Nothing enters the ERP without a human signing it
Nothing iLEAN captures enters the MES or the ERP without a human seeing it first. In a long steel mill that becomes an industrial tablet at every critical line: a two-tap digital andon — check to validate, pencil to correct — between what the AI captures and the central system master.
The fear is not AI — it is AI writing data on its own and contaminating the ERP master.
In a long steel mill, the MES and ERP master holds the traceability of every heat, every rolled lot and every bundle that leaves on a shipment. Any figure that enters wrong — a misidentified heat, a mistranscribed test, a badly recorded specification change — propagates downstream through the mill until somebody catches it. That is why the cultural blocker here is twofold:
- The fear of AI writing data on its own — it is the biggest cultural blocker to digitizing the plant. If a system writes into the master by itself and a figure later turns out to be wrong, the cost is not correcting a field: it is reopening the traceability of everything that depended on it. It is the committee's first question: who has seen this before it existed in the ERP?
- But the generational handover runs in the opposite direction from paper — the operator arriving on the line is not going to fill in paper forms the way the veteran they replace did. If the alternative to AI is transcribing by hand, the data arrives late, incomplete or not at all.
Full automation does not pass the plant's cultural filter. Going back to paper is no longer an option. The answer is not choosing an extreme — it is giving the operator a fast, human gate between the AI and the central system.
Connect in early-human-verification mode — the person before the master, not after.
Everything iLEAN does rests on three safety rings: Connect carries the data as it arrives (photo, voice, parsed email, panel reading), Agents interprets it and decides what action to propose, and the person validates before it crosses into the MES or the ERP. That separation is the backbone of the whole iLEAN long steel matrix — and this piece is where the third ring becomes visible and tangible: an industrial tablet at every critical line, with a two-tap gesture on every capture.
An industrial tablet at every critical line. A two-second visual summary. Two taps — check to validate, pencil to correct — before the data crosses into the MES or the ERP.
How early human verification works in a long steel mill:
- An industrial tablet at every critical line — a UI designed for a two-second validation, not for data entry. There is no keyboard and no nested menus: there is a visual summary and a two-tap decision.
- A visual summary (heat, lot, event) — whether the data came from a photo, from voice, from a parsed email or from a panel reading, Connect normalizes it into the same simple format before showing it to the operator.
- Check to validate, pencil to correct — if the capture came through clean, one tap and they carry on; if something does not add up, the correction is minimal and the original is filed alongside it.
- Latency between capture and a clean master: seconds — validated data reaches the MES or the ERP almost at the same time it was generated, not at the end of the shift.
- No validation, no master — if the operator does not validate, the data is held and contaminates nothing downstream. The line carries on; the data waits.
- Human jidoka with data support — the operator keeps their stop authority intact and exercises it on real information from their own line, not on intuition or on what gets remembered at shift end.
Unvalidated data (or data never captured out of mistrust) vs. data validated in seconds
| Aspect | Without early human verification | With iLEAN Connect + tablet |
|---|---|---|
| Capture by photo/voice/email/panel | Discarded out of mistrust, or enters straight through unreviewed | Always passes through a visual summary and two taps before touching the master |
| Operator recording | Paper forms the new generation no longer fills in | Check or pencil on the tablet, no typing |
| Wrong or doubtful data | Contaminates the MES/ERP master and drags traceability downstream | Held — it does not cross into the master until somebody validates it |
| Time between capture and a clean master | Hours or shifts, if it ever gets recorded | Seconds |
| Stop authority (andon) | Exercised on intuition, or late, with the defect already downstream | Exercised on real line information, in the moment — jidoka with data |
| The plant's stance on AI | Blocked — "we do not know what is going in by itself" | A clean master: every figure has passed through a human before existing |
The value of this piece — enabling, not a payback in months.
This case is deliberately not presented with a payback figure. Its value is structural: it is what lets everything else be signed off.
- It is the cross-cutting enabling piece: every Connect, Edge or Agents case in your long steel mill depends on this ring existing so the data they capture can enter the MES or the ERP without fear.
- It does not replace work — it replaces doubt. The cost it avoids is not an hour of operator time: it is the paralysis of "we are not going to let an AI write into our master on its own".
- It is also the piece of cultural change the generational handover calls for: the new operator does not inherit paper forms, they inherit a two-tap tablet — and the veteran keeps what was always theirs, the stop authority.
- The indicators that can be measured from the first pilot — as an estimate to be validated in each plant: the percentage of captures validated in under five seconds, and the percentage of held data that turned out to be real corrections — the latter measures whether capture is well calibrated, not whether "the AI fails".
- We do not put forward a payback in months here — that would be inventing a number only your plant's assessment can give. What we can state is that without this piece, the rest of the IRIS roadmap in your mill stays a pilot and never scales.
And the fair question from the committee
"Why not automate fully, if we already have AI that reads well?" — because reading well is not the same as deciding well, and deciding well is not the same as being authorized to write into the MES or ERP master with nobody reviewing it. iLEAN separates the three functions deliberately: Connect carries the data as it arrives, Agents interprets it and proposes the action, and the person validates before it crosses into the master. Those are the three safety rings, and this piece is where the third becomes physical: a tablet, a two-second summary, a check or a pencil. Hallucination is a problem of free generation, not of anchored tasks — in tasks where the AI merely recontextualizes a figure (reading a panel, transcribing a label, parsing an email), the best models brought the error below 1.5% [1]. And even so, on a long steel line what is critical is not recorded on its own: that is why the tablet exists — and why stop authority still belongs to the operator.
[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.
What people ask about the digital andon with human verification
Why not automate fully if there is already AI capable of reading the data well?
"Reading well" is not the same as "being allowed to write into the MES or ERP master with nobody reviewing it". iLEAN separates Connect (carries the data as it arrives), Agents (interprets and decides what action to propose) and the person (validates in two taps before the data crosses into the central system) — those are the system's three safety rings, and none replaces another. Automating fully sounds faster, but it turns a one-off error into a systemic one: in a rolling mill, a badly recorded heat or lot propagates downstream — bundle, labeling, shipment — until somebody catches it, and by then correcting it is no longer editing a field, it is redoing traceability. And there is a deeper Lean reason: jidoka means the machine detects and the person decides. Taking the person out of the decision is not more Lean — it is less. The tablet keeps that principle, with data: the AI proposes, the operator disposes.
How fast are the "two taps" really — or is it another form in disguise?
Two taps means exactly that: a visual summary — heat, lot, event — and two possible gestures, check to validate or pencil to correct. The UI is designed for a two-second validation, not for data entry: there is no keyboard in the normal flow, no nested menus, no mandatory fields to fill. If the capture came through clean — which is the usual case — the operator looks, taps the check and carries on with their line. Only when something does not add up does the pencil come in, opening a minimal correction on the specific field rather than a whole form. And it is measurable from the first pilot: the percentage of validations resolved in under five seconds is one of the startup indicators — as an estimate to be validated in each plant. If the operator takes longer, the flaw is in the capture design, not in the operator, and what gets fixed is the capture.
What happens when the operator corrects a figure — is the AI's original capture lost?
No. When the operator taps the pencil and corrects, the original figure Connect captured is filed alongside the correction, not overwritten — the MES or ERP receives the validated figure and the history keeps both, with who validated, when and what changed. This matters for two reasons. First, long steel traceability — from the heat to the shipped bundle — requires being able to demonstrate what happened, not only what was corrected. Second, it makes it possible to detect with evidence whether capture fails repeatedly at one specific point — a bundle tag covered in scale, a mill panel hard to read through the heat — and adjust the capture, rather than arguing over impressions. The operator's correction is not a patch: it is the signal that calibrates the system.
Does it work for every capture source (photo, voice, email, panel) on the same tablet?
Yes — the tablet is not specific to one channel: it is the point of convergence where the visual summary arrives, whether the data came from a photo (a bundle tag, a billet delivery note), from voice (the shift lead dictating an incident at the cooling bed), from a parsed email (a specification change sent by the customer), from a panel reading (the reheating furnace pyrometer, the shear counter) or from a spreadsheet of mechanical tests. Connect normalizes every source into the same simple summary — heat, lot, event — so the operator does not have to learn a different interface depending on where the data came from. One two-tap gesture, whatever the capture channel.
How does this relate to the operator's stop authority (jidoka)?
It reinforces it — it does not replace it. In the classic andon, the operator pulls the cord when they detect an anomaly, and that stop authority is theirs, not the machine's. iLEAN does not touch that hierarchy: the tablet never stops the line on its own, and it never writes into the central system on its own. What changes is the quality of the information on which the operator exercises their authority: instead of deciding on intuition or waiting until shift end, they see within seconds what has just been captured on their line — heat, lot, event — and decide with that in front of them. If something does not add up, they hold the data (which does not cross into the master) and stop the line if their judgment calls for it: it is human jidoka with data support, not with intuition. The digital andon does not transfer stop authority to the AI — it gives the operator better reasons to use it in time.
More cases from this series
- Every heat ships with its audit dossier already madeEvery heat ships with its auditable dossier: charge mix, chemistry, rolling record and certificates,…
- The coil marker stitched to the active order, with no retypingHeat number, grade and customer travel from the active order to the coil marker with no retyping: the…
- The customer alert, read within seconds instead of hoursOrder changes, quality alerts and scrap supply delays are read, classified and cross-checked against the…
- The electric arc furnace panel already talks — without touching its PLCA camera reads the electric arc furnace panel and every heat carries its real energy and power-on curve…
- Heat startup in the melt shop, digitized with a photographThe heat startup sheet is photographed and the heat exists in the central system at zero latency: scrap…
- The lab spreadsheet poured into the ERP, without retypingChemistry and mechanical readings move from the lab spreadsheet into the ERP with no retyping — and…
Two operator taps, a clean master every time — tell us your case and we will show you the tablet running in a plant like yours.
We work on your plant's real data, not ours. Assessment with no commitment.
See the tablet at the line ‹ See all 12 long steel cases See steel