The AI doesn't touch your ERP without your sign-off

The problem

At an IATF 16949 plant, a data point nobody signed is a finding waiting to happen.

the fear that AI will "contaminate the ERP master" is the most common cultural pushback to digitalization at an IATF 16949-certified plant. But operators don't want to type either.

• The most common pushback to digitalization in a certified tube plant is cultural, not technical: the fear that AI will contaminate the ERP master with a wrong heat number or weld result.
• At the same time, the welder and the inspector do not want to type: their job is the seam and the tube, not a terminal. So data is either entered unvalidated, or not captured at all out of distrust.
• Both outcomes break the same thing: an auditor or an OEM asking who confirmed a weld result and on what evidence.
• Without an answer to that fear, none of the other eleven cases gets signed off by quality, however good its business case looks on paper.

How it fits the IRIS system

Connect with early human verification — the AI proposes, the weld cell signs.

Connect's early human-verification mode. Every capture reaches the tablet as a visual summary before crossing into the master. UI designed for 2-second validation, not data entry.

It is the piece that decides whether the rest gets approved. Every capture reaches the tablet as a two-second visual summary, and nothing crosses into the master without a name attached to it. And every correction the welder makes is stored, so the system learns where its extraction fails.

See the full IRIS architecture →

Before and after

Unvalidated capture versus capture signed at the weld cell

Aspect	Today	With iLEAN
Entry into the ERP master	Typed later, or pushed unchecked	Only after a human confirmation
Welder's or inspector's effort	Data entry at a terminal	Two taps at the cell
A wrong heat number in a capture	Travels into the master	Corrected before it crosses
Answering an IATF auditor	Whoever typed it	Who confirmed, when, and what they saw
Quality's stance on AI	Blocks the project	Signs off, with the gate in place
Every correction made	Lost	Stored and fed back to the model

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

Impact estimate

Estimated impact — to validate with your own 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.

• No standalone payback: this is an enabling piece for the whole matrix, with strategic value rather than a directly monetizable saving.
• Without early human verification, none of the other pieces gets signed off at an audited plant.
• The ERP master stays clean and traceable, with data captured and validated within seconds. The alternative is a master that is either late or contaminated.
• And the question an auditor always asks — who confirmed this — gets a name, a timestamp and the exact summary that person saw when they confirmed.

the enabling piece for the whole iLEAN matrix — without early human verification, none of the other pieces get signed off at an audited plant. *Strategic value, not directly monetizable*.

And the fair question from the production manager

“Two taps at every capture, every shift — won't the welders just tap through?” — the summary is built to be read in two seconds, and what it asks is narrow: extracting a program, a heat number or a weld result is an anchored task, where the best models drop below 1.5% error [1], so most taps are confirmations. Fields the model is unsure of are highlighted, not hidden, so attention goes where it is needed. And every correction is stored with its author, which is how the system learns where its extraction fails.

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