Nothing enters the system until a person signs it

The problem

The most sensitive data in the plant is typed by whoever was not at the still.

A distillery's process data is sparse but very sensitive. One misread degree or one mis-noted volume shifts the mass balance of the whole batch, and with it the yield calculation and the declared stock. Manual transcription introduces errors that surface weeks later, when there is no reconstructing what happened or who wrote it down. Work-in-progress inventory stops being reliable, and with it any conversation about losses. And there is a deeper problem with AI systems that write without a signature: if the AI posts data straight into the company master, the master gets contaminated and nobody can tell which record has human authorship and which does not. In front of a certifying body or a contract-production client, that does not hold up.

• A distillery's process data is sparse but very sensitive. One misread degree of proof or one mis-noted volume of heads shifts the mass balance of the whole batch, and with it the yield calculation and the declared stock.
• Manual transcription introduces errors that surface weeks later, when there is no reconstructing what happened or who wrote it down. Work-in-progress inventory stops being reliable, and with it any conversation about losses.
• If an AI posts data straight into the company master, the master gets contaminated: nobody can tell which record has human authorship and which does not. That is not a technical nuance; it is what an auditor asks about first.
• In front of the certifying body or a contract-production client, a record nobody signed does not hold up.

How it fits the IRIS system

Connect with early human verification — two taps at the still.

Connect in early human verification mode. The flow:

It is the piece that guarantees the master is not contaminated and that every datum has a name attached. Without it, the other eleven cases on this page do not get approved, because nobody signs off on AI writing into the system the proof and the cut of a batch. The distiller confirms or corrects; iLEAN only assembles. It is also the piece the certifying body understands first, because it is the one that puts a name next to every number.

• An industrial tablet mounted at the still — resistant to humidity and wet hands — and a second one in the bottling area.
• It shows a visual summary, not a form: batch, fraction, liters, proof, cut time, with uncertain fields highlighted.
• Two taps: ✓ or correct the field that is wrong.
• Only after that signature does the data cross into central memory and from there into the ERP.

Every process record carries authorship: signed by whoever was standing at the still, with its timestamp. That turns the history into defensible evidence, and it is the requirement that makes every other case on this page defensible too.

See the full IRIS architecture →

Before and after

Today's record versus the signed record

Aspect	Today	With iLEAN
When the cut data enters	Days later	Seconds, at the still
Who appears as the author	Whoever typed it	Whoever stood at the still
A transcription error	Surfaces weeks later	Corrected on the spot
Uncertain fields	Invisible	Highlighted before signing
Work-in-progress inventory	Unreliable	Signed batch by batch
Who writes to the master	A person, typing	A person, confirming

From transcription errors surfacing weeks later to validation in seconds, right at the point where the data happens. From a process master with no authorship to every record signed by a person.

Impact estimate

Impact estimate — to be validated with 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.

• Estimated payback 4-9 months.
• Fewer administrative reworks: the error is corrected at the still, not chased weeks later by someone who was not there.
• A more reliable work-in-progress inventory, which is what every loss conversation depends on.
• And a process history that is defensible evidence in front of the certifying body, because every record carries who signed it and when.

Estimated payback 4-9 months from fewer administrative reworks and more reliable work-in-progress inventory. *Estimate to be validated.*

And the fair question from the production manager

"If a person signs in the end, what is the AI saving?" — the work is not in deciding, it is in assembling. Today somebody would walk to a terminal, find the batch and type fraction, liters, proof and time. Extracting those from the sheet or the still's readings is an anchored task, where the best models drop below 1.5% error [1], so the summary arrives assembled and the distiller only confirms or corrects. Every correction is stored and feeds back into the extraction, so the fields corrected most often become the ones flagged before asking.

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