Cellular manufacturing with AI — cells designed with your plant's real data, not with the consultant's spreadsheet.

The problem

The cell that was well designed two years ago is no longer the cell your plant needs today.

Moving from a functional layout to a cellular layout is the most profitable organizational change in Lean — and the hardest to sustain. Not because the method is weak: because keeping it alive, in the classical version, depends on manual reviews that almost nobody actually runs.

1. The initial design takes weeks. Part-family analysis (group technology), volumes per SKU from the ERP, operation times from the MES, physical constraints of the floor plan. A properly done project consumes several months of Lean team time.
2. The mix changes and the cell ends up mis-sized. A new SKU comes in from the US customer, an old catalogue item drops out, the ratio between two part families inverts. The cell is still there, but its Yamazumi is out of balance and nobody rebalances it.
3. Cross-training erodes silently. The operator who mastered the critical station leaves, a long absence opens a hole in the skills matrix, and nobody sees it until the day the cell cannot start the shift.
4. And lead time climbs again — the customer complains, a meeting is held, somebody asks to go back to a functional layout "because the cell does not work". The cell did work: it worked against the old mix.

It is the two-gaps syllogism applied to cell design: the data exists (routings in the ERP, times in the MES, cross-training records in HR, real workload on the floor), but it lives on islands; and even if you joined it all up, there was no agent able to simulate groupings and rebalances continuously. AI has just delivered both missing pieces.

How it fits the IRIS system

iLEAN does not propose "the perfect cell" — it keeps the one you already have alive.

Here iLEAN acts as the filler that binds the systems you already own (ERP with routings, MES with times, cross-training sheet, CAD floor plan of the shop) into a brain that simulates cells and rebalances continuously. Without throwing away anything you already have.

Connect reads routings, times and skills. Edge measures workload per station live. The agent proposes groupings and rebalances the Yamazumi against today's mix. The person decides the physical change — the agent does not move machines.

The three iLEAN pieces applied to cellular manufacturing:

• Connect — captures the real ERP/MES history (routings, times, volumes per SKU over the last few years), the live skills matrix of the workforce and the mix changes that arrive through external channels (customer email, EDI, a message from the sales rep). The data you need to design the cell properly stops living on five separate islands.
• Edge — in the cells you already run, it measures effective cycle time per station, micro-stops and real workload. It surfaces saturated and under-loaded stations against the current mix. It works offline: it keeps measuring even if the plant loses its WiFi.
• Agent — runs assisted group technology analysis (grouping SKUs by routing similarity and volume) and simulates the Yamazumi of each cell against the current mix. It proposes a physical move (relocate a machine, redistribute tasks) or a cross-training change (train station X on operation Y). You decide; the agent never rearranges the plant on its own.

See the full IRIS architecture →

Before and after

Classical cellular manufacturing vs. cellular manufacturing with iLEAN

Aspect	Classical design and upkeep	With iLEAN Connect + Edge + Agent
Part-family analysis (group technology)	Weeks with a consultant and a spreadsheet	Continuous, on the full ERP/MES history
Rebalancing when the mix changes	Reactive, once the customer complains	Proactive, the agent warns before the complaint
Skills matrix	Spreadsheet updated once a year	Live, with real data on who ran what
Workload per station in the cell	Cell leader's intuition	Measured continuously by Edge
On-the-job training plan for a critical station	Manual, once people are already missing	Anticipated, assisted by a voice copilot
Evidence for the steering committee	The consultant's slide deck	Real data from your own plant

Impact estimate

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

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the immersion.

• Machining / assembly plant with a functional layout or partially implemented cells, 3-6 main part families, mix varying quarter to quarter.
• Connect + agent pilot (data-assisted design on ERP/MES) across 2 cells. First value expected within a few weeks: the first grouping / rebalancing proposal the Lean team can actually argue with.
• Internal lead time cut ≥ 30% in redesigned part families. Work-in-process between operations cut ≥ 30%. Hard levers: working capital released plus service level recovered.
• Indicative payback between 4 and 9 months, depending on how much internal lead time weighs in your customer commitment and on the hourly cost of the pilot cells.

Why this works now

An AI model asked to opine in the abstract will fill the gaps; the same model anchored to verifiable data — real routings, measured times, closed training records — stops guessing and starts answering ^[1]. That is exactly the difference between a cell redesigned every quarter with plant data and a cell that stayed frozen in the consultant's drawing. The plants that win will be the ones whose people are trained to redesign their own cell, not the ones waiting for an outsider to redesign it for them.

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

Keep pulling the thread: Just in time · Toyota TPS · All Lean methods