TwinBI: Agentic Digital Twin Enhances BI Dashboard Interaction

Business intelligence (BI) analysis often involves a dynamic interplay between direct dashboard manipulation and natural language queries powered by Large Language Models (LLMs). A common challenge arises when these two interaction modes fall out of sync, leading to inconsistencies in the analytical state across filters, hierarchies, and metrics. This paper introduces TwinBI, an agentic digital-twin framework designed to address this issue. TwinBI tightly couples an LLM-based agent system with an executable representation of the BI dashboard's state. This integration allows for a unified analytical state, reconstructed from a comprehensive interaction log, which seamlessly combines conversational input, direct dashboard actions, semantic understanding, and provenance tracking. TwinBI also exposes key artifacts such as schema views, SQL queries, and an '/insights' command for state-grounded analytical summaries. Evaluations demonstrate that TwinBI significantly improves exact-match accuracy from 43.3% to 63.3% and partial-credit accuracy from 48.3% to 70.8% compared to using a dashboard alone, while also drastically reducing timeout rates. User studies confirm the benefits of its integrated dashboard-and-chat workflow, highlighting improved task accuracy and favorable ratings for its state-aware interaction mechanisms.