Mean time to repair down 38%, independent resolution by new technicians doubled

The group runs more than 40 service stations across China. Repair know-how lived in the heads of technicians with 20 years on the job and in scattered Word files. A new technician facing a difficult fault could only phone a colleague; a remote assist took 40 minutes on average, and in peak season the senior technicians were never off the phone.

The group was turning service from a cost center into a profit center. It wanted a higher first-line resolution rate without hiring more senior technicians, and it wanted every repair to become reusable knowledge.

The FDE team spent 5 weeks on site. It first cleaned the service manuals, historical work orders and technician group-chat logs for the 3 main equipment models into a structured knowledge base, then launched a maintenance assistant for the field.

• The assistant is built into the service app: a technician takes a photo, enters a fault code or describes the symptom, and the Agent returns a troubleshooting sequence, the parts needed and safety precautions, with the original manual page attached
• Closed-loop work orders: when a repair is done, the Agent drafts a structured summary, the technician confirms it, and it goes into the knowledge base, so the system gets smarter with every repair
• Senior technicians handle only the difficult cases the Agent flags as low confidence, so their time goes to the problems that genuinely need experience
• A monthly knowledge base review after go-live, where the service engineering department approves new entries and retires outdated practices

Retrieval is layered: exact match on fault code, then semantic search by model and symptom, then the general manual. A multimodal model identifies components and nameplates in photos from the field. The service app uses an offline cache on weak networks and connects only to sync.