Validate before the data reaches the ERP

The problem

A misread heat number does not stop at the ERP: it ends up on the delivery note.

Without a human gate, any recognition error — a misread size or heat number — would flow straight into the ERP and from there to the invoice and delivery note, the worst place to discover it.

• Without a human gate, any recognition error — a size read as 1990 instead of 1900, a heat number with one character wrong — would flow straight into the ERP.
• And the ERP is not the end of the line. From there it travels to the delivery note and to the invoice, which in this trade is priced by the kilo, and onto the label that goes with the pallet.
• That is the worst possible place to discover it: in front of the customer, with the material already loaded and the paperwork signed.
• It is also the fear that stops the whole project at management level, and it is a reasonable fear: bad data that looks official is worse than no data at all.

How it fits the IRIS system

Connect with early human verification — an industrial tablet and two taps.

An industrial tablet at the cutting station or supervisor's post shows a clear visual summary of what was captured by photo, voice or email. The operator confirms with 2 taps (✓ / correct) and only then does the data cross into the central system.

Humans in command written into the architecture rather than promised in a slide: whatever was captured by photo, by voice or by email appears as a short summary, and it does not exist in the ERP until somebody has confirmed it. This is the piece that gets the other eleven approved.

See the full IRIS architecture →

Before and after

Today's data entry versus the validated entry

Aspect	Today	With iLEAN
Who writes to the ERP	A person, typing later	A person, confirming now
Time per order	Minutes at a terminal	Two taps at the cutting station
A misread size or heat number	Found on the invoice	Corrected before it enters
Where the summary comes from	—	The order photo, a dictation or an email
Who appears as the author	Whoever typed it	Whoever looked at the order
Can the AI write on its own	The fear that blocks everything	No, by architecture

Silent capture error reaching the invoice → error caught and corrected before invoicing or cutting.

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 it is honest to say so: it is not monetized on its own.
• What it buys is the reduction in returns and claims for a wrong size or a wrong heat, which the brief flags as the expected benefit and which your own returns log will price better than we can.
• And the approval of everything else: no sensible management team lets a system write the heat number that travels to an invoice with nobody in front of it.
• Plus a record of who confirmed what and when, which is exactly what gets asked for the day a claim lands.

Expected reduction in returns and claims from incorrect size or heat. Estimate to be validated.

And the fair question from the production manager

“If a person confirms in the end, what is the AI actually saving?” — the work is not in deciding, it is in assembling. Today somebody walks to a terminal, finds the order and types grade, thickness, format, quantity and heat. With the gate, the extraction — an anchored task where the best models drop below 1.5% error [1] — arrives already assembled and the operator only confirms or corrects. Every correction is stored, so the system learns where it fails on your own documents.

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