The PET lab reaches the ERP without a single retyped figure

The problem

The ERP returns an old snapshot of the plant's quality state.

Two parallel worlds coexist. The real spreadsheet, in a shared folder, holding the live data on which decisions are actually made. And the official ERP, with retyped, late data and copy-paste errors. The visible cost is time: roughly half a person-day a month keying into the ERP what was already written. The invisible cost is worse: throughout that gap, any query to the ERP returns an old snapshot of the plant's quality status. And when the bottler challenges a batch, traceability has to be rebuilt by hand across both worlds, copy-paste errors included. That is not quality work, it is documentary archaeology — and it is exactly where the gaps an auditor raises as non-conformities appear.

• The laboratory decides: intrinsic viscosity, residual acetaldehyde, resin moisture, weight, top load, environmental stress crack resistance, perpendicularity, thickness distribution, color and recycled content. Release or block comes out of that.
• And it works in a spreadsheet the ERP cannot see, with a gap that can reach thirty days.
• The visible cost is half a day a month of somebody typing into the ERP what was already written.
• The invisible one is worse: during that gap, any ERP query returns an old picture. And when a bottler raises a complaint, traceability has to be rebuilt by cross-referencing by hand.

How it fits the IRIS system

Connect in inward listening mode — the folder, not the person.

Connect, "listen inward" mode. iLEAN observes the shared folder — the folder, not the person — and reads the file when it is saved. Step by step:

Only the delta is read: what is new since the previous reading. The whole sheet is not reprocessed each time, which is what makes the case viable on files that have been growing for years.

• Change detection. When the sheet is saved, only the delta is read: what is new since the previous reading.
• Grounded parsing. A grounded language model interprets the sheet's structure — columns, units, merged cells, margin notes — without requiring the lab to change its format.
• Early human verification. The summary is presented to the quality manager, who validates or corrects.
• Posting to the ERP. After sign-off, results enter the ERP via API, tied to the corresponding batch.
• Cross-reference. Test results are linked to the resin batch and its certificate of analysis, to drying conditions read off the panel, and to the real start-up recipe.

The lab technician carries on working exactly as before.

See the full IRIS architecture →

Before and after

Today's laboratory versus the connected laboratory

Aspect	Today	With iLEAN Connect
Gap between real data and ERP	Up to 30 days	Second zero
Manual transcription	Half a day a month	Zero
Batch release	Decided on stale data	Decided on live data
A bottler complaint	Manual cross-referencing of sources	Traceability already cross-referenced
Sheet format	Has to be respected	Tolerated as is: columns, units, margin notes
Quality module licenses	Five	One, the supervisor who validates

Impact estimate

Estimated impact — to validate against your actual invoice.

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.

• 💰 CFO angle: the ERP bills per active user. If five laboratory people log in purely to capture test results, with Connect only the validating supervisor does: typical 50-80 % reduction in recurring licenses.
• Half a day a month of staff recovered.
• From up to 30 days of gap between the real figure and the official one, down to second zero.
• Batch release decided on live data instead of on an old picture.

CFO angle — direct ERP license savings. The ERP bills per active user. If five lab staff log into the ERP purely to key in test results, that is five licences; with Connect only the manager who validates needs to go in. Typical reduction of 50-80% in quality/lab module licences, recurring every year. Estimated payback 4-9 months, combining license savings with the half day a month of retyping. *Estimate to be validated* against last year's ERP invoice.

And the fair question from the production manager

«Different people have been changing that sheet for years — does it hold up?» — that is exactly the anticipated case: parsing does not depend on fixed positions, it interprets columns, units, merged cells and margin notes without requiring anyone to standardize anything. It is an anchored task, where the best models drop below 1.5% error [1], and the quality supervisor validates before the push.

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