The Future of Visualization Dashboards in the Age of Generative AI

1 Introduction

• Participants questioned whether dashboards remain the right medium for exploration when AI can answer natural-language questions, and whether authors still need to elicit requirements and translate them into visualizations.The introduction also notes that dashboard use involves expertise, collaboration, and work across tools despite no-code promises.Source: §1 · pp. 1–2
• The study used semi-structured interviews with 16 experts and thematic analysis organized around two research questions: dashboards’ future and the future of their creators and users.Participants discussed current challenges, LLM use, interface expectations, and stakeholder roles.Source: §1 · p. 2
• Almost all participants expected dashboards to persist while changing form, with interaction remaining part of how people explore and make sense of data.Source: §1 · p. 2
• Participants expected authors’ responsibilities to shift toward work delegated less easily to AI, while raising concerns about understanding AI output and preserving shared understanding as dashboards become easier to generate and personalize.Source: §1 · p. 2
• The paper contributes seven research and practice opportunities spanning validation, education, maintenance, organizational guidance, adaptation, novel visualizations and interactions, and accountability for AI-generated content.Source: §1 · p. 2

2 Related Work

• The review organizes prior work around dashboard design and organizational practice, AI assistance in authoring and interaction, and human roles in AI-assisted analysis.Source: §2 · pp. 2–3

2.1 Dashboard Design, Use, and Organizational Practice

• Dashboards serve monitoring, exploration, communication, and decision support; this paper defines them as interactive visualizations on one screen, while stable layouts support recurring workplace routines [3].Prior work also reports that no-code tools did not remove dependence on experts or work across multiple tools.Source: §2.1 · p. 3

2.2 AI Assistance in Dashboard and Visualization Authoring

• Natural-language systems translate requests into visualization specifications, actions, or code, while multimodal tools combine language and graphical interaction for iterative edits [24] [44] [30] [7] [40] [43].The literature also frames a debate over whether dashboards will persist or be replaced by AI-delivered insights.Source: §2.2 · pp. 3–4

2.3 Human Roles and Responsibilities in AI-Assisted Analysis

• Prior research examines human and automated work, but often in biomedical visualization [46], one organization, or exploratory sessions; this study addresses dashboard practice across domains and organizations.The review connects authors’ responsibilities to data preparation, organizational needs, usability, and effective dashboard use [42].Source: §2.3 · pp. 4–5

3 Method

• The researchers conducted semi-structured interviews about current dashboard challenges, LLM use, and expectations for dashboards and stakeholder roles.Source: §3 · p. 5

3.1 Participants

• The 16 participants came from 14 countries and included academics, industry practitioners, and consultants with varied dashboard responsibilities and about 8 to 25 years of relevant experience.The sample comprised eight academics, four industry participants, and four consultants.Source: §3.1 · pp. 5–6

3.2 Interviews

• Interviews lasted about 30 to 45 minutes and covered participants’ backgrounds, current practices, AI influences, possible interface changes, and future human skills.Source: §3.2 · pp. 6–7

3.3 Analysis

• The team collaboratively coded transcripts with an iteratively refined shared codebook, then grouped codes into themes aligned with the two research questions.Coding was not independent and inter-coder agreement was not computed.Source: §3.3 · p. 7

4.1.1 Current Challenges with Dashboards

• Participants described three current challenges: understanding users’ tasks and audiences amid weak feedback, the effort of building and maintaining dashboards, and cases where dashboards are the wrong medium.Changing requirements, fragile data dependencies, and platform updates add to maintenance work.Source: §4.1.1 · p. 8

4.1.2 Current Integration of LLMs in Dashboards

• Most participants used LLMs for code, mock-ups, and design suggestions, but adoption varied with doubts about output quality, rigid workflows, and confidential data.Some constrained models to a single data source to reduce hallucinations, while noting this did not guarantee accuracy.Source: §4.1.2 · p. 9
• LLMs often improved output quality or expanded the work participants could attempt without necessarily saving time, while polished visualizations could conceal errors and opaque analysis.Participants described incorrect KPIs, misleadingly ordered axes, and difficulty spotting hallucinations in visual output.Source: §4.1.2 · p. 9

4.1.3 Expectations for the Future

• Almost all participants expected dashboards to persist for recurring questions, monitoring, reporting, and organizational habits, while becoming more adaptive or tailored to users and settings.Participants pointed to regular reporting needs and existing customer bases as reasons dashboards would remain in use.Source: §4.1.3 · p. 10
• Participants expected language, speech, text, and gestures to complement graphical controls, with modality choice shaped by task and users continuing to interact to explore and understand data.They noted that familiar filters can be quicker than prompts and that direct manipulation can provide immediate visual feedback.Source: §4.1.3 · pp. 10–11
• Participants worried that easier personalization could proliferate dashboards and undermine shared truth, and that reduced human involvement could obscure effects on people or expose systems to security risks.They also anticipated users needing explanations of how AI reached an answer.Source: §4.1.3 · pp. 11–12

4.2 RQ2: The Future of Stakeholder Roles

• Most participants expected authors and analysts to move from implementation toward specifying requirements, curating AI-generated work, and communicating with users while retaining design and human skills.Participants emphasized understanding users’ questions and learning goals, design knowledge, empathy, critical thinking, and creativity.Source: §4.2 · pp. 12–13
• Participants expected humans to evaluate and approve dashboards, but generating many candidates could make checking outputs and deciding readiness laborious.They noted that stakeholder discussion and sign-off remain necessary before production use.Source: §4.2 · p. 13
• AI could let some end users explore data without contacting analysts, but differences in analytical knowledge would remain important.Participants questioned whether users who struggle with dashboards could ask an LLM the right questions; they expected natural-language interactions to be easiest for people already skilled in data analysis.Source: §4.2 · p. 13
• Participants worried that relying on AI could reduce analytical thinking and learning, leaving dashboard production unmatched by understanding of how the output works.One participant reported that students built more complete applications but struggled to explain them; another worried that authors would put less thought into dashboards as they offloaded more work to AI.Source: §4.2 · pp. 13–14

Dashboards persist but evolve

• Most participants expected dashboards to persist but evolve: public-facing and monitoring dashboards would remain largely as they are, while analytical dashboards would change most.Expected changes include adaptation to user needs and a wider range of interactions, including gestures.Source: §5 · p. 14

Easier creation redefines user roles and responsibilities

• Participants expected authors to spend less time implementing dashboards and more time specifying requirements, choosing and checking generated options, and taking responsibility for production outputs.Understanding audiences, evaluating AI output, and applying design principles were still expected to matter.Source: §5 · p. 14

Dashboards persist but evolve

• Familiar dashboard layouts can support repeated use: situated action and distributed cognition frame spatial stability as a work practice and an external cognitive scaffold.A proposed middle path keeps structure stable while generating content for the question at hand, drawing on dashboard design patterns [3].Source: §5 · p. 14

6.1 Making Human Validation Manageable

• Validation must cover values, calculations, visual encodings, and whether views answer the intended question; interfaces could expose provenance and changes alongside automated checks and human approval.Proposed support needs evaluation for the effort it requires and the errors it leaves undetected.Source: §6.1 · pp. 15–16

6.2 Helping End Users Recognize Mistakes

• End-user authoring calls for education in how scales, aggregations, filters, and representations shape meaning, plus opportunities to examine how generated dashboards were constructed.The proposed learning support could be integrated into work with users’ own data.Source: §6.2 · p. 16

6.3 From Heirloom to Fishwrap: Managing Dashboard Proliferation and Rot

• Easier creation may proliferate dashboards while maintenance still requires backward compatibility and responses to platform updates; research could track reuse, ownership, errors, and obsolete assumptions.Lifecycle tools could help distinguish temporary explorations from maintained data products and flag outdated dashboards.Source: §6.3 · p. 16

6.4 Developing Organizational Guidance for AI-Assisted Dashboards

• Organizations need actionable guidance on which data AI may process, where processing occurs, and who can authorize it, especially for confidential data.The proposed research would involve dashboard teams, data owners, and privacy specialists in translating policy and regulatory requirements into guidance.Source: §6.4 · pp. 16–17

6.5 Balancing Adaptation with Shared Understanding

• Personalized views may better fit changing individual goals but can weaken shared understanding when people see different representations of the same data.Tools could retain a shared reference view while making personal adaptations visible and reversible.Source: §6.5 · p. 17

6.6 Considering the Future of Novel Visualizations and Interactions in Dashboards

• Novel interactions may be appreciated in demonstrations yet used infrequently, so long-term evaluation should examine task fit, continued use, and the support users need.The proposed studies distinguish novelty, complexity, and unfamiliarity rather than treating them as interchangeable.Source: §6.6 · p. 17

6.7 Accuracy, Hallucination, and Accountability

• Because authoritative-looking dashboards can contain subtle errors and influence decisions, research should address appropriate accuracy expectations, disclosure, provenance, and responsibility.Proposed tools could mark AI-generated views, retain provenance, and define approval roles; the opportunities still require empirical evaluation for benefit and cost.Source: §6.7 · p. 18

Sample representativeness

• All 16 participants reported using LLMs, limiting insight into people who chose not to use these tools; participants’ presence in 14 countries does not make the sample nationally representative.Source: §7 · p. 18

Perspectives of other stakeholders

• The study interviewed dashboard experts rather than end users, managers, or other stakeholders directly, so findings about those groups largely reflect experts’ perspectives.Specialized analytical-task expectations may not apply to everyday dashboard use.Source: §7 · p. 18

Temporal and methodological limitations

• Interviews combined current practice with future expectations and did not observe work over time or independently evaluate systems, so they cannot establish whether anticipated changes will occur or their effects.Descriptions of productivity and errors reflect participants’ experiences rather than independent measurements.Source: §7 · p. 18

8 Conclusion

• Almost all participants expected dashboards to persist for recurring questions, monitoring, and established reporting, with adaptive views and varied interaction supporting exploration and understanding.Source: §8 · p. 19
• Authors may shift toward specifying requirements, selecting and checking outputs, while audience understanding, design knowledge, and stakeholder communication remain important.Greater end-user independence also raises questions about users’ ability to assess what they create.Source: §8 · p. 19
• The paper develops seven research and practice opportunities, and concludes that future dashboards’ value depends on supporting responsibilities for recognizing mistakes, preserving shared understanding, sustaining dashboards, and answering for their content.Source: §8 · p. 19

• Gemini 3.8 Image Flash generated the teaser figure.Source: Use of Generative AI · p. 19
• Claude by Anthropic generated the tables and supported parts of the analysis.Source: Use of Generative AI · p. 19
• The authors reviewed and verified all outputs.Source: Use of Generative AI · p. 19