NOTHING ENTERS WITHOUT A HUMAN SIGN-OFF

The problem

Nobody objects to the capture. They object to what writes into the master.

The fear that AI will 'enter data on its own and someone will have to clean up the ERP master' is the main cultural brake on digitization — especially sensitive when several brands with different recipe specs coexist in the same plant. But operators don't want to type either.

• The fear is specific and reasonable: that an AI enters data on its own and somebody spends the next month cleaning up the ERP master. It is the sentence that stops most digitization projects in this sector.
• It bites harder in a plant where several brands with different recipe specs share the same lines, because the master is where the allergen declarations live and a wrong field there reaches the sachet.
• The alternative does not work either: operators will not type, so anything not captured automatically is simply not captured at all, and the plant keeps running on paper.
• The result is a plant that has decided not to decide, keeping the paper and the distrust at the same time, and paying for both.

How it fits the IRIS system

Connect with early human verification — two taps before anything crosses.

Connect early human-verification mode: every Connect capture (photo, voice, parsed email) arrives at an industrial line-side tablet as a 2-second visual summary. The supervisor confirms with 2 taps (✓ enters the ERP / ✏️ correct) before the data crosses into the master.

The model never writes into the master. A person does, confirming a summary that took two seconds to read at the line where the data was produced. That single sentence is what gets the other eleven cases signed off by quality, and it is why this one is built first rather than last.

See the full IRIS architecture →

Before and after

Data that enters by transcription versus data that enters signed

Aspect	Today	With iLEAN Connect
Path into the ERP master	Typing, hours after the fact	One confirmation, seconds after
Allergen fields of the recipe	Exposed to a typing slip	Only confirmed values get in
Time asked of the supervisor	A trip to the office terminal	Two taps at the line
A capture the model got wrong	Would reach the master	Corrected before it crosses
Who the record names	The transcriber	The person who was at the line
What the supervisor has to learn	ERP navigation	Two buttons

Unvalidated data contaminating the ERP, or data not captured due to distrust → data captured and validated within seconds, clean master.

Impact estimate

Impact estimate — to be validated with quality and the ERP owner.

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: this is the enabling piece of the whole matrix. Without it, the other eleven cases do not get signed off in a plant running several brands.
• Its value is risk, not productivity, and putting it in months would be dishonest. What it buys is permission for the rest of the matrix.
• The master stays clean because nothing enters it that a person has not confirmed on screen, at the line, seconds after the data existed.
• And the data nobody dared capture, for fear of contaminating the ERP, starts being captured at last: it was never a technical gap, it was a permissions one.

The enabling piece of the whole iLEAN matrix: without early human verification, the other 11 pieces don't get signed off. Strategic value, not directly monetizable.

And the fair question from the production manager

"If a person confirms everything, what exactly have we automated?" — the typing, which is where the time and most of the errors were. Reading a two-second summary is not transcription. And on the extraction itself, an anchored task, the best models are already under 1.5% error [1], so the confirmation works as a net rather than as a second job. What used to take a walk to a terminal and five minutes of typing now takes two taps at the line.

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