Manual heijunka leaves levelling to the veteran planner — and to their spreadsheet.

The problem

Levelling lives in the planner's head. And in their parallel spreadsheet.

Heijunka is a beautiful concept on paper: level production so there are no peaks and valleys, spread the mix across the hours, smooth the load upstream. In practice, in a multi-SKU plant with no AI, what actually happens is this:

1. The planner builds the best levelling they can on Friday afternoon, in a spreadsheet only they understand.
2. On Monday sales pushes in an urgent order, on Tuesday a critical raw-material supplier flags a delay, and on Wednesday a line stops for 2 hours with an unplanned breakdown. Friday's plan is already worthless.
3. The planner re-levels by hand whenever they can — and meanwhile the plant runs on the old plan, with overstock in some families and an imminent stockout in others.
4. The day the planner goes on holiday or retires, half the plant leaves with them.

The problem is not the method: it is that nobody can sustain heijunka by hand in a modern multi-SKU plant where things change every hour. You need an industrial nervous system that keeps levelling alive — and that gives the planner back the judgement, not the arithmetic.

How it fits the IRIS system

iLEAN doesn't replace the planner — it gives them an agent that never tires of recalculating.

IRIS (Industrial Reality Intelligence Systems) is the category; iLEAN is the system. Applied to heijunka the principle is clear: data joined together with no agent on top is a library without a librarian. Levelling needs someone to cross demand, capacity and stock in real time — and that is exactly what a well-designed agent does, without ever running out of stamina.

Manual heijunka levels once a week. AI-augmented heijunka levels hour by hour, with every change. The planner stops managing a spreadsheet and goes back to deciding.

The iLEAN pieces applied to digital heijunka:

• iLEAN Brain (Central) — the multi-agent brain that runs 24/7 and keeps the levelled plan alive. This is where the multi-objective optimization logic lives (lead time, stock, OEE, changeover cost) and where APS/ERP outputs are crossed with plant reality.
• iLEAN Agent (the planner's) — an assistant dedicated to the planner. It proposes the levelled plan for the coming day with the best achievable balance, flags critical SKUs and their constraints, and reworks the sequence when a change comes in. The planner validates, adjusts and signs off. It proposes; the person decides.
• iLEAN Connect — it captures the changes that break the old plan: a customer email with an urgent order, a supplier message about a delay, a breakdown logged by maintenance, a manual adjustment in the MES. At second zero the agent knows something changed, with nobody forwarding anything or calling a meeting.

See the full IRIS architecture →

Before and after

Manual heijunka vs. digital heijunka with iLEAN

Aspect	Manual heijunka + spreadsheet	With iLEAN Brain + Agent
Levelling frequency	Weekly (sometimes monthly)	Hour by hour, with any relevant change
Manageable SKU mix	5-10 comfortable, 50+ unmanageable	50, 500 — the agent carries the complexity
Rush order change	Half an afternoon of the planner recalculating	Rework proposed in minutes, person signs off
Planner's knowledge	In their head and their spreadsheet — it leaves with them	Captured in the agent, stays in the company
APS/MES integration	Manual export/import	Native read/write, nothing to rewrite
Total inventory	Inflated by psychological safety margin	Optimized per SKU with equal or better coverage

Impact estimate

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

This block is an estimate to be validated with the real data from your plant. It works as an order of magnitude for the steering committee; we refine it during the diagnostic.

• Plant with 50-300 active SKUs, a veteran planner running levelling on a spreadsheet plus a classical APS that recalculates at end of day.
• Immersion of 3-5 days, first planner agent operating within a few weeks, replanning in minutes the changes that arrive by email or messaging.
• Total inventory reduction between 15% and 30% in the first measurable cycle, with equal or better service level.
• Reduction of the planner's monthly hours spent on repetitive calculation ≥ 40%, releasing those hours to campaign decisions and customer negotiation.
• Indicative payback between 4 and 9 months depending on volume and the cost of tied-up working capital. The hard lever is releasing working capital without touching service level.

And the CAIO's doubt — does the AI invent the plan?

"What if the agent proposes an absurd plan based on a hallucination?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recombines and optimizes data from existing systems (crossing demand with capacity, recalculating a sequence), the best models brought error below 1.5% ^[1]. And even then, the plan does not come out of the AI — it comes from the planner who validates the agent's proposal. The three safety rings exist precisely for this.

The underlying data point — digitalization vs. survival

Between 2000 and 2021, the least digitalized sectors barely improved productivity; the most digitalized ones raised it by up to 40% ^[2]. Digitalization separates those who survive from those who fall — and digital heijunka is one of the most profitable entry points because it has a clear owner (the planner) and a measurable KPI (working capital / service level / OEE).

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

^[2] Fundación BBVA / Ivie — productivity and digitalization, 2000-2021.