AI 讀取送貨單的實作：真正難點在紙本
We taught AI to read delivery orders. The AI was the easy part.
作者在一家馬來西亞建築公司建立 AI 工作流程，自動讀取簽署送貨單、配對採購訂單或工地調撥，並把記錄輸入公司系統。首兩週處理超過 2,300 張送貨單，首日 68 張全數正確，並把約一名全職員工的資料輸入工作移至夜間處理。手寫、印章遮字及斜角拍攝等紙本問題是主要難點，流程會把不確定項目列出供員工核對。
Every construction site runs on paper. Concrete, steel, diesel, rebar: each delivery comes with a signed delivery order (DO), and someone in the office has to key every one into the system, match it to the purchase order, and file the scan. At a busy contractor that's hours of work every week, and it's the kind of work nobody notices until it falls behind.
Over the past few months I set up an AI workflow at a Malaysian construction company to take that job over. Here's what it does and what I learned.
How it works
- Site and store staff scan the signed DOs into one folder. That's the only manual step.
- Overnight, the AI reads each DO (supplier, items, quantities, dates) and matches it to the right purchase order or site transfer.
- It enters the record in the company's system with the signed scan attached.
- Anything it isn't sure about goes on a short list. In the morning, staff only check that list.
What I learned
The hard part isn't the AI, it's the paper. Handwriting, stamps over text, carbon copies, photos taken at an angle. Most of the work was teaching the workflow what "good enough to trust" looks like.
Letting the AI say "I'm not sure" is the most important feature. People trust the system because it doesn't pretend. A short list of flagged items is easy to check; a system that is quietly wrong is worse than doing it by hand.
Missing paperwork is as valuable as processed paperwork. Once every DO is tracked, you get a daily list of what hasn't been scanned yet, so the office knows exactly what to chase from site.
Start with one boring document. Not a company-wide AI strategy. One document type, one folder, one list. Once that works, the next one (store transfers, diesel, claims) is much faster.
Results so far
- Over 2,300 delivery orders read in the first two weeks
- 68 out of 68 correct on day one
- Roughly one full-time person's worth of keying work now done overnight
If you run a contractor or SME
You don't need a big IT project to start. If your team spends hours keying paper into a system, that's where AI pays for itself first. I'm happy to share what worked.
Pang Jun Han: operations & AI automation, founder of MyCreatorPulse. pangjunhan.vercel.app
來源：DEV Community · dev.to