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

1.1 Contextualization

• Prior LLM and LVLM analysis has improved analyst productivity by automating routine tasks, while domain-related insight generation remains less explored.The paper distinguishes data observations, factual insights, hypothesis-driven insights, and domain-knowledge insights linking analysis findings with user knowledge.Source: §1.1 · p. 1

1.2 Motivation

• Current analytics workflows often separate data-to-chart generation from chart interpretation, risking disconnection from the raw data and missed dataset-level insights.The paper also identifies limited exploration of domain-knowledge-driven insight generation by LLMs and agents.Source: §1.2 · p. 2

1.3 Problem statement

• The paper frames the challenge as building a robust, generalizable agentic system for domain-related insights and meaningful data visualization.Source: §1.3 · p. 2

1.4 Proposed Solution

• The proposed pipeline transforms raw data through domain identification and insight generation into a dashboard of charts.It uses specialized agents to detect domains and concepts, apply domain-relevant knowledge, generate visual insights, and iteratively reflect.Source: §1.4 · p. 2

1.5 Contributions

• The paper introduces an end-to-end agent-based framework for turning raw business tables into dashboards through domain-informed reasoning.Its contributions include domain detection and knowledge retrieval for feature selection and multi-perspective analysis, plus iterative reflection; managers can validate and refine domain-specific analytics.Source: §1.5 · p. 2

• Data-to-chart systems typically generate one or two charts from single tables and emphasize syntactic correctness or template alignment over insight utility.Source: §2 · p. 2
• Chart-interpretation systems summarize existing charts, often limiting insights to surface features without context-aware reasoning.Source: §2 · p. 2
• Chart QA systems prioritize factual accuracy or verification against known ground truth, not the decision of what to visualize or explore.Few systems ground metric selection or resulting insights in domain knowledge.Source: §2 · p. 2
• QA-pair workflows suit users who can articulate precise queries but constrain autonomous exploration beyond predefined questions.Source: §2 · p. 2
• InsightBench moves toward higher-level insights but remains question-driven and supplies enriched metadata context, leaving open-ended agentic analytical workflows underexplored.Source: §2 · pp. 2–3

3 Proposed Approach

• The system has sequential data-to-insight and insight-to-chart stages, with the first producing structured guidance for domain-tailored visualizations in the second.Here, intermediate “insight” means domain-knowledge insight, rather than question-based insight.Source: §3 · p. 3

3.1 Key Definition

• A domain is a broader business context, while a domain concept is a more granular element within that context.The system treats these as hierarchical relational constructs rather than fixed categories, prompting case-by-case inference to guide downstream analysis and visualization.Source: §3.1 · p. 3

3.2 Stage 1: Data-to-Insight

• Stage 1 takes raw data without supplementary context and produces a domain name, domain-specific concepts, and analytical insights through profiling, detection, generation, evaluation, and reflection.Source: §3.2 · p. 3
• The Data Profiler builds a structured statistical synopsis before questions are posed, inferring column types, ranges, units, dependencies, and potential keys.The resulting profile scaffolds later reasoning and insight generation.Source: §3.2 · p. 3
• The Domain Detector infers a dataset-specific business label from its structural profile using external reference knowledge rather than a closed taxonomy.The authors identify possible label inconsistencies as dataset diversity grows and propose, but do not implement, self-consistency verification.Source: §3.2 · p. 3
• The Concept Extractor identifies domain-relevant, actionable phrases from the domain label and structural metadata, and the Analysis Generator produces descriptive, predictive, and domain-related insights.The insights are organized in a unified JSON object aimed at novel, non-obvious analysis; the Evaluator scores domain inference, concept relevance and comprehensiveness, insightfulness, novelty, and depth, with explanations for ratings.Source: §3.2 · p. 3
• The Self-Reflector uses evaluation scores, feedback, and prior-iteration memory for up to n iterations, stopping early when all criteria meet the threshold.The authors set that threshold at 4 out of 4 across all criteria.Source: §3.2 · p. 3

3.3 Stage 2: Insight to Chart

• Stage 2 uses Tree-of-Thought reasoning and three-expert consensus to compare chart strategies and encodings against domain requirements before selecting a visualization.The authors report that transferring higher-quality Stage 1 domain insights can make chart generation accuracy, especially legends, more challenging.Source: §3.3 · p. 4

4 Evaluation Criteria

• The evaluation framework assesses the system’s ability to generate insightful visualizations from raw business tables through multiple evaluation approaches.Source: §4 · p. 4

4.1 Textual Insight Evaluation

• Textual outputs are scored with adapted G-Eval prompts on insightfulness, novelty, and depth, while InsightBench ground-truth insights provide a partial check of analytical direction.The authors do not adopt InsightBench’s metric because its workflow supplies more context than their raw-data-only system receives.Source: §4.1 · p. 4
• The framework offers domain experts and business analysts a unified rubric for human-in-the-loop validation, with statistically significant professional evaluation planned for future work.Source: §4.1 · p. 4

4.2 Chart Evaluation

• For selected Kaggle datasets, the chart evaluation reviews chart type, legend, and axis labels, and assesses insight quality using G-Eval.Source: §4.2 · p. 4

Experiment 1: Data-to-Insight

• The experiments address whether explicit domain identification matters, how the agentic pipeline compares with a GPT-4o prompt-only baseline, and how domain-knowledge insights differ from question-focused insights.The first two questions use dataset 1; the third compares against a question-focused study using one of its datasets.Source: §5 · p. 4

Experiment 2: Insight-to-Visualization

• A finance survey dataset with 19.9K downloads is used to qualitatively examine chart-type distribution and insight depth.Source: §5 · p. 4

Dataset Selection

• The selected datasets span diverse domains, data types, and analytical complexities, including a business-school marketing simulation, InsightBench, and a Kaggle dataset.The authors state that future work will examine a real-world industrial dataset.Source: §5 · p. 4

Result 1: Explicit domain identification matters

• Adding domain identification improves insight coverage, structure, and business relevance; the domain-aware analysis links customer churn to retention-program strategy.The baseline produces broad categories described as not materially supported by the raw data; Figure 3 says domain identification grounds insights and fends off hallucinations.Source: §6 · p. 4, Figure 3 · p. 5

Result 2: Our agentic system works better compared to the baseline.

• The agentic system significantly outperforms domain-aware non-agentic GPT-4o in insightfulness, novelty (0.599 vs. 0.390), and depth (0.942 vs. 0.803).The reported insightfulness advantage includes identifying nuanced effects of acquisition channels and premium status on customer retention and lifetime value.Source: Table 1 · p. 5, §6 · p. 5

Result 3: Insights generated with domain-knowledge can capture the right direction of analysis,

• Against InsightBench ground truth, the system captures the core analytical direction and broader concept coverage but misses variation across users and departments.Its generated concepts include cost variability, processing dynamics, optimization opportunities, automation, compliance, and AI integration, while InsightBench focuses on operational behavior and expense brackets.Source: Figure 4 · p. 5, §6 · p. 5
• The authors cannot establish the validity of the system’s additional analysis directions without domain-expert evaluation.Source: §6 · p. 5

Result 1: Explicit domain identification matters

• The authors interpret the results as suggesting domain knowledge is essential for insightful analysis, while identifying causal experiments as a future test.Source: §6 · p. 5

Result 4: An observation for Human Analyst and Our System

• Against the Kaggle user’s analysis, the system significantly outperforms on the task, although two generated charts were plotted incorrectly because of coding errors.The system generated five chart types; the Kaggle users’ charts included eight bar charts and one box plot.Source: Table 2 · p. 5, §6 · pp. 5–6