Page 331 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

innovation and technology integration, ensure that every emergency can be quickly responded
to and properly handled.

Hidden Disease Rectification: A series of automated operations and maintenance tools, such
as configuration auditing and log auditing, are generated based on the audit LLM. Incremental 4.3 - 5G
tuning of the audit LLM based on real-time network configuration improves the audit accuracy
rate by over 95% and further reduces the audit time by approximately 50% compared to the
existing baseline.

Timely discovery: Through the synergy of large and small models, real-time monitoring of gold
indicators, and rapid identification of abnormal events in cloud-network operation. The next
step is expected to build an anomaly dataset covering the entire specialty of cloud-network
operation, assess the anomaly detection model more accurately, and improve the model
accuracy rate by over 95% through the introduction of time series LLM and model fine-tuning.

Accurate positioning and evidence-based disposal: After emergency events are disposed of,
fault cases are automatically stored in the database, and the knowledge graph is automatically
generated, so as to realize 100% of the experts' experience in the management and to provide
more reference experience for the autonomous learning of the intelligent body. The cross-
specialty fault location scheme is further refined to ensure that the decision-making accuracy
of the intelligent body is improved by more than 95%.

2�2 Benefits of the use case

The use case focuses on evolving China’s Cloud-Network O&M model and has the following
impacts:

(1) Reduced network maintenance costs‐The introduction of digital employees has cut
telecom O&M personnel by 30%. Take China Telecom’s promotion as an example: 400+
registered intelligent agents, 800+ associated API applications, 800+ daily user visits, and
5,000+ daily usages—saving approximately 5,000+ hours/day on query, emergency, and
work order handling.
(2) Enhanced cloud-network operation quality and efficiency: Leveraging existing
configuration standards, the system automatically generates configuration templates. In
practical production, the abnormal configuration recall rate reaches 94%, precision rate
85%, and China's network device configuration compliance rate has improved to over
95%, ensuring devices "go online without defects" and "cut over without issues".
(3) Improved user satisfaction: 90% of network issues achieve the "1-5-10" fault capability,
pioneering nearly one year of "0 public opinion" and "0 failures" for general and
above-level faults in the telecom industry. The large-small model collaborative anomaly
detection method has reduced original 10-minute detection to 1 minute, with an average
accuracy of over 90%. Through intelligent agents for fault localization and handling, fault
localization time is shortened by 95%, ensuring fault resolution within 10 minutes and
boosting user friendliness to 99.8%.
(4) Cross-industry deployment in China: Deployed in O&M for China's government, culture-
tourism, healthcare, and other industries. It has provided innovative digital workforce
O&M services to hundreds of clients and trained over 2,000 development-oriented O&M
teams.
(5) Implementation in foreign telecom operators: Primarily applied to fault detection and
root cause location scenarios.

295