Toyota TPS in a non-automotive plant — copying TPS literally kills the project. Adapting it makes it win.

The problem

You've been through two TPS consultancies and the system still doesn't stick.

The scene repeats itself in many non-automotive plants that decided to make their leap to lean: they hire a consultancy that knows TPS "from Toyota", they build a beautiful kanban board on the wall, they mark the takt on a whiteboard, they install andon lights — and six months later the board has yellowed, the whiteboard is smudged and nobody touches the andon lights because stopping the line for every deviation is far too expensive.

It isn't that TPS doesn't work outside the car. It's that what was contextual to the car was copied literally, instead of being adapted to what really rules in food, pharma or chemical:

1. Fixed takt — on a cured-cheese line or a syrup reactor, the rhythm is set by fermentation, cooking or drying, not by the end customer. Imposing a fixed takt on something with a biological/chemical cadence is violent.
2. Physical card kanban — in a plant with dozens or hundreds of SKUs and variable recipes, the paper kanban board becomes ungovernable by hand. It ends up as a parallel spreadsheet that only the veteran planner understands.
3. Aggressive andon — in automotive, stopping the line for anything was culturally accepted. In your plant the operator tries it for a week, sees the shift supervisor's face when it sounds, and stops pulling the cord.

The problem is not the method. It's that nobody translated the method to your sector. And meanwhile there is still process deviation that isn't caught in time, batches that slip below the expected quality level, and a shift supervisor managing in their head what a system could guarantee for them.

How it fits the IRIS system

iLEAN doesn't ask you to throw away what you already have — it adapts TPS to your reality.

IRIS (Industrial Reality Intelligence Systems) is the category; iLEAN is the system that implements it. And the category was designed precisely for this: to capture the reality of the plant from the bottom up, not to impose a model from the top down. iLEAN doesn't bring you TPS "from Toyota": it brings you the principles of TPS connected to the concrete reality of your sector, keeping the pieces that do work and letting go of the ones that don't.

Classical TPS needs sustained human discipline for years. iLEAN turns that discipline into an industrial nervous system that sustains it for you — and leaves the judgement to the person.

The iLEAN pieces applied to TPS in a non-automotive plant:

• iLEAN Brain (Central) — the multi-agent brain where the agents that sustain the TPS pillars live 24/7. This is where fixed takt is translated into dynamic takt per family, physical kanban into orchestrated digital kanban, and aggressive andon into andon with intelligent escalation that doesn't overwhelm the operator.
• iLEAN Agent — one assistant per key user (planner, shift supervisor, quality, maintenance). The planner's agent levels heijunka by crossing demand, capacity and stock; the quality agent runs jidoka by crossing process, recipe and deviation; the maintenance agent anticipates TPM. They propose; the person signs.
• iLEAN Connect — the filler that seals the cracks between the ERP, the MES (if there is one), the shift supervisor's spreadsheets and the channels through which critical information arrives (a customer email, a supplier's message). Classical TPS assumed the data was on the board; iLEAN assumes it is scattered and brings it together.

See the full IRIS architecture →

Before and after

"Literal copy" TPS vs. TPS adapted with iLEAN

TPS pillar	Literal copy of the automotive model	Adapted with iLEAN Brain + Agent
Takt time	Fixed on a whiteboard, ignores biological/chemical cadence	Dynamic per family, recalculated by an agent with real data
Kanban	Physical board the planner ends up duplicating in a spreadsheet	Digital kanban orchestrated by an agent, integrated with ERP/MES
Jidoka	Andon cord nobody pulls after 6 months	In-line deviation detection + escalation by severity
Heijunka	Static monthly plan signed by the planner	Hour-by-hour levelling with real demand, capacity and stock
Kaizen	Suggestion wall nobody reviews	Continuous loop: agent proposes, plant validates, system learns
Project owner	The external consultant — they leave and TPS leaves with them	The trained internal team — it replicates the next lines on its own

Impact estimate

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

The block below is an estimate to be validated with the concrete data of your plant. It exists so the committee has an order of magnitude; we refine it during the diagnostic.

• Mid-sized food, pharma or chemical plant with one or two clear bottlenecks and multi-SKU production. A previous attempt at "classical" TPS that never got institutionalized.
• A 3-5 day immersion of the mixed team (plant + AI) to identify the Pareto and set up the first two agents (typically jidoka at the bottleneck + heijunka for the planner).
• First value expected within a few weeks; reduction of scrap/critical deviation of ≥ 30% in the first measurable cycle.
• Indicative payback between 4 and 9 months depending on the maturity of your starting data and the cost per deviation in your sector. The hard levers are: scrap avoided, rework avoided, OEE and reduction of unplanned downtime.

And the honest contrast — the Mecatherm lesson

Mecatherm's "Baguette Factory" showcases an entire factory running by itself — bread without people. It is TPS taken to the extreme of automation in order to replace. The sector's figures tell the consequence: flour mills went from 1,647 to 152 between 1970 and 2009; the frozen-dough segment is operated by ~40-50 companies against the ~169,000 of traditional bakery. Technology that replaces and technology that augments have coexisted for decades — and we already know which one builds something that lasts. iLEAN plants its flag on augmenting the people you already have, not on replacing them.

And the CAIO's doubt — does AI make things up?

"What if the agent hallucinates a levelling recommendation and wrecks the plan?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where AI limits itself to recontextualizing data from one system into another (crossing demand with capacity, recalculating takt per family), the best models brought the error below 1.5% ^[1]. And even so, what is critical is never decided alone: the agent proposes, the person signs. The three safety rings are there for exactly this.

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