24/7 OEE in food manufacturing with AI — the micro-stops nobody writes down are your biggest loss pool.

The problem

The weekly spreadsheet OEE is calculated on what the shift remembers — and the shift does not remember micro-stops.

The routine is the same in most food plants: at the end of the shift, the operator fills in the production report from memory — units, stops, causes — and on Friday someone in the engineering office keys those reports into a spreadsheet to rebuild the week's OEE. The number arrives late, and it arrives wrong, through four specific cracks:

1. Micro-stops do not exist for the system — the film jam on the wrapper, cartons running out at the end of the line, the doser adjustment. Three minutes here, four there, fifteen times a shift. None of them makes it into the report, because nobody is going to drop what they have in their hands to write down three minutes. Added up, they are the line's biggest loss pool.
2. The causes are generic — “breakdown”, “adjustment”, “cleaning”. At that level of detail nothing can be prioritized: a breakdown of what, on which machine, with which SKU running? The detail exists in the head of whoever was standing there, and it is lost at 10 p.m. when they clock out.
3. Memory rounds off — a report filled in eight hours after the fact compresses three stops into one, forgets the one at start-up and estimates durations by eye. It is not bad faith; it is how memory works at the end of a packing shift.
4. The data arrives when it is no longer useful — Monday's OEE is known on Friday. By then the problem SKU is off the line, the format has already changed, and the committee conversation is archaeology, not improvement.

The result: investment and continuous-improvement decisions are taken on an OEE that nobody has actually measured. And the biggest loss of all — micro-stops — does not even appear in the picture.

How it fits the IRIS system

iLEAN does not ask the operator for more discipline — it takes the report out of their hands and measures the line instead.

The bottleneck is not calculating OEE (that is a division); it is that the underlying data does not exist: the line does not emit it, the report rebuilds it from memory and the spreadsheet consolidates it too late. iLEAN attacks the source: it captures the data where and when it happens — without changing your ERP or your MES, and without adding one more sheet of paper to the shift. And, importantly: iLEAN measures the line — counters, stops, speeds — not people. It does not hold the operator to account; it takes paperwork away.

Connect listens to the line counter and logs every stop to the second. The operator classifies the cause by voice, without dropping what they are doing. The agents consolidate real 24/7 OEE and flag the 3 stops that cost the most points. Nothing enters the system without human validation.

The iLEAN pieces applied to 24/7 OEE:

• Connect — connects to the line counter: if the machine is modern, by direct integration (OPC-UA, Modbus, encoder signal); if it is vintage and exposes nothing, with a camera pointed at the panel that reads the counter from the image. Every start-up, every stop and every unit is logged to the second, without touching the machine. And when the line stops, the operator simply says: “iLEAN, log this: film jam on the wrapper” — the cause is linked to the exact stop, by voice, without putting down what they are holding.
• Tracer agent — consolidates real OEE continuously, 24/7, cross-referenced by line, SKU and shift: availability, performance and quality calculated on stops that were measured, not remembered. Every morning it flags the 3 stops that cost the most OEE points, with machine, cause and SKU — the top 3 to act on today, not next month.
• Writer agent — drafts the shift report and the committee summary by itself from the captured data: stops, classified causes, open “no cause” items still to close. The shift lead validates before anything is consolidated — without a signature, the data does not go in. The twenty minutes of paperwork per shift disappear; validation takes two.

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

See the full IRIS architecture →

Before and after

OEE rebuilt in a spreadsheet vs. OEE measured live with iLEAN

Aspect	Report from memory + weekly spreadsheet	With iLEAN Connect + Tracer + Writer
Source of the data	Report filled in from memory at the end of the shift	Line counter to the second (integration or camera on the panel)
Micro-stops (2-5 min)	Invisible — never logged	All logged, with exact duration and cause
Stop causes	Generic: “breakdown”, “adjustment”	Classified by voice by whoever was there: machine + specific cause
OEE frequency	Weekly, rebuilt on Friday	Live 24/7, by line, SKU and shift
Prioritizing improvements	By intuition and by whichever breakdown is best remembered	Top 3 stops by OEE points lost, flagged every morning
Admin load on the operator	15-20 minutes of paperwork per shift	Ten seconds of voice per stop — the report writes itself and is validated in two minutes

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

• Food packing plant with 2-8 lines and 2-3 shifts, OEE calculated today in a weekly spreadsheet from end-of-shift reports.
• Connect deployed on the counters (integration or camera on the panel) + Tracer/Writer agents. First value expected within a few weeks on one pilot line — the one that hurts most.
• OEE goes from a rebuilt weekly figure to live data. As micro-stops that were never logged come to the surface and the identified top 3 is attacked, the expected recovery is several OEE points — an estimate to be validated with your data.
• Indicative payback between 4 and 9 months depending on the number of lines, shifts and the value of an OEE point in your plant — an estimate to be validated in the diagnostic.

And production's reasonable doubt

“What if the AI classifies the cause of a stop wrongly?” — hallucination is a problem of free generation, not of anchored tasks. Linking the phrase “film jam on the wrapper” to the stop the counter has just registered and classifying it against the plant's catalog of causes is exactly an anchored task: the AI recontextualizes one piece of data (the operator's phrase) using another (the stop measured to the second) against a closed list. In this kind of task, the best models brought error below 1.5% ^[1]. And even so, the person validates the report — without a signature, the data is not consolidated. 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.