The human gate sits at the bench, not in an office

The problem

Badly captured data is worse than missing data.

In a mold shop, badly captured data is worse than missing data. If the system believes a detail is finished when it is not, the entire schedule is built on that belief and the date is defended with false data until the mold reaches the bench and does not fit. This is exactly what happens when an "autonomous" AI agent is deployed that writes into the system unsigned: it ends up contaminating project history, and from then on nobody trusts any figure. The only person able to say whether something is genuinely done is the one holding it. The problem is that this person cannot afford an approval flow that steals their time.

• If the system believes a detail is finished when it is not, the whole schedule is built on that belief and the date is defended with false data until the mold reaches the bench and does not fit.
• That is exactly what an "autonomous" agent writing into the system unsigned produces: it contaminates the project history, and from then on nobody trusts any figure.
• The only person able to say whether something is genuinely done is the one holding it. That person cannot afford an approval flow that steals their time.
• So today the choice is between data nobody signed and approvals that wait until end of day.

How it fits the IRIS system

Connect in early human verification mode — two taps at the bench, not a form in an office.

Connect in early human verification mode — the safety ring that defines the IRIS architecture.

It is the safety ring that defines the IRIS architecture. Every capture from the other cases appears as a card on the bench tablet with the doubtful field highlighted; the fitter confirms or corrects, and only then does the data cross into central memory stamped with who validated it and when. Seconds of his time for a history management can commit a date on. Nothing to search for, no forms, no walk to a terminal: the card comes to the bench where the part is.

• The bench tablet shows cards with what iLEAN has just captured, with the doubtful field highlighted.
• Two taps: confirm or correct. No forms, nothing to search for.
• The correction feeds the system too, which learns which fields it gets wrong on each document type and improves the next capture.
• Only then does the data cross into central memory, stamped with who validated it and when.

The human gate costs seconds, not approvals that wait until end of day.

See the full IRIS architecture →

Before and after

Today's data versus the signed data

Aspect	Today	With iLEAN
Who says a detail is done	Whoever types it	Whoever holds it
Time it costs the fitter	A walk to a terminal, or nothing	Two taps
A wrong capture	Found weeks later, or never	Corrected on the spot
What the correction does	Nothing	Improves the next capture of that document
Arguing whether the system is lying	Every morning	One version of the truth
Who writes to central memory	An unsigned agent	A person, confirming

from an unquantified share of unreliable project data → data signed by the person who knows. From corrections made weeks later (or never) → made on the spot. From arguing about whether the system is lying → a single version of the truth.

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, through planning quality rather than throughput.
• Fewer late replans, because the schedule is built on details genuinely closed.
• Fewer management hours reconciling what the system says with what the shop sees.
• And every figure in the project history has a name and a time behind it, which is what a customer asks for when the date is discussed.

estimated payback 4-9 months through planning quality: fewer late replans and fewer management hours reconciling what the system says with what the shop sees. *Estimate to validate*.

And the fair question from the production manager

"If a person signs anyway, what is the AI saving?" — the work is not in deciding, it is in assembling. Today somebody would have to leave the bench, find the mold and the detail in the system and type the result. Extracting the fields is an anchored task, where the best models drop below 1.5% error [1]; the fitter only confirms or corrects what arrives assembled, and every correction is stored so the system learns where it fails.

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