The human gate that makes artificial intelligence auditable

The problem

Automatic capture, by itself, is not enough in a certified environment.

A system that introduces data without an identified signature creates two risks at once: the technical risk of contaminating the genealogy with an erroneous reading, and the regulatory risk of not being able to demonstrate who validated what:

• The data demanding that validation is very specific — the reel batch mounted on each feeder, the first-article approval result, an out-of-range measurement, the assigned serial number, the loaded engineering revision.
• Any of them badly captured breaks unit traceability — which in an automotive panel is what allows scoping a campaign to the affected units instead of the entire production.
• But the human gate cannot cost line time — if validating costs more than typing, nobody validates. And then you are back at the starting point.

Hence the solution is not choosing between automating and controlling, but separating the capture — which can be automatic — from the validation, which must be human and identified but can take seconds.

How it fits the IRIS system

Early human verification — Connect's fourth mode, which applies to the other three.

It is not just another case: it is the doctrinal mode that makes all the others signable. Everything captured by photo, by panel, by file or by voice passes through here before existing in the central system.

An industrial tablet at the line — startup station and final assembly station. An at-a-glance visual summary of what was captured, whether it came from a photo, a dictation, a panel reading or a laboratory file. Two taps: confirm, or correct the specific field that does not add up. Only then does the data cross, with signature and timestamp.

How early human verification operates on the line:

• At the line, at two stations — startup and final assembly, which is where data is born and where it closes. A validation that demands walking to an office piles up and ends up done in bulk, without looking.
• An at-a-glance visual summary — not a form. The interface is designed for reviewing what was already extracted, not for transcribing.
• Two taps: confirm or correct the specific field — the whole record is not rewritten, only what does not add up.
• Identified signature and timestamp — which is exactly what an auditor asks for and what today goes unrecorded when the data is typed in.
• What gets corrected improves recognition — so the number of corrections drops with use: validation gets cheaper over time, not more expensive.

See the full IRIS architecture →

Before and after

Unsigned capture vs. early human verification

Aspect	Data entering directly	With iLEAN + validation tablet
Entry into the central system	Unacceptable in a certified environment	Only with an identified signature
Traceability of who validated	Does not exist	Signature and timestamp per data point
Product genealogy	Risk of permanent contamination	Protected
Technician's work	Complete manual transcription	Review of what was already extracted
Cost of validating	—	Seconds, and decreasing with use
Before a process auditor	A data point with no backing	Data with origin, confidence and signature

Value of this piece

Value of this piece — enabling, with no isolated return.

This case is deliberately presented without its own payback: its value is that it makes the others auditable.

• Certified electronic panel plant with unit traceability and process data that cannot enter the system without a signature.
• Pilot of the tablets at the startup station and at final assembly, together with the first capture case. It is not deployed on its own.
• It has no isolated return: it is the enabler that makes the photo, panel, file and voice capture cases auditable. Without this gate, automatic capture does not pass a process audit. Its ROI is accounted together with those cases'.
• And there is an effect that improves over time: what gets corrected feeds the improvement of the recognition itself, so the number of corrections — and therefore the cost of validating — drops with use.
• Direct regulatory benefit: every process data point keeps its origin, its confidence level and the person who validated it, with a timestamp.

And the fair question from the line manager

"If everything must be validated, don't we lose rate?" — no, because validating and transcribing do not cost the same: nothing is entered here, an already-extracted summary is reviewed and confirmed. And it is that validation which lets the AI work on anchored tasks safely: even with an error below 1.5% ^[1], that remainder never reaches the master data or the genealogy. On top of that the system learns from every correction, so the number of interventions drops with use.

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