First-piece release with early human verification

The problem

The release is signed fast because it asks for data that already exists.

First-piece release and heat start-up decide whether the shift comes out good or goes to scrap. Today they are signed on paper or on a screen that forces you to retype data that already exists in another system. The result is fast signatures, fields filled in out of habit, and releases that prove very little when a problem appears. The alternative many AI projects offer — letting the system write on its own — is worse in this context: it contaminates the master data with unverified figures and breaks the trust of customer and auditor exactly where it matters most.

• First-piece release and heat start-up decide whether the shift comes out good or goes to scrap. Today they are signed on paper or on a screen that forces the operator to retype what the heat sheet and the lab already hold.
• The result is fast signatures, fields filled in out of habit and releases that prove very little when a caliper comes back from the customer.
• The alternative many AI projects sell — letting the system write on its own — is worse here: it contaminates the master data with unverified figures, and a contaminated master is worse than no data at all.
• And it breaks the trust of the customer and the IATF auditor exactly where it matters most: on the record that says who released a safety part and on what.

How it fits the IRIS system

Connect with early human verification — the gate everything else passes through.

Connect in early-human-verification mode, the gate everything else passes through:

This is the piece that makes the other eleven approvable. Everything captured by photo, voice, folder parser, panel reading or receiving lands in a validation queue, and the tablet shows it next to its evidence. The operator confirms or corrects in under thirty seconds; only then does the figure cross into the ERP or the MES. Without this gate, no quality manager signs off on AI writing into the system a release depends on. With it, authorship moves from whoever had time to type to whoever actually looked at the disc, and that is the name the customer will ask for.

• Everything captured by the other cases — photo, voice, folder parser, panel reading, receiving — lands in a validation queue prioritized by criticality.
• The industrial tablet shows the extracted data next to the original evidence: heat number, chemistry, mounted pattern, first-piece measurements.
• The operator confirms or corrects in under thirty seconds. Only then does the data cross into the central system.
• The correction is kept as a signal: the system learns from the craft of the person on the line.

See the full IRIS architecture →

Before and after

Today's release versus the signed release

Aspect	Today	With iLEAN
Time per release	3-6 minutes of typing	Under 30 seconds of reviewing
What the signature proves	That somebody signed	Which chemistry, pattern and measurements they saw
Heat number on the release	Retyped from the sheet	Already there, next to its photo
Who writes to the master	The AI, or whoever types	A person, confirming
Capture quality	Unknown	A measured correction rate
When the auditor asks	A checkbox	Evidence and a name

from 3-6 minutes of typing per release to under 30 seconds of reviewing · from data with no evidence to every figure with its source photo and signature · from no visibility on capture quality to a measured correction rate.

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 4-8 months.
• A 60-80% reduction in the time spent on releases and recording, at the line, without walking to a terminal.
• Above all, zero unsigned data in the central system: every figure has its source evidence and the name of who confirmed it.
• And every correction is kept as a signal, so capture improves where the person on the line says it fails.

estimated payback 4 to 8 months, with a 60-80% reduction in time spent on releases and recording and, above all, zero unsigned data in the central system. *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 the operator walks to a terminal, finds the order and retypes the heat, the chemistry and the measurements. With the gate it all arrives assembled from sources where the best models drop below 1.5% error on anchored extraction [1], and the operator only confirms or corrects. Every correction is stored with its author, which is what an auditor wants to see and what the model learns from.

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