Production waste and food waste with AI — waste should not surface only at month-end.

The problem

Waste is produced in twenty places and discovered in one: the P&L.

Ask any plant manager how much waste they have and they will give you a percentage. Ask them where it is generated and the answer turns into a list of suspicions. Because real waste is spread across dozens of points that nobody records at the moment they happen:

1. Format trim — the edges of the dough, the cuts that fall outside the size grade, the fillet that comes up short of weight. It goes into the bin with nobody logging how many kilos or from which batch.
2. Line startups and stoppages — the first minutes after a startup produce out-of-spec output; what gets discarded before the line stabilizes is almost never weighed.
3. Quality rejects — the detector ejects, the checkweigher rejects, the inspector pulls a unit aside. Every reject is a data point; almost none of them is logged with its cause.
4. Expiry in the cold store — raw material and work in progress that expire waiting for their work order. It is discovered at stocktaking, weeks after the decision that caused it is beyond fixing.
5. SKU changeover leftovers — whatever is left in the mixer, in the hopper or in the pipework when you switch reference. Kilos purged at every changeover, multiplied by every changeover in the month.

At month-end, accounting subtracts actual consumption from theoretical consumption and one single aggregate number appears. No batch, no line, no cause. With that number nothing can be attacked: the waste meeting turns into an exchange of hypotheses. And the waste reduction targets — plus the sustainability reporting the retailers already ask for and the CSRD demands — stay as estimates that nobody can defend in an audit.

How it fits the IRIS system

iLEAN does not add another waste log nobody fills in — it captures the data in the gesture that already exists.

The bottleneck is not analyzing waste (that is arithmetic); it is that the data never comes into existence. The operator who has just finished a format changeover is not going to walk over to a terminal, open a screen and fill in a form for 12 kilos of trim. Any system that depends on that form is dead before it starts. iLEAN inverts the problem: it brings capture to the point where waste happens, at a logging cost close to zero — and without asking you to change ERP, MES or scales.

Connect captures every kilo where it is generated — scale, photo or voice. Edge counts line rejects automatically. The agents attribute every record to its cause and rank the 3 that hurt most. The person validates — without a signature, the data does not enter.

The iLEAN pieces applied to production waste and food waste management:

• Connect — the connected scale sends the weight on its own; the trim bin is logged with a photo; and the operator with an earpiece dictates it without putting down what is in their hands: “iLEAN, log this: 12 kilos of trim from a format changeover”. Every record is born with line, shift, batch and time. Waste stops being an accounting number and becomes a stream of events with context.
• Edge — at the points where there is already a camera, a detector or a checkweigher, Edge counts the rejects automatically: units ejected, kilos set aside, minutes of out-of-spec startup. With no operator involvement and without taking critical data out of the plant: processing happens in ring 1.
• Tracer and Writer agents — Tracer cross-references every record with the line context (was there a startup? an SKU changeover? a quality reject? a cold store release past its date?) and attributes every kilo to its cause; when in doubt, it asks instead of inventing. Writer aggregates, computes the Pareto of causes, ranks the 3 that weigh most and prepares both the weekly action plan and the food waste report built on primary data. The person in charge validates the attribution — without a signature it does not enter the system.

The Brain that orchestrates the agents lives in Central; critical data stays in ring 1 of the plant, not in some random cloud (see the IRIS architecture and the three safety rings).

See the full IRIS architecture →

Before and after

Waste aggregated at month-end vs. waste captured at the point with iLEAN

Aspect	Accounting close + spreadsheets	With iLEAN Connect + Edge + Agents
When waste surfaces	At month-end, aggregated in the P&L	The moment it is generated, point by point
Attribution by cause	Hypotheses in the monthly meeting	Every kilo with its cause: startup, SKU changeover, defect, expiry
Detail by batch and line	Does not exist — one single global number	Record with line, shift, batch and time
Cost of logging a discard	A form or paper report nobody fills in	Scale, photo or voice — seconds, without putting the tool down
Food waste reporting (CSRD, retailers)	An estimate built on the close, hard to defend	Auditable primary data, broken down by cause and batch
Prioritizing actions	Every cause at once, none of them properly	A living Pareto — the 3 heaviest causes, tracked weekly

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. It is there so the committee has an order of magnitude; we refine it during the diagnostic.

• Plant with 2-8 lines and aggregate waste of between 2% and 6% of consumption, with no reliable breakdown by cause. Frequent SKU changeovers and growing pressure from retailers and the CSRD for defensible waste figures.
• Deployment of Connect + Edge + Tracer/Writer agents on one pilot line, integrated with the existing ERP/MES and the scales already in place. Visibility of waste by cause from the very first week — before touching anything, you already know where every kilo goes.
• Expected reduction of ≥30% on the first causes attacked (estimate to be validated: every plant's Pareto is different, but the top 3 causes usually concentrate most of the kilos). Food waste reporting goes from estimated to real data with no extra work.
• Indicative payback between 4 and 9 months (estimate to be validated), depending on current waste volume, raw material cost and the number of lines brought in after the pilot.

And production's reasonable doubt

“What if the AI attributes waste to the wrong cause?” — hallucination is a problem of free generation, not of anchored tasks. Attributing a waste record to its cause is precisely an anchored task: the AI recontextualizes one data point (kilos, line, time) using another (the line event happening at that moment) and applies verifiable rules. In this kind of task, the best models brought error below 1.5% ^[1]. And even so, the person validates the attribution — without a signature, the data does not enter the system. The three safety rings are there for exactly this.

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