Nothing enters the ERP without your confirmation

The problem

The fear that blocks the whole project is not technical.

The fear that AI will enter data on its own and contaminate the ERP master is the main cultural blocker, especially when the data affects batch, vintage or denomination of origin.

• The main objection in a winery is not whether AI can read a log. It is the fear that it will write into the ERP on its own and contaminate the master data everyone relies on.
• That fear is reasonable when the data is batch, vintage or denomination of origin: one wrong field can end up printed on thousands of back-labels.
• The usual outcome is the worst of both worlds: either unvalidated data enters, or nothing is captured at all out of distrust and the paper keeps winning.
• And when the regulatory council asks who entered a figure and on what basis, "the system" is not an acceptable answer. The answer has to be a name, a time and what that person saw.

How it fits the IRIS system

Connect with early human verification — a two-second summary and two taps.

Connect early human-verification mode. Every capture (photo, voice, parsed email) reaches an industrial tablet as a two-second visual summary. The supervisor confirms with one tap, or corrects with another.

This piece earns nothing on its own, and it is the one that lets everything else be signed off. No winery with origin requirements approves a system that writes batch and vintage data without a named person behind every single entry. With it, the cellar team stops seeing AI as something that writes behind their backs and starts seeing it as something that prepares their work for a signature.

See the full IRIS architecture →

Before and after

Unchecked capture versus capture confirmed on the bottling line

Aspect	Today	With iLEAN
Data entering the ERP	Typed late, or entered unchecked	Confirmed by a person in seconds
Who answers for each figure	Unclear	The supervisor who confirmed it
A wrong vintage or batch	Found on the back-label	Corrected on the tablet
Captures from photo, voice and email	Each one handled differently	One summary format, one gesture
The ERP master	At risk	Clean
The team's trust in AI	Low, so nothing gets captured	Built on control, not on promises

unvalidated data contaminating the ERP, or data not captured out of distrust → data captured and validated with seconds of latency, clean master data.

Impact estimate

Why this piece has no payback of its own.

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.

• No standalone payback · enabling piece: we do not put a number on it, because its value is letting the other eleven cases be approved.
• In a winery with denomination-of-origin traceability requirements, no other piece of the matrix gets signed off without it.
• It turns unvalidated data, or data never captured out of distrust, into data captured and validated with seconds of latency.
• And it keeps the ERP master clean, which is what every later audit, every blend decision and every back-label depend on.

enabling piece — without it, no other piece of the matrix gets signed off in a winery with denomination-of-origin traceability requirements. Strategic value, not directly monetizable.

And the fair question from the production manager

"If a person confirms everything anyway, what is the AI doing?" — the work it removes is assembling, not deciding. Extracting the fields is an anchored task, where the best models drop below 1.5% error [1]; the supervisor only looks at a summary that is already built and confirms or corrects it. Every correction is stored with its author, so the system learns exactly where it got it wrong.

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