The big order calls for one consistent shade — and the plant produces batch by batch with unavoidable variance.

The problem

Three systems, one personal spreadsheet and a decision only the veteran knows how to make.

An organic pigment order for a serious customer arrives with a tight COA: shade, tinting strength, particle size distribution, residual monomer. The plant produces the pigment batch by batch in a reactor, and every batch comes out with its natural variation. To serve the order with a single consistent shade, somebody has to decide which batches to blend so they offset each other before packing. That decision depends on three sources that are almost never in the same place:

1. The order and its expected COA — in the ERP / CRM, with the customer's history and their real tolerances (not just the theoretical ones).
2. The batch QC — in the LIMS, with shade and tinting strength measured on the lab spectrophotometer.
3. The sequence of upcoming reactors — in the MES or in the shift lead's head: what is charging right now, what starts next shift, which reactor is free for an offsetting run.

The production manager joins the three islands in his head, helped by a personal spreadsheet holding the last few months of decisions. That is the island of the veteran's knowledge — the most valuable and the worst protected. The day he retires, half the plant walks out with him. And in the short term, when a last-minute order change comes in, replanning happens by phone and meeting, not by data.

How it fits the IRIS system

iLEAN does not replace the manager — it captures his judgment and amplifies it.

The multi-batch coordination problem is not a lack of information: it is information living on islands plus a personal spreadsheet filling the gaps. iLEAN acts as the putty that seals those gaps without asking you to change ERP, LIMS or MES — and, above all, without asking the manager to change the way he works. The personal spreadsheet can go on existing; it just stops being the only source of truth.

Connect carries. The Agents propose the grouping. The manager reviews, adjusts and signs. The veteran's knowledge is captured as a permanent capability.

The iLEAN pieces applied to multi-batch organic pigment:

• Connect — captures the three islands at second zero: order and COA from the ERP/CRM, batch QC from the LIMS, reactor status from the MES/SCADA. And it also captures what arrives from outside (the customer email with the special note, the sales rep's WhatsApp with the order change) without anyone forwarding anything.
• Grouping agent — proposes which batches to blend for each order to minimize the variance of the final shade, against the customer's COA and their real history. It builds the blend record with batch-by-batch traceability and presents it to the manager. The manager accepts, adjusts or rejects with a note — that decision stays with the agent as learning.
• Sequencing agent — proposes which reactor to charge next to offset the average of available stock. If the customer changes the order, it replans the queue without calling a meeting. Connect carries, the Agents propose, the person signs. The Agents do not discharge a reactor on their own.

See the full IRIS architecture →

Before and after

The veteran's parallel spreadsheet vs. Agents with the veteran's judgment inside

Aspect	Manual coordination	With iLEAN Agents
Source of the customer COA	ERP customer record + personal note	COA and real tolerances from history, at second zero
Grouping decision	The manager's personal spreadsheet	Agent proposal on joined data; the manager signs
Batch QC	Lab sheet that gets printed	LIMS read instantly, cross-referenced with the live order
Last-minute order change	Meeting, phone calls, manual replanning	Automatic replanning proposal; the manager decides
Capturing the veteran's judgment	It leaves with him the day he retires	Every decision informs the agent — the capability stays
Final blend record	Handwritten sheet with transcription risk	Complete dossier with batch-by-batch traceability

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 diagnostic.

• Organic pigment plant with 4-8 active reactors, azo and/or phthalocyanine and/or quinacridone families, a mix of customers with tight COAs and large orders that require multi-batch blending.
• Agents pilot on the Pareto family (highest volume or tightest color requirement). First value expected within a few weeks: a grouping proposal for a large order with a consistent final blend record.
• Expected reduction in batches downgraded to a lower grade because of shade deviation ≥ 30%, with a defensible floor in kg/month recovered at premium price.
• Indicative payback between 4 and 9 months, depending on monthly volume, the price gap between premium and downgraded shade, and how often orders change at the last minute.
• The hard lever is the veteran's knowledge captured as a permanent capability and the blend dossier ready for audit with no extra manual work.

And the production manager's reasonable doubt

“What if the agent proposes a blend that does not work on the floor — the theoretical composition is not the real one?” — hallucination is a problem of free generation, not of anchored tasks. The agent does not propose invented compositions: it cross-references the real QC from the LIMS with the history of blends signed by the manager. Every decision you adjust goes back to the agent as learning, with nothing to program. In anchored tasks, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: the proposal is there for you to sign, not to run by itself.

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