Unlocking Actionable Insights with Jupybara: A Multi-Agent AI Assistant for Data Analysis and Storytelling

In today’s data-driven world, extracting insights isn’t enough—what matters is turning those insights into informed action. Whether you’re analyzing customer behavior, optimizing manufacturing pipelines, or evaluating public policy outcomes, one needs to go beyond charts and statistics to communicate actionable insights clearly and persuasively. That’s where Jupybara comes in; Tableau Research presented at the 2025 ACM CHI conference on Human Factors in Computing Systems, the premier international conference of human-computer interaction. This was work done by research intern, Will (Huichen) Wang, a Ph.D. student at the University of Washington in collaboration with Vidya Setlur and Larry Birnbaum, a professor at Northwestern University.

From Exploration to Communication

Jupybara is an AI assistant designed for the full lifecycle of data analysis, from exploratory data analysis (EDA) to insight-driven storytelling. It’s implemented as an extension to the Jupyter Notebook environment and integrates large language models (LLMs) to collaborate with analysts directly where they work. Figure 1 shows the system interface.

Jupybara is based on a three-dimensional design space that draws from theory in data visualization, narrative discourse, and communication. These dimensions guide how Jupybara generates, critiques, and refines data narratives—transforming insights into impact. 

• Semantic Precision: Ensures that data insights are expressed accurately and grounded in analytical facts. For instance, when visualizing a sharp drop in average hourly wage, Jupybara offered alternatives like “fell,” “declined,” or “crashed”, each carrying a different semantic nuance. This reflects the challenge of aligning language with the statistical patterns of the data.
• Rhetorical Framing: Supports persuasive, goal-driven messaging to align insights with audience expectations. For example, when analyzing the gender pay gap, Jupybara's storytelling agent begins the narrative with a compelling headline-style summary -  “Women are paid 17% less than men in full-time roles,” and follows it with interpretive framing, such as, “Even after adjusting for industry and education level, this gap persists, suggesting structural bias rather than compositional effects.”
• Pragmatic Relevance: Anchors insights in real-world decision-making, factoring in domain context and intended actions. For example, identifying a significant pay gap in full-time roles, Jupybara makes a suggestion to the user, “Organizations should implement transparent salary bands and review promotion rates by gender every six months.”

Figure 1. The interface of Jupybara, an AI-enabled assistant for actionable EDA and data storytelling implemented as a Jupyter Notebook extension. (A) For a complex user query in EDA, Jupybara identifies and presents an analysis plan before producing code. (B) In the data story generated by Jupybara, the system uses precise language to convey analytical results; appropriate hooks, connectives, and narration of analytical strategies to bolster actionable insights; and relevant domain knowledge to connect data facts to actionable insights.

Multi-Agent Collaboration for Supporting Human-AI Data Analysis

Jupybara introduces a multi-agent architecture where different language model agents assume specialized roles in the analysis and storytelling process. One agent may interpret statistical results and visualizations, another may critique clarity and rhetorical framing, and a third ensures relevance to domain-specific contexts. This division of labor enables iterative refinement, drawing inspiration from collaborative workflows among human analysts. While Jupybara also supports a single-agent mode for simpler tasks or time-sensitive situations, the multi-agent approach offers a promising direction for future research—particularly in analytical scenarios that benefit from layered reasoning, peer-like critique, or adaptive storytelling strategies.

Insights from the Field

To evaluate how analysts interact with AI-assisted tools for EDA and storytelling, the researchers evaluated Jupybara with nine experienced data analysts across various domains, including finance, healthcare, and retail. Participants completed an EDA task and generated data stories using both single- and multi-agent versions of Jupybara, working with datasets relevant to their professional backgrounds.

Participants appreciated how Jupybara reduced the mental burden of tracking evolving insights and managing analysis history with the "Insight Tracker". The "Clarify Tab" further allowed them to engage in threaded conversations to understand AI output. They also found Jupybara helpful in refining the language used to convey findings. The tool's semantic and rhetorical guidance helped users make nuanced word choices and construct persuasive narratives. They valued the support Jupybara provided in choosing analytical strategies. The system surfaced rationale for operations (e.g., why missing values were treated a certain way), helping participants reflect on and justify their choices. Participants consistently rated responses generated via the multi-agent architecture as higher quality across all three dimensions of the design space—semantic, rhetorical, and pragmatic, compared to the single-agent responses. 

A Tool for Analysts, A Platform for Research

Beyond helping analysts, Jupybara also serves as a research probe for studying human-AI collaboration. Our evaluation highlights Jupybara’s usability, steerability, explainability, and reparability—all important traits for collaborative human-AI systems.

As multi-agent AI tools become more prevalent, we see Jupybara as a first step in this direction. Future work could explore agents that specialize in different storytelling genres or domains (e.g., healthcare vs. finance), or dynamically adapt to evolving data streams. Jupybara is open-sourced, and we encourage researchers, educators, and practitioners to explore new ways to analyze, interpret, and communicate data insights using this tool.