The classic MES replans at the end of the day. Your plant changes every hour.

The problem

The MES knows what should be happening. It doesn't always know what is happening.

In any multi-product plant running MES + MRP there are two truths operating at different speeds:

1. The MES/MRP truth — the signed-off plan, the released orders, the schedule leveled at the end of the day. Stable, auditable, reliable.
2. The truth of the plant today — line 3 has been down for 90 minutes because of a breakdown only the shift supervisor knows about; the sales rep passed along a rush order by email at 8:00; a critical supplier gave notice in a message that the truck leaves tomorrow instead of today. All that information exists — but it isn't in the MES at second zero.

The planner ends up acting as human glue between the two truths: half the day spent reading email, calling the floor, checking a spreadsheet, adjusting orders by hand in the MES. And they do it well — but only at human speed. The day they go on holiday, the plant runs on the old plan and starts losing orders, lead time and raw material.

The problem isn't that the MES is bad. It's that the MES was never designed to live with a shifting reality in real time — it was designed to impose the plan. And between the plan and reality there is a gap that until recently could only be filled with human hours.

How it fits the IRIS system

iLEAN doesn't tear out your MES — it seals the cracks between what it predicts and what happens.

IRIS (Industrial Reality Intelligence Systems) is the category; iLEAN is the system. Applied to multi-product MES/MRP the metaphor is exact: iLEAN is the filler that closes the joints between the systems you already have. You don't tear down the wall or the tiles: you fill what was left hollow. And that gap is enormous — because the cross-cutting work between MES, ERP, the plant floor and external channels is still done by hand with email, phone calls and spreadsheets.

The MES says what should happen. iLEAN sees what is happening and replans in minutes what the MES would replan at the end of the day. The planner signs off. The plant runs on a live plan.

The iLEAN pieces applied to multi-product MES/MRP:

• iLEAN Brain (Central) — the brain where the agents that keep the plan alive 24/7 live. It reads the MES and MRP outputs, cross-checks them against the reality captured by Connect, and maintains a shadow version of the plan that moves at the pace of events.
• iLEAN Agent (planner's / shift supervisor's) — one assistant per key role. The planner's agent proposes replans in minutes when a change lands; the shift supervisor's agent translates the plan into operational instructions for the line, validating against the MES at every step. They propose; the person signs off.
• iLEAN Connect — the multichannel capture the MES doesn't do: the sales rep's email, the supplier's message, the transcribed customer call, a photo of the analog panel on the old machine. Three integration levels (manual / intermediate / integrated) so no island is left out, without asking you to change a machine or buy a license.

See the full IRIS architecture →

Before and after

MES/MRP "alone" vs. MES/MRP augmented with iLEAN

Aspect	MES/MRP alone	MES/MRP + iLEAN Brain + Agent + Connect
Replanning frequency	End of shift / weekly close	Minutes, on any relevant change
Rush order by email	Forwarded by the sales rep to the planner	Captured at second zero, agent proposes the adjustment
Unplanned breakdown	Shift supervisor calls it in by phone	Logged, automatic replan of the affected plan
Supplier delay	A message that arrives at 8:00 and is acted on at 11:00	Connect captures the message, agent replans downstream
Legacy machine with no interface	A side spreadsheet only the shift supervisor knows	Intermediate capture from the local PC or a photo of the panel
Lead time met for the customer	Whatever percentage the old plan allows	Measurable improvement from running on a live plan

Impact estimate

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

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

• Multi-product plant with an MES (vertical or standard) + a large ERP, with the planner as the human bottleneck.
• A 3-5 day immersion of the mixed team on the plant floor. Pareto identified: typically the 5-10 SKUs or families that concentrate 80% of the cost and the lead time.
• First agent (dynamic replanning on rush orders + multichannel capture via Connect) running in a few weeks.
• Reduction in average multi-product lead time between 15% and 35%. Reduction in the planner's monthly hours spent on reactive management ≥ 40%.
• Indicative payback between 4 and 9 months. The hard levers are: on-time delivery to the customer, lead time, rework avoided and human hours freed up.

The underlying fact — the cost of integration hit the floor

"I'm not integrating that old machine because it costs more than it's worth" — that excuse has expired. If there is one thing LLMs do exceptionally well with minimal hallucination, it is recontextualizing information from one format and returning it in another. Integration is not a magic button, nor is it free: connecting to each island is still real work. What changed is that it is now feasible and affordable, not instantaneous. And that shift is what makes augmenting your MES/MRP with AI economically sensible today and not five years ago.

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

"What if the agent pushes a wrong order into the MES because of a hallucination?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI simply recontextualizes data between systems (read the email, extract the order, compare it against capacity), the best models brought the error rate below 1.5% ^[1]. And the agent never writes directly into the critical OT side of the MES: it proposes, the planner signs off, and the write passes through ring 2 validation. The three rings exist precisely for this.

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