P/L, falling number and destination. Industrial customers use that sheet to adjust their own process to each delivery.
• If the blend or the batch destination enter wrong, the error travels to the customer on an official document. And once a spec sheet has been sent, correcting it is far more awkward than getting it right.
• That is a commercial risk, not just a quality one: an industrial customer who receives a sheet that does not match the flour stops trusting all the others.
• And AI capturing data straight into the ERP with no control would be the fastest way to multiply that risk. Quality managers know this, which is why they block any project that does not answer it.

How it fits the IRIS system

Connect with human validation — two taps on an industrial tablet.

An industrial tablet with a clear visual summary of what was captured (photo, voice, email). The mill supervisor confirms or corrects with two taps before the data crosses into the central system.

This is the piece that makes the other eleven approvable. Nobody in a mill signs off on an AI writing directly into the system that issues the customer's spec sheets; everybody signs off on an AI that prepares and a person who confirms. That is why we deploy it with the first capture case, not as an afterthought.

See the full IRIS architecture →

Before and after

Data entry today versus the validated gate

Aspect	Today	With iLEAN
Who enters the blend	Whoever types it later	The supervisor who saw it, confirming
What the supervisor reviews	Nothing, or a printout	A visual summary: photo, voice, email
Time per validation	Walk to a terminal	Seconds, two taps
Corrections	Overwritten, no trace	Stored with author and time
Spec sheet sent to the customer	Fed by unchecked data	Fed by confirmed data
AI writing to the ERP	—	Never without a person

From AI-captured data entering directly with no control, to a seconds-long human gate that adds no friction to the operation.

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.

• No standalone payback — this is an enabling piece, and we do not put a number on it on purpose.
• It adds no implementation cost: the tablet and the gate come with the capture cases it protects. It is part of the architecture, not an extra line on the quote.
• Its value is risk removed: the quality history and the customer's spec sheets stop being contaminated with unvalidated data. Every record that reaches the ERP has a name and a time behind it, which is what an auditor or a customer will ask for.
• And it is what allows the photo, voice and email cases to be approved by quality management.

No added implementation cost; removes the risk of contaminating the quality history with unvalidated data. Estimate to validate.

And the fair question from the production manager

"If a person confirms everything anyway, what is the AI saving?" — the work is not in deciding, it is in assembling. Extracting the blend from a photo or an alert from an email is an anchored task, where the best models drop below 1.5% error [1], so the supervisor receives it already assembled and only confirms or corrects. Every correction is stored: the system learns exactly where its extraction fails. After a few weeks, the corrections themselves show which fields of which sheets need attention.

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