The shift starts on a sheet of paper and stays there for eight hours

The problem

The sheet that defines the whole shift ends up in an office folder.

With around a hundred active variants, shift startup and intermediate changes generate several handwritten sheets a day per line: startup sheet, variant changeover checklist, first-part record required by the control plan. All of it travels to a folder in the office. The consequence is not administrative. When the customer opens a claim about a specific part three days later, the plant needs to prove which tooling and which parameters it was made with. If the paper is not digitized and linked to the order, there is no possible proof: the charge is accepted without being able to argue it technically. The classic attempt to fix it — asking the line lead to also key it into a terminal — always fails for the same reason: it competes with cycle time and with the attention the line demands at startup, the busiest moment of the shift.

• Startup sheet — which variant is coming in, which tooling has been fitted and what parameters the press starts with.
• Variant changeover checklist — the proof that what had to be checked was checked before releasing the line.
• First-part record — required by the control plan, and the evidence asked for months later in a claim.

How it fits the IRIS system

Connect in photo-the-analog mode — the line lead's phone digitizes the paper without touching the line.

Connect in photo-the-analog mode.

Two photos at startup. Connect extracts variant, tooling, parameters and incidents, and links them to the active shift and order. The line lead keeps filling in the paper exactly as always.

• The line lead photographs the startup sheet and the variant changeover checklist with a phone or an industrial tablet. They keep filling them in by hand exactly as always.
• An anchored language model extracts the structured fields: variant, fitted tooling, press parameters, station operators, open incidents and first-part result.
• Connect links that record to the shift, the order and the variant active in the production system.
• The summary appears on the line tablet for human validation in two taps before crossing into the central system.
• From there, any query by shift, by variant or by tooling is instant.

See the full IRIS architecture →

Before and after

Startup retyped after the fact vs. startup captured at second zero

Aspect	Startup on a paper sheet	With iLEAN Connect
Data latency	4-8 h, or never digitized	Second zero
Link to the order	None: the sheet goes to a folder	Anchored to shift, order and variant
Facing a claim	“I think we started with tooling B”	Queryable, signed record
Press parameters	Handwritten, uncorrelated	Linked to that shift's parts
Load on the line lead	Fill in the paper and also key it in	Fill in the paper and take a photo
Query by tooling or variant	Search the office folder	Instant

From four to eight hours of latency — or never being digitized — to second zero. From "I think we started with tooling B" to a queryable, signed record. From a claim accepted for lack of evidence to a claim bounded with process data.

Impact estimate

Impact estimate for your plant — to validate against 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.

• Plant with around a hundred active variants and several changes per shift and per line.
• Connect pilot on one line, with the phone or tablet the line lead already carries. No construction work and no downtime.
• Indicative payback between 4 and 9 months, depending on variant changes per shift.
• The hard lever is not the retyping time: it is being able to argue a claim technically instead of accepting the charge for lack of evidence.

Estimated payback 4 to 9 months depending on the number of variant changes per shift, combining recovered retyping time and the ability to defend a claim technically. *Estimate to validate*.

And the fair question from the production manager

“What if the model misreads a parameter on the handwritten sheet?” — hallucination is a problem of <b>free</b> generation, not of <b>anchored</b> tasks. Here the AI moves a specific data point from one medium to another, and in that kind of task the best models dropped below <b>1.5%</b> error ^{<a href="#source-openai">[1]</a>}. And nothing enters the system without the supervisor validating it on the line tablet.

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