TwinBI: Agentic digital twin boosts BI dashboard accuracy by 20%

Business intelligence (BI) dashboards increasingly integrate LLM-based assistants, but switching between direct manipulation and natural-language queries often breaks analytical consistency. Enter TwinBI, a framework from researchers Jisoo Jang and Wen-Syan Li that creates an 'agentic digital twin' — a shared analytical state that couples an LLM agent with the live dashboard. This allows users to filter, drill-down, and ask questions without losing context, as the system reconstructs state from a unified interaction log. TwinBI exposes artifacts like schema views, SQL queries, and an /insights command for state-grounded summaries.

In controlled A/B testing with the same backbone agent, TwinBI improved exact-match accuracy from 43.3% to 63.3% and partial-credit accuracy from 48.3% to 70.8%. Timeout rates plummeted from 40% to 10%, showing far more reliable multi-step analysis. A usability study confirmed high task accuracy and favorable ratings for the state-aware interaction. The dataset and source code are publicly available on arXiv. TwinBI points to a future where BI tools maintain coherent analytical narratives across both manual and conversational interactions.