Data-to-Dashboard: Multi-Agent LLM Framework for Insightful Visualization in Enterprise Analytics

Ran Zhang, Mohannad Elhamod

TL;DR

Researchers have developed a multi-agent LLM framework that automates the journey from raw business data to insightful visual dashboards. By simulating the reasoning process of a human analyst—including explicit domain detection and iterative self-reflection—the system generates visualizations that are significantly more insightful, novel, and deep than standard single-prompt AI baselines.

What Problem Does This Paper Address?

While Large Language Models (LLMs) are increasingly used for data analysis, most current systems suffer from "tunnel vision." They typically focus on low-level factual extraction (e.g., "What was the highest sales month?") or simple chart summaries.

The authors identify a critical gap: domain-knowledge grounding. In a real enterprise setting, an analyst doesn't just look at numbers; they interpret them through the lens of the business (e.g., connecting a drop in user activity to a specific retention strategy). Existing AI tools often lack this context, leading to "hallucinated" insights or generic observations that don't help decision-makers. Furthermore, the process of creating a chart and interpreting it is often disconnected in AI workflows, which can lead to visual representations that fail to capture the most important underlying trends.

Methods

The authors propose a two-stage agentic framework designed to mimic the cognitive workflow of a business analyst:

Stage 1: Data-to-Insight

Instead of jumping straight to charts, a suite of specialized agents collaborates to understand the data:

  • Data Profiler: Uses "Tree-of-Thought" prompting to create a statistical synopsis (types, ranges, dependencies).
  • Domain Detector: Infers the business theme (e.g., "Finance" or "Marketing") using external knowledge like Wikipedia, rather than relying on a fixed list of categories.
  • Concept Extractor: Identifies salient business concepts (e.g., "monthly active users," "unit cost").
  • Analysis Generator: Synthesizes insights through three lenses: descriptive, predictive, and domain-related.
  • Evaluator & Self-Reflector: Using the Reflexion framework, the system scores its own insights and iterates until they meet a high quality bar (4/4 on all metrics).

Stage 2: Insight-to-Chart

The system uses a Tree-of-Thought (ToT) reasoning framework to transform insights into visualizations. Three "expert" agents must reach a consensus on the best chart type, rationale, and annotations. This ensures the final dashboard isn't just a collection of random plots, but a cohesive narrative tailored to the business context.

Datasets

The framework was evaluated on three distinct datasets:

  1. Marketing Simulation (Wharton): A complex dataset focused on customer centricity.
  2. InsightBench: A benchmark designed to evaluate business analytics agents through multi-step insight generation.
  3. Personal Finance Survey (Kaggle): A popular real-world dataset used to compare the AI's performance against high-ranking human analysts.

Key Insights and Ideas

  • Domain Grounding as a Lever: The most important "aha moment" is that simply instructing the model to identify the domain first acts as a powerful prompt-engineering lever. It triggers a specific "knowledge frame" in the LLM, leading to deeper topic granularity.
  • The "Fishing" Problem: Without domain context, LLMs tend to "fish" for general business axes (like "regional focus") that might not actually be supported by the raw data.
  • The Complexity Trade-off: Interestingly, the authors found that the quality of generated insights was sometimes inversely related to chart generation accuracy. As insights become more nuanced and complex, the technical task of plotting them accurately (e.g., creating complex legends) becomes more challenging for current models.

Key Results

The framework was benchmarked against a non-agentic GPT-4o baseline and human-generated content:

  • Vs. GPT-4o Baseline: The multi-agent approach showed a +12% lift in insightfulness, a +28% lift in novelty, and a +31% lift in analytical depth.
  • Vs. Human Analysts (Kaggle): In a head-to-head comparison with the most "upvoted" human analysis on a finance dataset, the system outperformed the human in every G-Eval metric, including a staggering +147% improvement in insightfulness.
  • Alignment: The system successfully captured core analytical directions found in ground-truth benchmarks, such as identifying cost variability and optimization opportunities.

Limitations and Future Work

The authors are transparent about several hurdles:

  • Technical Execution: Due to the complexity of Stage 2, some charts were plotted incorrectly due to coding errors, highlighting that the "Insight-to-Chart" pipeline needs more robust debugging.
  • Evaluation Bottlenecks: While G-Eval (LLM-based evaluation) is useful, the authors emphasize the need for more "human-in-the-loop" validation from actual domain experts in future iterations.
  • Scalability: The current work is a proof-of-concept; testing it on massive, multi-table industrial databases remains a future goal.

Paper Citation: Ran Zhang and Mohannad Elhamod.

  1. Data-to-Dashboard: Multi-Agent LLM Framework for Insightful Visualization in Enterprise Analytics. In Proceedings of 2nd Workshop on Agentic AI for Enterprise (ACM SIGKDD). GitHub Repository

Summary updated

Abstract

The rapid advancement of LLMs has led to the creation of diverse agentic systems in data analysis, utilizing LLMs' capabilities to improve insight generation and visualization. In this paper, we present an agentic system that automates the data-to-dashboard pipeline through modular LLM agents capable of domain detection, concept extraction, multi-perspective analysis generation, and iterative self-reflection. Unlike existing chart QA systems, our framework simulates the analytical reasoning process of business analysts by retrieving domain-relevant knowledge and adapting to diverse datasets without relying on closed ontologies or question templates. We evaluate our system on three datasets across different domains. Benchmarked against GPT-4o with a single-prompt baseline, our approach shows improved insightfulness, domain relevance, and analytical depth, as measured by tailored evaluation metrics and qualitative human assessment. This work contributes a novel modular pipeline to bridge the path from raw data to visualization, and opens new opportunities for human-in-the-loop validation by domain experts in business analytics. All code can be found here: https://github.com/77luvC/D2D_Data2Dashboard
BibTeX
@misc{zhang2025data,
  author = {Ran Zhang and Mohannad Elhamod},
  title = {{Data-to-Dashboard: Multi-Agent LLM Framework for Insightful Visualization in Enterprise Analytics}},
  year = {2025},
  eprint = {2505.23695},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.23695},
}

Loading related content...

Figure