The shift lead dictates while walking the melt shop

The problem

The gemba dies at the end of the shift — and the next shift starts from zero.

In a long products steel mill, the shift lead is not sitting in front of a screen: they are walking the gemba — furnace, continuous casting, rolling mill, cooling bed — with their hands full, all shift. And so are the maintenance lead and the quality lead. They are the people who first see the patterns no sensor reports:

• A mill stand that starts vibrating differently — a change in behavior the shift lead notices long before the data triggers any formal alarm.
• A guide marking the bar since the campaign change — "since that heat, line 2 puts out more marked material" — which nobody notes down at the time because there is nowhere to type on the floor.
• The well-founded sense that the line is running worse this week — real operational knowledge, reconstructed days later from memory, if it gets reconstructed at all.

All of that lives in three people's heads and dies at the end of the shift. It does not enter the system, it is not cross-referenced with MES data and, above all, it does not feed the next shift's A3s, katas or PDCAs: every improvement cycle starts without that context, rebuilding the problem from scratch — and that is how you always end up fixing symptoms rather than causes.

How it fits the IRIS system

Connect in hands-free voice mode — the shift lead takes notes by dictating, without stopping.

The earpiece is not a system that records the shift lead throughout the shift: it is a tool they activate deliberately when they want to note something on their walk through the gemba. They decide when to speak and what to dictate. Connect structures it, ties it to the active plant context and returns it within seconds.

The shift lead says "iLEAN, take a note". They dictate the observation in their own words, without stopping. Connect structures it as an incident tied to the active asset, heat, line and shift. Seconds later the response arrives by voice through the earpiece, or as a push to the line manager's tablet.

How Connect works in earpiece mode in a long products steel mill:

• A personal wake word — the shift lead opens the dictation with their own phrase, one only they say deliberately. Outside that phrase, the earpiece transcribes nothing and sends nothing.
• Free dictation, no forms — they speak in their own words, exactly as they would tell the incoming shift: "stand 6 is vibrating more since the campaign change, have maintenance take a look before the next heat".
• The LLM structures the incident — Connect turns the dictation into an incident with clear fields, automatically tied to the asset, the heat, the line and the shift active at that moment, without the shift lead having to recite any of it.
• Response by voice or push — the confirmation arrives through the same earpiece or, if more context is needed, as a push alert to the line manager's tablet. The shift lead keeps walking, with the observation already in the system.
• A natural enabler of the improvement cycle — every dictated incident feeds the kata, the shift Kaizen and the digital A3 that nourishes the evidence pack: the next shift starts its PDCA with real context, not with a reconstruction.

See the full IRIS architecture →

Before and after

Knowledge that dies at the end of the shift vs. information that enters the system in real time

Aspect	Without an earpiece	With iLEAN Connect in earpiece mode
Observation spotted on the floor	Stays in the head until the end of the shift, if it makes it	Dictated in the moment, hands free, without stopping
Link to asset, heat, line and shift	Depends on somebody remembering and reconstructing it	Automatic — the system already knows what is active
Shift handover	Five minutes of hurried conversation beside the mill	The incoming shift inherits structured, dated incidents
Vibration in a mill stand	Spotted once the line has stopped or marked material	Noted and tied to the asset the first time somebody notices it
A3, kata and PDCA for the next shift	Start by rebuilding the problem from memory and fix symptoms	Start with real gemba context and attack causes
Cross-reference with MES data	Does not exist: human observation and data live apart	The incident is cross-referenced in real time with live MES data

Impact estimate

Impact estimate for your plant — to be validated 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.

• Long products steel mill with a shift lead, a maintenance lead and a quality lead who walk the gemba most of the day, hands full and with no practical access to a keyboard.
• Connect earpiece pilot on one line or one shift — without changing the shift lead's role and without adding administrative tasks.
• The main return is not an accountable saving in hours: it is that every A3, every kata and every PDCA start with real gemba context instead of memory-based reconstructions at the handover meeting.
• There is an expected operational return too: early detection of anomalous behavior in mill assets — the pattern the shift lead notices days before the line stops. That improvement is an estimate to be validated with the plant's continuous improvement lead, not a generic figure invented here.

And the fair question from the mill director

"Does this not end up as a system that records everything my shift lead says?" — no. It is exactly the reverse: the earpiece hears nothing until the shift lead activates it with their wake word. There is no background recording, no continuous transcription, nobody listening to their conversation with a mill operator. The only thing that enters the system is what they themselves choose to dictate — because they want it on record, not because somebody above asked for it. And the second question, "what if the AI makes half an incident up?": for anchored tasks like this one — structuring a dictation against the active asset, heat, line and shift — the AI operates far from the free-generation hallucination problem ^[1], and even so every incident is signed by whoever dictated it and is reviewable before it feeds any A3.

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