AI Invoice Processing — Auto-Extract from LINE or Google Drive

Back-office teams manually keyed invoice data into accounting systems. AMCOLAB built an AI extraction pipeline supporting two input paths: LINE chatbot photo upload or Google Drive file drop.

Tech Stack

Project Info

Platform Mobile (LINE) + Web (Google Drive)

User Type Internal (accounting team)

Client Type SME

Engagement Fixed Price

Region Japan

The Challenge

Accounting teams are not paid to type. But that is what this one spent a significant portion of its time doing — reading paper invoices and re-keying every field into the accounting system. The volume was high enough that the task consumed meaningful hours each week, and errors introduced during manual entry went undetected until month-end reconciliation.

The added complication was input fragmentation. Invoices arrived through two entirely separate channels — LINE photos from field staff, and PDF files uploaded to Google Drive — with no unified process to handle both. Any automation solution had to work across both input paths without adding complexity for the team.

What We Built

AMCOLAB built a two-channel AI extraction pipeline that processes invoices automatically from the moment they arrive — regardless of how they were submitted. The accounting team receives structured, ready-to-import data. They review exceptions, not every document.

LINE Chatbot Input

Staff photograph a paper invoice and send it to a LINE chatbot. The image enters the extraction pipeline immediately and a confirmation is returned once processing is complete.

Google Drive Input

Files dropped into a designated Drive folder trigger the pipeline automatically. The folder acts as a live inbox — no additional action needed from staff.

AI Extraction (OCR + Vision)

PaddleOCR processes standard printed invoices. OpenAI Vision handles complex layouts and handwritten content. Extracted fields include vendor name, dates, line items, tax amounts, and totals.

Structured Output to Google Sheets

Extracted data is written to a structured Google Sheet — one row per invoice, fields separated and labelled — ready to import into the accounting system.

Confidence Flagging

Documents below a confidence threshold are held for manual review. The team sees only the cases the system flagged as uncertain, not the full volume.

Key Outcomes

Invoice processing time cut substantially for standard documents

Accounting team effort concentrated on exceptions rather than routine data entry

Two separate input channels handled through one consistent process

Pipeline runs continuously — invoices processed overnight are ready when staff arrive in the morning - Error rate in structured output reduced compared to manual keying

Error rate in structured output reduced compared to manual keying