Muda, mura and muri detection with AI — the 3M before they cost money, not after.

The problem

The 3M are not visible — until they cost money.

The continuous improvement manager knows muda, mura and muri are there. They can smell it on every gemba walk. But quantifying them by hand is impossible: measuring the mura of station 4 across a shift would take a technician with a stopwatch for the whole day, and again the next day. Detecting muri in an operator before they burn out would take continuous observation. Mapping muda in material flow would mean a fresh spaghetti diagram every week.

So the manager works with last month's aggregate OEE and the eye of the shift leaders. OEE says the line is at 72%. It does not say why, at which station, at what point in the shift, with which SKU, or who is absorbing the overload so the number doesn't drop further. By the time a 3M problem explodes — an accident, a systemic defect, a veteran who resigns — the cost is already paid, and reconstructing it after the fact takes weeks. Invisible 3M is expensive 3M.

How it fits the IRIS system

iLEAN turns 3M detection into a by-product of continuous capture.

The 3M are reality-capture problems: if you measured every cycle of every station continuously and at no cost, mura, muri and muda would surface on their own. iLEAN was designed for exactly that — the R in IRIS is the capture of reality from the floor upwards. Without Edge, without Connect and without agents cross-referencing them, the 3M are still only visible after the fact. With the three pieces working together, they are a natural by-product of the system. It is the filler that closes the cracks between ERP/MES/SCADA and the real plant floor.

Edge measures cycle time continuously. Connect captures the rhythm of the chain (shift changes, material shortages, customer messages). The agent cross-references and attributes root cause. The person decides what to fix.

The three iLEAN pieces applied to 3M detection:

• Edge — vision terminals per station detect the milestones of the cycle (part in, operation starts, operation ends, part out) without touching the PLC. Every cycle is compared with takt and with the history of that specific operator and SKU. That is how mura (cycle variance) and muri (cycles sustained above takt) emerge. It works without network: even if the WiFi drops, the Edge keeps measuring.
• Connect — captures the rhythm of the chain the Edge cannot see: shift changes, raw material arriving late, a supplier message that changes the plan, the operator's voice saying "the grease on line 5 is giving us trouble". Without that capture, mura would be seen but not attributed. Connect is the filler.
• Agent — cross-references Edge data with Connect data and with ERP/MES. It identifies patterns: "the mura at station 3 appears when we change to the green SKU and the operator is new", "the muri of operator X happens when material arrives 8 min late and they compensate by rushing the cycle". It attributes a probable root cause. The person prioritizes and acts.

See the full IRIS architecture →

Before and after

The 3M with classic OEE vs. with continuous iLEAN detection

Aspect	OEE + the manager's eye	With iLEAN Edge + Connect + Agent
Muda visibility	Only what the eye catches or what becomes a logged defect	Continuous, per station, per SKU, per shift
Mura detection	One-off audit, days or weeks	In the moment, with a timestamp
Muri detection	When the operator burns out or resigns	Earlier — a sustained pattern within a few shifts
Root-cause attribution	Manual reconstruction, biased by memory	Agent cross-references Edge, Connect and ERP — a concrete proposal
Measurement frequency	Monthly (OEE), quarterly (studies)	Every cycle
Marginal cost of measuring more	Every extra measurement = a stopwatch and a technician	Zero — the camera is already there

Impact estimate

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

An estimate to be validated with the concrete data of the line. Order of magnitude, not a commitment.

• Production line with OEE stable at 65-78%, continuous improvement active but with no traction to move up, suspected mura on SKU changeovers and muri at one or two workstations.
• Edge pilot over 6-10 critical stations + a 3M detection agent. First quantified map in a few weeks.
• Indicative payback between 4 and 9 months. The hard lever is not raising OEE — it is understanding why it does not rise, attributing cause, and removing it. The continuous improvement manager stops timing and starts correcting.
• Expected reduction of undetected muda ≥ 30% in the first months, depending on the real starting point.

The underlying data

According to research by Fundación BBVA / Ivie, the most digitized sectors raised their productivity by up to 40% compared with the least digitized ones between 2000 and 2021. Invisible 3M is exactly the kind of loss that continuous data attacks and a monthly paper report does not. And the reasonable question ("what if the agent attributes the root cause wrongly?") has an answer: in anchored tasks (cross-referencing time series, comparing against history), the best models pushed the error below 1.5% ^[1]. The agent proposes a probable cause; the continuous improvement manager validates it and acts. The decision stays human.

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

Related: 8 wastes with AI vision · Live Yamazumi · Heijunka