When the supplier flags a lot, the clock starts running

The problem

Between the email being read and the line being blocked, units go out of the door.

The alerts that do the most damage are precisely the ones that do not arrive by electronic data interchange. They arrive as prose, in an email or a message, addressed to one named person, sometimes in another language and sometimes out of hours. And in passive safety the clock really counts: between the alert being read and the line being blocked, units go out of the door. This is not a negligence problem, it is a channel problem: nobody can keep an inbox open twenty-four hours a day. Answering the next question — how many units are affected and where are they? — today requires a traceability query that is neither instant nor within reach of whoever is on the night shift. And while that answer does not exist, the only prudent decision is to over-contain.

• The gas generator supplier flags a deviation on a propellant lot. The notice arrives as prose, in an email addressed to one named quality person, sometimes in another language and sometimes at eleven at night. The alerts that do the most damage are exactly the ones that do not come by EDI.
• While the inbox stays closed, that lot keeps being assembled into modules and keeps shipping. Every hour of delay multiplies the number of affected units and with it the cost and the scope of the containment.
• This is not negligence, it is a channel problem: nobody can keep an inbox open twenty-four hours a day, and the night shift has no traceability specialist to ask.
• Answering the next question — how many units carry that lot and where are they? — needs a traceability query that is neither instant nor within reach of whoever is on shift. Until it exists, the only prudent decision is to over-contain.

How it fits the IRIS system

Connect in chaotic-sources mode — a silent copy that never closes its inbox.

Connect in chaotic-sources mode.

The difference between reading the alert and acting on it is the cross-check: finding out sooner is not enough, you need to know which serial numbers carry the lot, how many are in the warehouse, how many in transit and how many already at the carmaker. That is what comes back, in seconds, together with a concrete proposal — and a person decides whether to block and signs.

• A dedicated mailbox (and a messaging number where relevant) sits on silent copy around the clock.
• The LLM extracts intent — lot deviation, specification change, non-conformity, delay — and entities: supplier, part reference, lot identifier, declared scope.
• It cross-references live unit traceability data: which serial numbers carry that lot.
• It returns an actionable answer, not an acknowledgement: how many units in the warehouse, how many in transit, how many already delivered, and a concrete proposal (block the warehouse, open containment, notify the customer).
• Humans in command: iLEAN proposes, a person decides and signs.

See the full IRIS architecture →

Before and after

Today's alert versus the cross-checked alert

Aspect	Today, in one inbox	With iLEAN Connect
When the alert is read	2-6 h, or a whole shift	Seconds, around the clock
Which serial numbers carry the lot	A specialist's query, next morning	Cross-referenced against live traceability
Where the units are	Unknown until somebody looks	Warehouse, in transit, delivered — counted
Scope of the containment	A month of production, out of caution	The units that actually carry the lot
Notice in another language or out of hours	Waits for its reader	Read like any other
Who decides the block	Whoever opened the email	A person, on a proposal with numbers

from 2-6 h (or a whole shift) of human latency to seconds · from containment sized out of caution to containment bounded to the genuinely affected units · from a traceability query that needs a specialist to an answer available on the spot.

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 how often critical alerts arrive; the first properly bounded incident usually pays for it.
• The dominant saving is the containment scope: from blocking a month of production to blocking one lot.
• From 2-6 hours of human latency — or a whole shift — to an answer in seconds.
• A traceability query that needed a specialist becomes an answer available to whoever is on shift.

the dominant saving is the reduction in containment scope — going from blocking a month of production to blocking one lot. Estimated payback 3-7 months depending on how often critical alerts arrive; the first properly bounded incident usually pays for it. *Estimate to validate.*

And the fair question from the production manager

"Are you going to read my quality team's email?" — no. iLEAN sits on silent copy of a dedicated mailbox and, where relevant, a company messaging number: it only sees what suppliers send to those operational channels. On extraction, classifying the intent of a notice and pulling out supplier, part reference and lot identifier is an anchored task, where the best models drop below 1.5% error [1] — and the block itself is decided and signed by a person.

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