LG CNS Case Study: Designing and Automating APQR Systems with Agentic AI

Chong Kun Dang and LG CNS applied Agentic AI to reduce the time for creating Annual Product Quality Review (APQR) reports from the previous 12 hours to approximately 1 hour. By shifting repetitive tasks—where quality assurance personnel previously gathered data across multiple systems—to system-level automation, staff can now focus on reviewing evidence and exceptions.

System Architecture

The AWS-based platform consists of two workflows: data preparation and report generation. Files from QMS, LIMS, SAP, and EDMS are standardized into evidence datasets, which then drive the report generation workflow.

• Business Agent: Handles 12 quality tasks, including deviations, CAPA, and stability testing.
• Unit Agent: Performs 28 types of unit functions, such as queries, calculations, interpretation, and document generation.
• AWS Services: Utilizes Amazon S3 (storage), Athena (querying), EKS (execution), Bedrock (interpretation), RDS (history management), and others.

Core Design Principles

The verification methods for numerical calculations and generative AI narratives are separated. Queries and statistical calculations are handled by reproducible code, while the LLM is restricted to interpretation and narrative generation based on verified results. This approach allows for clear identification of root causes in case of result errors, and the workflow, composed of over 66 execution nodes, operates stably.