Paperless Work Permit and demolition report

The problem

The site is a cost center that moves, and its only record is on paper.

Every demolition site is a mobile cost center with no fixed office. The Work Permit and the daily report are the only source of truth for "which machine did what, where and when," and today they depend on someone physically collecting them from the site office at day's or week's end. When the paper gets lost or is transcribed late, the technical office reconstructs what happened on each work zone from memory, with all the errors and gaps that implies.

• Every work zone opens with a signed Work Permit and closes the day with a handwritten report: machine assigned, attachment fitted, operator, incidents, progress against the demolition plan.
• That paper is the only source of truth for which machine did what, where and when — and it lives in the site office or in the foreman's pocket until somebody physically goes out to collect it at day's or week's end.
• The site has no fixed office and the crew rotates: the person who could explain a line in that report may already be on another site, with another subcontractor, by the time the technical office gets around to reading it.
• When a sheet is lost or typed up a week late, progress per zone is reconstructed from memory, and the certification that comes out of it inherits every one of those gaps — in front of a developer who is paying against exactly that figure.

How it fits the IRIS system

Connect in photo mode — the foreman keeps filling in the same paperwork.

Connect's photo mode: the foreman photographs the permit and the report with their phone, no habit change required. A grounded LLM recognizes the document's structure and extracts the key fields — work zone, assigned machine, attachment, operator, issues, progress — and stores them in the site's central memory.

It is the least intrusive of the twelve and the one the crew accepts fastest, because nothing is taken away from them: the same permit, the same pen, the same signature, plus one photo. Nobody has to learn an app halfway through a shift and nothing depends on a tablet surviving the dust. That is usually why it is the case that gets approved first, and why it is still running six months later.

See the full IRIS architecture →

Before and after

Today's site paperwork versus captured site paperwork

Aspect	Today, on paper	With iLEAN Connect
Where the permit lives	The site office or a pocket	In the system from minute one
Capture latency	Until someone drives out to collect it	Zero
Which machine worked which zone	Rebuilt from memory	Recorded zone by zone
Progress against the plan	Weekly and approximate	Daily, with its evidence
A report that gets lost	Nothing survives it	The photo is the record
The foreman's routine	—	Unchanged: one photo

Before: paper sitting in the site office until the week closes, memory-based reconstruction. After: information is available in the central system at zero latency, site by site.

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 how many sites you run concurrently.
• Technical-office transcription time recovered, which today piles up at week's end and gets done against the clock, which is when the mistakes go in.
• From paper sitting in the site office to information available at zero latency, site by site.
• And the demolition report stops being a sheet of paper that has to survive a dusty site, a van and a week in a folder before it counts for anything.

Estimated payback of 4-9 months depending on how many sites run concurrently, counting recovered technical-office transcription time. Estimate to validate.

And the fair question from the production manager

"What if it misreads the zone or the machine on the permit?" — reading fields from a document with a known structure, printed by your own technical office, is an anchored task, where the best models drop below 1.5% error [1]. And there is a second net: the extracted summary reaches the site-office tablet and the manager confirms or corrects it before anything crosses into the central system.

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