Nothing enters the system without a human having signed for it

The problem

The brake is not technical: it is trust, and it is reasonable.

The real blocker to digitising a shop floor is almost never technical: it is trust. The objection is always the same — if the AI enters data on its own, somebody has to clean the ERP master afterwards — and it is a reasonable objection. A contaminated master costs more to fix than the problem it was meant to solve. But the alternative does not work either: asking the operator to type. On a blow molding line at real speed, any manual entry is done late, done badly or not done at all. In a PET plant the risk is highly concentrated. The four critical fields — weight, mold and cavities, resin batch, recycled content — decide whether the container is conforming and whether the recycled content declaration handed to the customer is true. An error there is not a record-keeping error: it is a product conformity problem.

• The objection is always the same — if the AI writes data on its own, someone has to clean up the ERP master afterwards — and it is reasonable: a contaminated master costs more to fix than the problem it was meant to solve.
• But the alternative does not work either: on a blow molding line at real cadence, any manual capture happens late, happens badly or does not happen.
• In a PET plant the risk is highly concentrated in four data points: preform and bottle weight, mold and active cavities, resin lot and the blend's recycled content.
• Those four underpin quality, traceability and the material balance. An error in any of them contaminates all three at once.

How it fits the IRIS system

Connect with early human verification — a summary, not a form.

Connect, early human verification mode. An industrial tablet at the blow molder and the injection machine, or at the supervisor's station. Step by step:

The interface is designed for validating, not for entering data. That difference is what makes it get used: asking an operator to fill in a form at real cadence is asking them not to do it.

• Capture by any route. Photo of the set-up sheet, dryer panel reading, voice note, parsed email or inline weight measurement.
• Visual summary. What was captured appears on the tablet as a summary, not a form. The interface is designed for validating, not for data entry.
• Two taps. The operator confirms with one tap, or corrects the specific field before it goes in.
• Crossing to the master. Only after the human signature does the data cross into central memory and, where applicable, into the ERP via API.
• Signature traceability. Who validated what and when is recorded, feeding the audit evidence pack directly.

See the full IRIS architecture →

Before and after

The three ways of resolving it, compared

Aspect	Autonomous AI	Operator types it
ERP master protected	No	Yes
Actually done at cadence	—	No: late, badly or never
Operator time	Zero	Minutes
Training required	—	High
The four critical data points	Unsigned	Half captured
Approvable by the board	No	Yes, but unused

Impact estimate

An enabling piece — no payback of its own, and stated as such.

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.

• This case has no payback of its own worth calculating: it is the enabling piece of the whole matrix.
• Without early human verification, the other eleven cases do not get signed off by the board.
• It is what turns "the AI writes data" into "the AI prepares and a person signs", which is a different conversation.
• It removes both unvalidated data reaching the ERP and data never captured out of distrust of the AI.

This case is the enabling piece of the whole matrix and has no payback of its own worth calculating. Without early human verification the other eleven cases do not clear a steering committee: it is what turns "the AI enters data" into "the AI prepares and the person signs". *Strategic value, not directly monetisable.*

And the fair question from the production manager

«If a person signs in the end, where is the saving?» — the work is not in deciding, it is in assembling. Today somebody would have to weigh, check the panel, look up the resin lot and type it. With the verification gate it all arrives assembled in a two-second summary and they only confirm or correct the specific field. And every correction is stored: the system learns where its extraction fails.

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