CJ Olive Young Builds Operational Automation with AI Agents on Amazon Bedrock

CJ Olive Young built two AI agents using Claude on Amazon Bedrock and the Strands Agents SDK to enhance the stability of global promotional operations. The Incident Analysis Agent analyzes root causes and customer impact during outages to propose compensation candidates, while the Daily Health Check Agent regularly monitors operational data to proactively detect anomalies.

Architecture Separating Code and LLM Roles

The core of the agent implementation is the separation of roles: deterministic calculations and judgments are performed by code, while the LLM interprets evidence to assist operator review. Initially, there were issues with response latency and inconsistent conclusions, but consistency was achieved by hardcoding customer impact amount calculations and metric status judgments in code, limiting the LLM to focus solely on analysis and explanation.

Operational Efficiency and Validation Results

Validation with real operational data showed that incident analysis time was significantly reduced from 1-2 days to approximately 5 minutes. Notably, in an incident where 320,000 errors occurred over 6 minutes, root cause identification and damage assessment were completed within 1 minute. For daily checks, a 7-day baseline and minimum traffic threshold were applied to reduce false positives in low-traffic periods. Results are automatically notified via Slack, allowing operators to focus on compensation decisions without waiting for developer analysis.