A day in a mid-sized manufacturer's purchase order team — documents are images, can't even copy

First, we sat next to the staff and watched how they processed purchase orders every day. Log into groupware, search through completed approvals by date and title to find shipment-related ones, open the attached request.

The document was embedded as an image, impossible to copy. "You can't copy this?" we asked. "It always comes like this," they said. The staff was reading numbers on screen and typing them into Excel, one by one. And it wasn't just transcription. Some fields needed manual calculation, especially pricing. On top of that, every vendor needed different values in different formats. This was the daily routine.

The first step was automatic groupware login. The internal groupware didn't have standard APIs, so we had to replicate the actual user flow. Once logged in, the system finds all shipment-related approval documents on its own. The manual process of checking dates and document titles one by one was gone.

The real challenge was that each vendor required completely different data and Excel formats. Vendor A focuses on delivery dates and quantities, Vendor B on unit prices and totals. Hard-coding these differences would mean needing a developer every time a vendor changes. So we built a flexible configuration system. Staff can modify each vendor's key fields, fixed values, unit conversion rules, and date formats anytime. When we first showed the settings screen, the response was: "I can change this myself?" Yes, that was the whole point. Adding a new vendor? Just add a configuration. No developer needed.

Documents in the approval box had subtle format variations even from the same vendor. Date positions shifted, item orders changed, layouts differed slightly. Humans get tired of this. The intuition a person uses when reading a document — 'this number is probably the quantity,' 'that must be the unit price' — that heuristic judgment is what the LLM handles. We used multimodal AI combining natural language and vision processing, but we don't rely on LLMs alone. We combined our own data pre-processing and post-processing pipeline built through years of projects. Image text recognition, key data extraction, and vendor-rule-based transformation all run automatically.

The final output is vendor-specific custom Excel files. Each file matches the vendor's upload format exactly. Date formatting, quantity unit conversion, fixed values, price calculations, all filled in automatically. The generated data also goes straight into the ERP system. When the staff asked, "I really just press this one button?" the answer was yes, that's it. A task that took over an hour daily, and several hours during quarterly peaks, now takes one minute.