Incident capture — the one that dies in a chat message is the one that happens again.

The problem

The plant reports through every channel except the official one.

The official incident system in most plants is a screen in the MES or a module of the CMMS that people do not use. And it is not bad faith: asking someone to stop the line, walk to the computer, open the form, look up the machine code and fill in seven fields to report that "the filler is vibrating oddly" means that incident never gets reported. It gets told to the shift leader over chat, it ends up in an unread maintenance email, in a note in a notebook — and it shows up again a month later, the same or worse.

The result is paradoxical: the plant reports through every channel except the official one. The MES says there were 12 incidents last month; the shift leader, looking at their phone, knows there were 60. The gap between what happens and what gets recorded is the most expensive gap in the plant: without a record there is no 5 whys, without 5 whys there is no continuous improvement, and without continuous improvement the same incident comes back.

The classical system works 20 % of the time — the rest gets told however people can.

How it fits the IRIS system

iLEAN does not impose a new channel — it listens to the one people already use.

The problem with incident capture is not a lack of software: it is that the software forces the person to come to it, and the person does not come. iLEAN acts as the filler that listens across every channel where the plant already talks (voice, photo, email, chat, transcribed phone call) and turns it into usable data — without asking you to change MES, CMMS or Andon.

Connect listens through the channel the plant already uses. The agent categorizes and prioritizes. The Writer writes the order into the MES, CMMS or Andon with the evidence behind it. A person validates — the line does not restart on its own when the incident is critical.

The iLEAN pieces applied to incident capture:

• Connect — listens through the channel the plant already uses: the operator's full-duplex earpiece, a photo of the panel, the maintenance email, the shift leader's chat message, the corporate messaging group, a transcribed phone call. The capture setup adapts to each zone (in heavy pressing with extreme noise, photo capture; in packaging with moderate noise, voice).
• Agent — transcribes, reads, sees, categorizes against the plant master data (line, machine, code, severity), asks back through the same channel when information is missing, prioritizes using history and proposes the escalation. The literal quote from the message stays embedded in the incident, for audit.
• Writer — writes the downtime with code and duration to the MES, the work order with its evidence to the CMMS, and the supervisor alert to the Andon with the configured SLA. Write autonomy is configured plant by plant — from proposing a draft to opening an order with automatic sign-off for low-criticality incidents. Anything critical always goes through a person.

See the full IRIS architecture →

Before and after

Classical incident system vs. multichannel capture with iLEAN

Aspect	Classical system (MES/CMMS/Andon)	With iLEAN Connect + Agent + Writer
Capture channel	A screen on the shift leader's computer	Voice, photo, email, chat — the channel people already use
Operator effort	Stop + walk + open + fill in 7 fields	Dictate what they see through the earpiece, without stopping
Real capture rate	20 % of what actually happens	>90 % — because it is captured through the channel the person already uses
Categorization	Manual, depends on how tired the shift leader is	Automatic, with the literal quote from the message as evidence
Multilingual	Everything in one language, friction for subcontractors	Each person in their own language, reply in their own language
Traceability	An incident in the MES without the conversation that started it	Incident + literal quote + photo + similar history

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the actual data from your plant. We lay it out so the committee has an order of magnitude; we refine it during the diagnostic.

• A standard industrial plant with 3-6 lines, MES + CMMS + Andon already installed, and structural under-reporting (20-30 % of what happens reaches the official system).
• Connect pilot on one pilot line (earpieces + integration with MES/CMMS). First value expected within a few weeks — the rise in the incident capture rate shows up very fast.
• Indicative payback between 4 and 9 months, depending on the current cost of MTTR and the frequency of recurring incidents.
• Hard levers: MTTR reduction on the order of 30-50 % (prioritized action in seconds instead of minutes), 4-5× more incidents captured (which feeds the 5 whys far better), and recurrence reduction on the order of ≥30 %.

And the plant manager's fair objection

"What if the AI categorizes it wrong and opens an order for something it wasn't?" — every categorization carries the literal quote from the message that justifies it. The shift leader validates it with one tap (the quote is right there, not hidden in a black box). On anchored tasks such as reading a message and cross-referencing it with the plant master data, the best models brought the error rate below 1.5 % ^[1]. And the system lets you configure autonomy by incident type: low-criticality ones can be opened with automatic sign-off; critical ones always go through a person. That is what the three safety rings are for. The over-validation that kills the system (everything goes through a human and nobody signs in time) is the antipattern iLEAN avoids by design — autonomy graduated by criticality, not uniform.

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