Yamazumi and line balancing with AI — from the notebook stopwatch to a live plant-floor chart.

The problem

The Yamazumi is obsolete the day it gets printed.

The industrial engineer spends three days with a stopwatch in hand, times five different operators on five different SKUs, averages the results, draws the stacked bars and presents the balance to the committee. The following Monday a new SKU comes in, an old one goes out, a veteran operator leaves on holiday, and the Yamazumi on the wall stops matching what the line actually does.

The supervisor knows it. They can feel it in the first shift: one station is always rushing (muri), another is standing by (muda), and the bottleneck moves on its own every time the mix changes. But a formal rebalance costs another three days of stopwatch work, so the line keeps running on a balance calculated months ago, patched with craft and instinct. The problem is not that people don't know Lean — it's that measuring reality is so expensive that it gets measured once a quarter, and the plant spends the rest of the year drifting.

How it fits the IRIS system

iLEAN turns the stopwatch into a permanent sensor — and the engineer into a decision-maker.

Line balancing is a classic reality-capture problem: if you could measure every cycle of every operator at no cost, you would balance continuously. Measuring used to be expensive, so it was done rarely. iLEAN acts as the filler that seals the gap between the ERP/MES (which knows which SKU is due), the line (which knows how long it really takes) and the industrial manager (who decides how to move the work). It replaces no system — it connects them.

Edge times every station at once, without stopping. Connect cross-references the plan and the changes that arrive by message. The agent proposes the rebalance to the supervisor. The person signs off — the agent never changes a balance on its own.

The three iLEAN pieces applied to line balancing:

• Edge — a vision terminal per station detects the cycle milestones (part in, operation starts, operation ends, part out). There is no need to touch the PLC or wire new sensors; the camera learns to read the operation the way we would read it. Real times per operator, per SKU and per shift, continuously. It works offline: the measurement cycle keeps running even if the WiFi drops.
• Connect — captures the plan from the ERP/MES and the last-minute changes wherever the data comes from: a customer email pulling an order forward, an instant message from the production manager moving a SKU, a spreadsheet from the planner changing the sequence. The filler that seals the cracks between the official systems and the real channels where decisions actually happen.
• Agent — recalculates the Yamazumi live per SKU and per shift, compares it with the current takt, identifies the station drifting into muri and the one drifting into muda, and proposes a concrete rebalance to the supervisor (move this sub-operation from station 3 to station 2, split this inspection between 4 and 5). The person decides.

See the full IRIS architecture →

Before and after

Stopwatch balancing vs. a live iLEAN Yamazumi

Aspect	Stopwatch + notebook + spreadsheet	With iLEAN Edge + Connect + Agent
Measurement frequency	A study every 3-6 months, days of work	Every cycle, continuous, no marginal cost
SKU coverage	One sample of the "representative" SKU	Every real SKU that runs down the line
Reaction to a mix change	The Yamazumi on the wall is not updated	The agent recalculates and proposes a rebalance per shift
Muri detection	When the station breaks down at the end of the shift	When the cycle drifts systematically from takt
Muda (waiting) detection	Seen by eye, never quantified	Quantified per station, per SKU and per shift
Who runs the time study	A technician, 3 days, everything else on hold	The camera — the technician decides what to do with the data

Impact estimate

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

An estimate to be validated with the specific data of your plant. Order of magnitude, not a commitment.

• Pilot line: 6-10 stations, multi-model assembly (tier-2 automotive, electronics, medical device, appliances). A mix of 8-15 active SKUs per week.
• Edge over each station + Connect against the ERP/MES + Agent working with the supervisor. First live Yamazumi in a few weeks.
• Indicative payback between 4 and 9 months, depending on the current cost of your time studies, how often the mix changes and how much muri/muda the line carries today.
• Expected reduction of total cycle time ≥ 30% of the imbalance the line absorbs today (the slack between the slowest station and takt), not of takt itself.

And the underlying fact that closes the case

Continuous line balancing is exactly the lever that turns digitization into plant-floor productivity — the same team, the same machines, a different rhythm. The reasonable doubt of "what if the AI gets a cycle wrong?" is legitimate, and the answer is direct: hallucinations are a problem of free generation; in tasks anchored to the source (reading the cycle and comparing it with the standard), the best models brought error below 1.5% ^[1]. And even so, nothing critical is decided alone: the agent proposes, the supervisor signs off.

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

Related: Heijunka (production leveling) · Dynamic takt time · Muda/mura/muri detection