DeepInsight Copilot: AI‑Powered Data Analysis Platform Overview and Technical Evolution

DeepInsight Copilot is a long‑term BI product from Ant Group that focuses on the data analysis stage of the data‑to‑application pipeline, dividing the process into four key parts: Data, Information, Insight, and Decision Recommendation.

The platform supports over 14 data sources, provides data preparation, report creation, self‑service analysis, and intelligent recommendation capabilities through a suite of eight intelligent agents covering tasks such as metric definition, query, analysis, chart generation, summary interpretation, asset search, product Q&A, and tutoring.

Copilot and intelligent agents have a many‑to‑many relationship; a single Copilot can reuse multiple agents and vice versa. Five Copilot assistants combine relevant agents to address specific user needs, improving analysis efficiency.

Four generations of Copilot technology are described: 1) FAQ mode with high precision but high barrier; 2) Dialogue mode with moderate barrier and clarification strategies; 3) Guided dialogue mode that leverages business knowledge, user history, enhanced analysis algorithms, and large models; 4) Multi‑mode fusion analysis that integrates natural language commands with drag‑and‑drop components for seamless report creation.

The underlying architecture evolved from a task‑oriented multi‑turn dialogue system (NLU → DM → NLG) to a standardized intelligent‑agent framework that includes perception, reasoning, and execution modules, enabling more flexible and efficient task handling.

Key AI capabilities include a gated intelligent agent powered by advanced NLU (now using DeepSeek), a Text2Chart model that converts natural language instructions into DSL for chart rendering, and natural‑language generation of analysis reports that combine predictive, anomaly, and distribution analyses.

Model fine‑tuning pipelines for NLU, Text2Chart, Text2DAL, and analysis models are detailed, covering corpus synthesis, training, evaluation, automated testing, reinforcement learning, and continuous iteration.

The article also outlines four major challenges for large‑model‑based data analysis—accuracy, domain knowledge, integration with traditional BI, and scalability—and presents DI Copilot’s solutions such as explicit reasoning traces, user‑guided clarification, customizable knowledge injection, hybrid natural‑language‑and‑drag‑and‑drop workflows, and delegation of heavy computation to specialized engines.

Finally, the future vision is to enable every user to become a data analyst through progressive steps: widespread Copilot adoption, introduction of Pilot assistants, and the construction of virtual analyst teams, while emphasizing the transition from traditional BI to intelligent, AI‑augmented analytics.