Project Background
East China Pharmaceutical Group, a leading domestic pharmaceutical manufacturer, manages annual procurement exceeding RMB 5 billion, generating a massive volume of daily purchase documents. All procurement data relied on manual recognition from PDFs/images and manual entry into the SAP ERP system, followed by approval workflows initiated via DingTalk. The lack of data integration between the two systems led to duplicate data entry, information delays, and approval bottlenecks.
Core Pain Points
Solution
AI Intelligent Recognition and Automatic Entry
Deployed an OCR+LLM dual-engine purchase order recognition system supporting intelligent parsing of multiple formats including PDFs, images, and scanned copies. The system automatically extracts key fields such as supplier information, material codes, quantities, and prices, and auto-fills them according to SAP field mapping rules.
ERP-DingTalk Data Integration
Developed a standard API integration middleware between SAP ERP and DingTalk to enable bidirectional procurement data synchronization. Once data enters ERP, an approval workflow is automatically triggered in DingTalk; approval results are written back to ERP in real time, eliminating duplicate entries.
Intelligent Anomaly Alerts
Built a procurement anomaly detection model based on historical data to automatically identify and alert on risks such as price deviations, quantity abnormalities, and duplicate orders.
Results Data
| Metric | Before | After | Improvement |
|---|---|---|---|
| PO Entry Time | 2 hrs/PO | 5 mins/PO | ↓96% |
| Entry Error Rate | 8% | 0.5% | ↓94% |
| Procurement Team Size | 8 persons | 3 persons | ↓63% |
| Approval Completion Time | 3 business days | 0.5 business days | ↓83% |
> Quantitative Summary: PO entry time reduced by 96% to 5 minutes per order, error rate lowered by 94% to 0.5%, procurement team downsized from 8 to 3 people, and approval completion time shortened by 83% to 0.5 business days.
Technology Stack
SAP ERP, DingTalk Open Platform, OCR Engine, Qwen Large Language Model, Node.js Middleware, PostgreSQL
FAQ
How is the OCR+LLM dual engine better than pure OCR?
Pure OCR achieves about 80%-85% recognition accuracy for complex formats (handwriting, stamp overlays, non-standard templates). With LLM, semantic understanding fills in missed or misrecognized fields, boosting overall recognition accuracy to over 99%. In this case, the error rate dropped from 8% (manual) to 0.5% (AI entry).
Does the AI intelligent entry support multiple purchase order formats?
Yes. The OCR+LLM dual engine supports PDFs, images, scans, faxes, and more, without requiring predefined templates. The LLM automatically understands the structure of purchase orders from different suppliers, extracts key fields, and maps them to standard SAP fields.
How much manual intervention is needed post-deployment?
Only 3 procurement clerks (down from 8) handle exception documents where AI confidence is low, accounting for about 5% of total documents. 95% of purchase orders now flow through fully automated entry → approval → archiving without any manual intervention.
