Two taps before any data reaches the ERP

The problem

Nobody wants AI writing a dangerous-goods rating into the ERP unchecked.

The fear that AI "enters data on its own and someone has to clean up the ERP master later" is the main cultural brake on digitalization at a drum factory — especially when what's at stake is a dangerous-goods rating. But the operator doesn't want to type either.

• The fear that AI "enters data on its own and someone has to clean up the ERP master later" is the main cultural brake on digitalization at a drum factory.
• It is a reasonable fear when what is at stake is a dangerous-goods rating: a wrong packing group in the master travels to labels, certificates and the customer. Cleaning it up afterwards is slow and leaves doubts about every record around it.
• But the operator does not want to type either. He is there to assemble drums, not to fill in screens. Every form added at the line is a form filled in late or not at all.
• So plants stay stuck between unvalidated data contaminating the ERP and data that is never captured out of distrust. Neither option supports a rating that customers audit.

How it fits the IRIS system

Connect with human verification at the line — two taps, no form.

An industrial tablet at the assembly line shows a 2-second visual summary of any recent capture — photo, voice or parsed email. The operator confirms with two taps: ✓ goes to the ERP, ✏️ fix before it enters. No training, no change of habit.

This is the piece that makes the others approvable. Without it, nobody signs off on photos, dictations or emails feeding the system a UN rating depends on; with it, every record has a person who saw it before it entered. That is what the plant's quality system needs before it lets any automation near the rating.

See the full IRIS architecture →

Before and after

Data entry today versus data confirmed at the line

Aspect	Today	With iLEAN
Who sees a capture before the ERP	Nobody, or nobody captures it	The operator at the line
Time to confirm	Minutes at a terminal	A 2-second summary, two taps
A wrong packing group	Found when the label is printed	Corrected before it enters
Photos, voice and emails	Three separate worlds	One confirmation screen
Corrections made by people	Lost	Stored to improve extraction
Learning curve	A new screen to master	None: confirm or fix

Before: unvalidated data contaminating the ERP, or data not captured at all out of distrust. After: data captured and validated in seconds, the master always clean.

Impact estimate

An enabling piece — no standalone payback, by design.

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: it is the enabling piece of the whole iLEAN matrix, and its value is strategic, not directly monetizable. We say so plainly instead of inventing a figure for it.
• Without it, the other pieces don't get signed off by quality or by management. The return of the matrix runs through this tablet even though none of it is booked here.
• From unvalidated data contaminating the ERP — or no data at all — to data captured and validated in seconds.
• And the master stays clean, which is what every certificate and every rating record downstream depends on. A clean master is what lets the dossier for a UN-rated batch be trusted.

The enabling piece of the whole iLEAN matrix — without it, the other pieces don't get signed off. Strategic value, not directly monetizable.

And the fair question from the production manager

“If a person confirms everything, what does the AI save?” — the assembly, not the decision. Extracting batch, format and rating from a photo or an email is an anchored task, where the best models drop below 1.5% error [1]; the operator only checks a summary that arrives already built. Two taps replace a form, and each correction is kept so the system learns where it fails. Over time, the operator corrects less and confirms more. That is the sign the system is learning your line.

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