The customer alert, read within seconds instead of hours

The problem

The critical alert arrives on time. The person who has to read it does not.

In a long steel mill the alerts that break the schedule — a customer chemistry specification change, a scrap supplier delay, a shipment brought forward — already arrive, but they arrive by email or WhatsApp to specific people:

• An automotive Tier 1 notifies by email a chemistry specification change affecting a heat already scheduled for this week.
• In parallel, the main scrap supplier flags on WhatsApp that the truck with the material planned for the next shift is running hours late.

The melt shop lead and the scheduler cannot watch an inbox or a phone in real time: they are at the furnace, on the heat in progress or in the schedule meeting. By the time somebody reacts, the basket has already been charged with the usual mix, the next heat has already started or the shipment has already been scheduled. The whiteboard heijunka breaks every day for the same reason: the schedule is set in the morning meeting and then the world changes — and the messages announcing it sit unread for hours. It is the natural bottleneck of the inbox, not of the plant.

How it fits the IRIS system

Connect in chaotic-sources mode — it reads the email and the WhatsApp before the next basket is charged.

The customer alert and the supplier alert need no EDI integration project and no new portal. They need the messages that already arrive to be read, understood and cross-referenced against live MES and ERP status at the very instant they land — without waiting for the right person to have a gap. That is what Connect does in chaotic-sources mode.

A Tier 1 email and a supplier WhatsApp arrive at the same time. Connect reads them at zero latency, extracts what is changing and for which customer, grade and material, and returns a concrete re-levelling action to the melt shop lead. Nothing is touched in the MES or the ERP without a person confirming it.

How Connect works in chaotic-sources mode on customer and supplier alerts:

• Mailbox and WhatsApp Business on silent copy, 24/7 — Connect receives a copy of the mail and of the WhatsApp Business channel where alerts already arrive, replacing nobody and changing nothing about the channel the customer and the supplier already use.
• The LLM extracts intent and entities — a chemistry specification change, a supply delay or a shipment brought forward as intent; customer, grade, material, tonnage and dates as entities. Within seconds the message stops being loose text and becomes structured data.
• Immediate cross-check against live MES and ERP status — Connect compares the extracted intent against the heat schedule, the baskets pending charge and the material actually available, and detects which heats and which shipments are affected.
• A concrete re-levelling action returned to the right role — not "you have a message", but a proposal grounded in the real state: "reassign heat C-4821 to the construction customer; bring C-4823 forward with the high-Mn recipe, which already matches the material on hand". The melt shop lead sees the option already built, without reconstructing it from two threads.
• Human confirmation before touching the MES or the ERP — Connect resequences nothing and holds nothing on its own. The person confirms the action, and only then does it move into the heat schedule.

See the full IRIS architecture →

Before and after

Whiteboard heijunka vs. real-time heijunka with Connect

Aspect	Inboxes with no automatic cross-check	With iLEAN Connect
Detecting the critical alert (email or WhatsApp)	Read when the melt shop lead or the scheduler has a gap — 2-4 h of human latency	Read and extracted within seconds, 24/7
Cross-check against the heat schedule and available material	Manual, reviewing MES, ERP and the message thread by hand	Automatic, against live MES and ERP status at zero latency
Basis for the heijunka decision	The status as it was three hours ago, in the morning meeting	The real state of the schedule at the moment of the alert
Heat affected by the specification change	Discovered with the basket already charged or the heat already started	Reassignment proposed within seconds, before the next basket is charged
Scrap delay from the supplier	Discovered when charging — hours of downtime or an improvised mix	Re-levelling proposed with the recipe that already matches the material on hand
Confirmation and traceability	The whole loop is manual; the thread is reconstructed after the fact	Human confirmation required; message, extraction and action filed per heat and per order

Impact estimate

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

• Long steel mill with specification-demanding customers — automotive, construction — and critical alerts arriving by email and WhatsApp to specific people.
• Connect pilot in chaotic-sources mode on the customer mailbox and the supplier WhatsApp Business channel — without touching the MES or the ERP except on human confirmation. First value expected within a few weeks.
• Indicative payback of 3 to 7 months, depending on how many critical alerts your plant handles. Estimate to be validated.
• Expected avoidance of out-of-specification heats caused by a change not received in time — every heat that comes out to the old standard is downgraded material or rework that does not happen. Estimate to be validated against your data.
• Expected avoidance of downtime hours from scrap that did not arrive when expected — the re-levelling is proposed before the basket is charged, not when the furnace is already waiting. Estimate to be validated against your data.

And the fair question from the production manager

"What if Connect misreads a WhatsApp and re-levels a heat that should not have moved?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recontextualizes a specific figure from one medium to another (reading a customer or supplier alert and comparing it against the MES and ERP heat schedule), the best models brought the error below 1.5% ^[1]. And even then, the critical call is not made alone: Connect proposes the re-levelling and a person confirms before the schedule is touched. The three safety rings exist precisely for this.

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