Paper picking sheet, digital data at zero-second latency

The problem

What actually shipped is written on a sheet sitting in a tray.

The picker marks it by hand while walking the aisles with the forklift, then leaves it in a tray until someone transcribes it at shift close — if there's time that same day. If there's a mismatch between what was picked and what the order called for, it usually surfaces when the client calls to complain, not before the truck leaves.

• The WMS prints the picking sheet with reference, quantity, location and client, but the picker corrects it by hand on the move: a short location, a substituted reference, a quantity split across two pallets because the first one was full.
• Those handwritten marks are the real record of what left the warehouse, and they sit in a dock tray until someone transcribes them at shift close, if there is time that same day. On a busy Friday, the tray waits until Monday.
• In a multi-client warehouse the same aisle holds goods for three or four clients with similar packaging, so a reference picked from the neighboring location is easy to make and hard to spot by eye once the pallet is wrapped.
• Between the pick and the transcription the error travels unseen: it usually surfaces when the client calls to complain, not before the truck leaves the dock, and by then the only fix is a return and a second delivery.

How it fits the IRIS system

Connect in photo mode — the picker keeps marking the sheet exactly as before.

The picker takes a photo of the completed sheet before wrapping the pallet. An anchored LLM extracts reference, quantity, location and client from every line, and stores it centrally at zero-second latency. No new screen to learn, no process to change: they keep marking the sheet exactly as before. The only difference is that, alongside the paper, the digital record now exists instantly.

It asks the picker for one photo before wrapping the pallet, nothing else. No new screen, no scanner workflow to learn, no change to how the WMS prints the sheet. The handwritten corrections stop dying in a tray and become lines the system can check against the order while the pallet is still at the dock. That is why it is usually the first case a warehouse approves: it changes the data, not the habit.

See the full IRIS architecture →

Before and after

The paper picking sheet versus the photographed picking sheet

Aspect	Today	With iLEAN Connect
When the pick exists in the system	Hours later, at shift close	Zero-second latency
Handwritten corrections	Lost unless someone types them	Extracted line by line
Wrong location or reference	Found in the client's complaint	Visible before the truck leaves
Admin transcription time	Every order, every shift	Recovered
Picks per client and per warehouse	Rebuilt from the tray	Queryable by order
The picker's routine	—	Unchanged: one photo

Before: the picking sheet sits in a dock tray for hours until manual transcription. After: the picking data exists in the central system at zero-second latency, the moment the photo is taken.

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 4-9 months, depending on the number of orders picked per day and on how many people transcribe sheets today.
• Administrative transcription time recovered at every shift close, which in a warehouse running several hundred order lines a day is a large part of one person's afternoon.
• From hours of latency to zero-second capture per order: the pick exists in central memory the moment the photo is taken.
• And the hours-long window in which a picking error travels unseen is closed before loading, which is where a mistake still costs minutes instead of a return.

Payback 4-9 months · from hours of latency to zero-second capture per order. It recovers administrative transcription time and closes the hours-long window in which a picking error travels unseen.

And the fair question from the production manager

“What if it misreads a handwritten quantity?” — reading fields from a sheet with a known layout is an anchored task, where the best models drop below 1.5% error [1]. And there is a second net: every line is checked against the order in the WMS, and any mismatch or low-confidence value goes to the dock tablet for a person to confirm before it counts. Every correction made there is stored, so the model learns the handwriting habits of your own pickers.

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