LINE Ads engineers build AI agent to automate advertising analytics reports

LINE Ads developers created an AI agent system to automate the generation of advertising KPI reports, replacing manual data collection and root-cause identification for LINE OA and Display Ads revenue.
The architecture strictly separates LLM and code responsibilities: code handles precise calculations, aggregation, and candidate selection, while the LLM focuses on business interpretation and natural language generation to prevent numerical errors and hallucinations.
A ReAct agent structure replaces fixed analysis paths, using an OpenAI model with high reasoning effort to execute Trino queries and select analysis dimensions dynamically based on observed data.
Five specific tool functions (drill_down, compare_pattern, entity_context, peer_comparison, volume_decompose) are defined via tool calling to decompose revenue changes and compare trends.
An evidence grading system classifies findings into driver_eligible, caution_only, or monitoring_signal to distinguish confirmed causes from weak signals before they are included in the final report.
Context is separated by section type (causal, statistical, business knowledge) to prevent information overload and ensure the executive snapshot only includes verified drivers.
Quality is measured through a dual evaluation system using rule-based quantitative checks and rubric-based qualitative assessments, with Langfuse traces linking execution steps to specific code versions and model parameters.

Key point