Milk collection alerts, read at second zero

The problem

Milk on a clock, and alerts that arrive through the most chaotic channels there are.

A carrier writes at five forty in the morning: the northern route tanker is running two hours late and short on volume because a farmer had a cooling tank failure. The message lands on the receiving manager's phone. They read it at seven fifteen, because at five forty they were asleep and at half six they were on the floor wearing gloves. By then a process schedule has been built on a volume that will not exist: the process room waiting, milk from another silo pushed out of its window, and a product rescheduled on the fly. The same pattern repeats with the export customer's email changing container quantities, with the farmer's phone call reporting an incident and with the packaging supplier's notice. They all reach specific people who cannot watch an inbox while they are on the floor, and they are all read once decisions have been made that now have to be broken.

• At five forty in the morning the hauler writes that the northern route is two hours late and short on volume because a farmer's cooling tank failed. The message lands on the receiving manager's phone, who reads it at seven fifteen.
• By then a process schedule has been built on a volume that will not exist: the process room waiting, milk from another silo pushed out of its window, a product rescheduled on the fly.
• The same pattern repeats with the export customer's email changing container quantities, the farmer's call reporting an incident and the packaging supplier's delay notice. They all reach people who cannot watch an inbox while wearing gloves on the floor.
• Every alert read late is paid for in a lost window or in milk downgraded to a lower-value range.

How it fits the IRIS system

Connect in chaotic-sources mode — silent copy, twenty-four seven.

Connect in "chaotic sources" mode. The flow:

Finding out sooner is not enough: what changes the morning is the cross-check. The message is read against open orders, the shift's process plan and the real level of each silo, and what comes back is what to bring forward, which silo to use and which export batch stays untouched. The process manager still decides; what changes is the information in front of them — current instead of three hours old, and already reconciled with what the silos actually hold.

• A dedicated mailbox and a messaging number sit on silent copy, twenty-four seven. Nobody changes how they communicate: the carrier keeps writing where they always did.
• When the message arrives, a language model extracts the intent — collection delay, volume shortfall, cooling incident at origin, export order change — and the entities: route, farmer, liters, estimated time, destination.
• That information is cross-referenced with live data: open ERP orders, the shift's process plan and the real level of each silo.
• iLEAN returns a concrete, actionable instruction to the relevant role, not a notification that a message has arrived.
• The process manager decides. The decision remains theirs; what changes is the information they hold.

See the full IRIS architecture →

Before and after

Today's alert versus the cross-checked alert

Aspect	Today, on a phone	With iLEAN Connect
When the hauler's message is read	When its recipient wakes up or takes off the gloves	In seconds, at arrival
Human latency	1.5 to 3 hours	Seconds
The process schedule	Built on volume that will not arrive	Reordered before the delay is felt
Milk pushed out of its window	Downgraded to a lower range	Brought forward in time
What the manager receives	One more notification	Which silo, what to bring forward, what to leave
Who decides	The recipient, late	The process manager, on time

from 1.5 to 3 hours of human latency, to seconds. From rescheduling on three-hour-old information, to deciding on current information. From a process room standing idle, to a proposed reordering before the delay is even felt on the floor.

Impact estimate

Estimated impact — to validate 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.

• Estimated payback 3-7 months, depending on the frequency of incidents on the collection routes.
• Process room downtime caused by missing information that you stop paying for.
• Milk that no longer degrades outside its window, and product that no longer gets downgraded to a lower-value range because the schedule was built on stale volume.
• And decisions taken on current information instead of on information three hours old.

estimated payback 3 to 7 months depending on route incident frequency, counting reduced downtime caused by missing information and milk that no longer degrades outside its window. *Estimate to validate.*

And the fair question from the production manager

«Are you going to read my team's messages?» — no. iLEAN sits in copy on a dedicated mailbox and a company messaging number: it only sees what haulers, farmers, customers and suppliers send to those operational channels. On extraction, classifying the intent of a collection delay or a volume shortfall is an anchored task, where the best models drop below 1.5% error [1] — and the reschedule is executed by a person.

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