The shift quality sheet stops being retyped by hand

The problem

Two parallel worlds, and decisions get made according to who is making them.

Two parallel worlds coexist in every plant. The real spreadsheet, holding the shift's data, current and reliable. And the official ERP, with data transcribed days or weeks later, carrying the copy errors you'd expect from repetitive work. Decisions get made looking at one or the other depending on who makes them, producing arguments that aren't about the operation but about which number is the right one. Someone spends half a day a month per plant on that transcription. That's a qualified person's time spent copying cells, and it introduces latency exactly where it's least affordable: quality parameters are what let you react to a deviation while there's still time. In a multi-plant group there's a worse consequence than lost time. Each lab has its own sheet, with its own columns and sometimes its own units. Quality parameters can't be aggregated across plants even when they measure exactly the same thing, so management can't compare quality between two factories in the same group.

• The real spreadsheet, where the shift's data lives, current and reliable.
• The official ERP, with figures keyed in days or weeks later and the copying errors you would expect from repetitive work.
• Half a day a month per plant of a qualified person copying cells — and latency exactly where it hurts: quality parameters are what let you react while you still can.
• Every lab has its own sheet, with its own columns and sometimes its own units: management cannot compare the quality of two factories in the same group.

How it fits the IRIS system

Connect in listen-inward mode — it watches the folder, not the people.

Connect in "listen inward" mode. The flow:

The decisive step is normalization: the plant does not change its sheet, the translation into the group's common model happens on iLEAN's side.

• iLEAN observes the shared folder — the folder, not the people — and detects when the file has been saved.
• It reads the delta and parses it with a grounded language model, tolerant of each plant having its own sheet layout.
• It normalises to the group's common model. This is the decisive step: the plant doesn't change its sheet, normalisation happens on the iLEAN side.
• The manager validates on screen (early human verification).
• The data is inserted into the ERP through its API, and becomes aggregable with the other plants'.

See the full IRIS architecture →

Before and after

An isolated sheet vs. an integrated sheet

Aspect	Today	With iLEAN Connect
Spreadsheet → ERP lag	Days or weeks	Second zero
Keying	Half a day a month per plant	None
Copying errors	Those of repetitive work	There is no copying
Sheet format	One per lab	Theirs, normalized
Comparing quality across plants	Not possible	Aggregable from the first shift
ERP licenses on the module	4-5 per plant	1, the validator's

weeks of lag → second zero. Half a day a month per plant transcribing → zero. Quality parameters not comparable across plants → aggregable from the first shift.

Impact estimate

Impact estimate — to validate against 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.

• One quality sheet per shift and per plant: cupping and moisture in coffee, pH, Brix, vacuum and drained weight in canning, moisture and sizing in culinary.
• CFO angle: the ERP is licensed per active user. If 4-5 lab people log in only to enter results, with Connect one person logs in.
• Typical reduction of 50 to 80% in quality and lab module licenses, recurring and multiplied by the number of plants.
• Indicative payback of 4 to 9 months per plant, plus the annual saving. To be checked against the real ERP invoice.

— *CFO angle, direct licence saving*: ERP is licensed per active user. If each plant has 4-5 lab people who log in purely to enter quality results, that's 4-5 licences per plant; with Connect only the manager who validates logs in → 1 licence. Typical reduction of 50-80% in quality and lab module licences, recurring and multiplied by the number of plants. Estimated payback 4-9 months per plant plus the annual saving. *Estimate to validate* against the real ERP invoice.

And the fair question from the production manager

“What if I change a column on the sheet?” — nothing breaks, and that is the reason for parsing with a language model instead of a fixed cell mapping. A rigid parser breaks the first time somebody inserts a row; here the structure is interpreted. If a change leaves something ambiguous, it is flagged in the validation rather than entering wrong.

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