One tanker collects from eight farms — and when the inbound analysis tests positive, nobody knows for sure which ones are inside.

The problem

The route takes eight hours and is documented on paper — the positive arrives in twenty minutes.

A collection tanker does not carry the milk of one farm: it carries the blend of a route. Six, eight, twelve stops, each with its liters, its time, its cooling-tank temperature and its sample. When the tanker enters the plant dock, the receiving lab runs its analysis — inhibitors, somatic cells, aflatoxin M1, freezing point, fat and protein — and with that result someone has to decide in minutes what to do with those 25,000 liters.

The problem is not the analysis. The problem is that, at that moment, the exact composition of the tanker lives on a paper route slip that rides in the cab folder and does not get typed up until the next day. The receiving manager knows which route it is and roughly which farms make it up, but does not have in front of them the liters stop by stop or the mapping of which sample belongs to which farm.

1. Reactive: the analysis tests positive for inhibitors. Without a reliable composition, the defensive answer is to reject the whole tanker. The milk of the compliant farms is dumped along with the responsible one's, the day's volume is lost, and the liability dispute with the farmers gets settled weeks later, arguing over delivery notes.
2. Worse still: the tanker is unloaded before the result arrives, because the dock is full and the truck cannot sit idle. When the positive shows up, the problem is no longer 25,000 liters: it is the 100,000 in the silo, plus everything manufactured from that silo that morning.
3. Deferred: weeks later, an official control flags a specific farm. You need to know which tankers carried its milk, which silos they went to and which batches came out of those silos. On paper, that reconstruction takes days of work — and the result is a date-range sweep, not a list of batches.

The classic system works most of the time. The problem is the exception — and the exception costs a whole rejected tanker, a blocked silo or a recall far wider than necessary.

How it fits the IRIS system

iLEAN Tracer is the filler between the truck cab, the receiving lab and the silo.

The dairy industry already has almost all the data — the problem is that each data point lives on an island and arrives late: the routing software with the planned stops, the tanker's flow meter with the actual liters, the paper delivery note with what really happened, the LIMS with the results, the receiving SCADA with the state of the silos, the ERP with the farmer's payment settlement. iLEAN Tracer does not ask you to change anything. It sits on top and joins what already exists around the identity of each stop, each tanker and each silo.

Connect captures the stop at the farm even without coverage. Edge reads the dock and the state of the silos. The agent proposes the unload destination with the analysis in hand. The person signs — never the other way round.

The iLEAN pieces applied to multi-farm tanker traceability:

• Connect in the cab — each collection stop is closed at the farm itself: farm identified, flow-meter liters, start and end time, cooling-tank temperature and sample code. The driver confirms on screen what the equipment has already read, instead of filling in a paper someone will type up tomorrow. If the farm is on a road with no coverage, it records locally and syncs when the signal returns.
• Edge at the receiving dock — the dock terminal cross-references in real time the composition of the tanker that just arrived, the inbound analysis result as the lab publishes it, and the actual state of each silo (level, content, planned destination). Everything on one screen, at the moment the decision has to be made. Without a network, the Edge keeps recording the unload locally.
• Genealogy agent — keeps the farm→tanker→silo→batch chain live and walks it in both directions. Given an alert about a farm, it returns the exact list of affected silos and batches; given a suspect batch, it returns the farms that fed it. The agent proposes the scope of the recall; the quality manager validates and signs.

See the full IRIS architecture →

Before and after

Paper route slip vs. a tanker with known composition in iLEAN Tracer

Aspect	Route documented on paper	With iLEAN Tracer
Tanker composition	Reconstructed afterwards from delivery notes	Known stop by stop on arrival at the dock
Sample per farm	Handwritten code, manual matching	Sample tied to the stop from the origin
Positive on inbound analysis	Whole tanker rejected	Narrowed to the responsible farm
Moment of the decision	Sometimes after unloading to the silo	Before the valve opens, with a proposed destination
Scope of a recall	Date-range sweep, wide out of precaution	Specific list of affected silos and batches
Farmer payment settlement	Double keying from delivery note and LIMS	Liters and grades already in the system, no rekeying
Operation without coverage on the route	n/a — paper	Captures locally, syncs when the network returns

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.

• Dairy plant with its own or subcontracted collection, dozens of farms, several daily routes and a receiving dock with several silos. Stop capture in the cab + an Edge terminal at the dock as a pilot on two or three routes.
• First value expected within a few weeks: the pilot routes' tankers arrive with known composition and the segregation decision is documented before the unload.
• Indicative payback between 4 and 9 months, depending on the history of liters rejected in bulk over inbound positives and the average cost of the recalls you have had to run by dates instead of by genealogy.
• Hard levers: compliant milk that today gets rejected along with the non-compliant; silos that are not blocked because the decision is made before unloading; surgical recalls instead of sweeps; hours of manual route reconstruction and of liability disputes with farmers.

And the quality director's reasonable doubt

“What if the AI assigns a stop to the wrong farm?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recontextualizes a piece of data (reading the flow-meter liters and tying them to that stop's farm), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: the agent proposes the unload destination, the person signs. The three safety rings exist precisely for this.

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