From the setup paperwork to the central system, at second zero

The problem

Paper is not the problem. The hours it takes to reach the system are.

The production order leaves the business system with the part number and the schedule, but everything that actually happens at start-up is written on top of it by hand: the heat number of the coil being consumed, the code of the mounted die set, the stroke count, the dimensions of the approved first piece and any authorised deviation. That sheet can sit four to eight hours without being digitised, or a whole shift. Meanwhile the lot exists physically — it is coming off the press — but it exists in no queryable system. The cost surfaces weeks later. When the OEM returns a cracked wheel and asks for traceability back to the steel heat, someone has to rebuild it by hand from a folder. The same happens when an auditor asks for the first-piece records for a given period.

• Those three documents — order, first-piece sheet and die tag — are the real traceability of the lot. The business system only holds half of it: the part number and the schedule.
• For four to eight hours, or a whole shift, the lot exists at the press but exists nowhere you can query.
• The cost does not show up that day. It shows up weeks later, when the OEM reports a cracked wheel and asks for traceability back to the steel heat.
• And because the answer sits in a folder in the office, rebuilding it takes a quality technician hours — with the customer waiting.

How it fits the IRIS system

Connect in photo mode — the setter keeps filling in exactly the same paperwork.

Connect in photo-to-analogue mode.

The habit change is zero, and that is not a detail: it is the reason the case is still alive in month six, which is where anything that asks a setter to learn a new app at the press dies.

• The setter photographs the order, the first-piece sheet and the tooling tag with his phone or industrial tablet. Zero change of habit.
• A grounded model recognises the structure of each document and extracts the fields: part number, heat number, die code, first-piece dimensions, press settings and the approver's signature.
• Records are structured in central memory and matched against the live order in the business system.
• The image of the original document stays attached as evidence, so any later query shows the real sheet rather than a transcription.

See the full IRIS architecture →

Before and after

Setup on paper versus setup captured

Aspect	Today, on paper	With iLEAN Connect
When the lot exists in the system	4-8 h later, or at shift close	Second zero
Steel heat of the coil	Rebuilt from paper, if legible	Tied to the lot from setup
First-piece dimensions	Sheet filed in a folder	History queryable by part number and die set
Answering a field claim	Hours of a quality technician	One query by lot
An illegible or mislaid sheet	The lot data is lost	The photo of the real document stands as evidence
Setter's routine	—	Unchanged: two photos at setup

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.

• Estimated payback 4-9 months, depending on how many changeovers you run per week.
• From 4-8 h of latency between the data and the system, down to second zero.
• Direct recovery of transcription time from the production supervisor and the quality technician.
• First-piece documentation gap closed in customer audits — the large saving sits here, and it shows up the day there is a claim.

estimated payback 4-9 months, depending on the number of changeovers per week. It comes from transcription time recovered from the production supervisor and the quality technician, and from the documentation gap removed in customer audits. *Estimate to validate.*

And the fair question from the production manager

«What if it misreads the heat number and ties the coil to the wrong lot?» — reading fields from a document with a known structure is an anchored task, where the best models drop below 1.5% error [1]. But the real answer is the other one: the extracted summary goes to the tablet and the supervisor confirms or corrects it before anything crosses into the system. Nothing enters on its own.

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