The alert you read three hours late, read at second zero

The problem

The alert is read when the knives are already mounted.

From customers come pulled-forward shipments, strip width changes and warnings that their press is down waiting for material. From mills come coil delays, reassigned heats and widths delivered that are not the ones ordered. Cutting schedules in a service center are a jigsaw of widths: each run combines different orders to make the most of the master coil. An alert read late is not an administrative nuisance, it is a knife set mounted for nothing and a whole setup lost. Or worse: a customer with a stopped press. And it is not solved by asking people to check their email more often. It is solved by putting something in place that listens all the time.

• From customers come pulled-forward shipments, strip width changes and the warning that their press is down waiting for material. From mills come coil delays, reassigned heats and a width delivered that is not the one ordered.
• Each one lands in one person's inbox, competing with a hundred other messages, and gets opened two or three hours later.
• A cutting schedule is a jigsaw of widths that combines several orders on one master coil. An alert read late is a knife set mounted for nothing and a whole setup lost. Or a customer with a stopped press.
• It is not solved by asking people to check their email more often. It is solved by putting something in place that listens all the time.

How it fits the IRIS system

Connect in chaotic-sources mode — silent copy, around the clock.

Connect in chaotic sources mode. A dedicated mailbox and a business messaging number sit silently in copy, around the clock. When a message arrives, the LLM extracts the intent — pull-forward, specification change, supply delay, claim — and the entities: customer, order, heat, width, gauge. It then cross-checks against the ERP and against real line load, and returns a proposed action to the operations manager: which run to resequence, which alternative coil in the warehouse covers that specification, and which customer to notify. Not a notification that an email arrived: the action.

The difference between reading the alert and acting on it is the cross-check: it is not enough to know sooner, you need to know which run it hits, which coil in the warehouse covers the new specification and which customer has to be told. That is what comes back, to the operations manager, as a proposed action.

See the full IRIS architecture →

Before and after

Today's alert versus the cross-checked alert

Aspect	Today, in the inbox	With iLEAN Connect
When it is read	2-4 hours later	In seconds
A width change from a customer	Discovered with the knives mounted	Caught before the setup
A coil delay from the mill	A line waiting for material	An alternative coil proposed from stock
Rescheduling against	This morning's snapshot	Real line load
What you receive	One more email	Which run to resequence and whom to notify
Who decides	Whoever opens the email	The operations manager — iLEAN proposes

from 2-4 h of human latency to seconds. Rescheduling against real line load rather than against this morning's snapshot.

Impact estimate

Impact estimate — to be validated with your numbers.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 3-7 months, depending on the real frequency of schedule changes.
• Measured in knife setups avoided: every alert caught before the changeover is a setup not lost.
• And in deliveries saved: the customer with a stopped press gets material from an alternative coil instead of an apology.
• From 2-4 hours of human latency to seconds, with the schedule resequenced against real line load rather than against this morning's snapshot.

estimated payback 3-7 months depending on the real frequency of schedule changes, measured in knife setups avoided and deliveries saved. *An estimate to be validated*.

And the fair question from the production manager

“Are you going to read my team's email?” — no. iLEAN sits in copy on a dedicated mailbox and a business messaging number: it only sees what customers and mills send to those operational channels. On extraction, classifying the intent of a message — pull-forward, width change, coil delay, claim — and pulling out order, heat, width and gauge is an anchored task where the best models drop below 1.5% error [1], and the resequencing is executed by a person.

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