From Dashboards to Dialogues: Using AI and AI Agents to Build Self-Service Business Intelligence Reports

For two decades, “self-service Business Intelligence (BI)” has been a promise the industry has consistently under-delivered. Modern BI suites give business users a thousand knobs to turn, but answering even a moderately novel question still requires a data analyst, a semantic-layer engineer, or a dashboard developer in the loop. Large Language Models (LLMs), and increasingly agentic compositions of them, change the economics of that loop. An LLM can read a schema, infer relationships, propose visualizations, narrate insights, and orchestrate a series of dependent steps that previously demanded hours of human effort. This paper articulates a concrete pattern for combining LLMs, deterministic data tooling, and high-fidelity visualization libraries to produce genuinely self-service BI reports. We use JasusViz-AI-Lib, an open-source MIT-licensed reference implementation, as a worked example. The library combines GitHub Copilot (gpt-4o), Google Gemini, Anthropic Claude and OpenAI behind a unified client interface; pandas-based data processing; FastAPI endpoints; and 14 hand-crafted D3.js v7 visualizations rendered by a React 18 frontend. We describe its three-layer architecture (schema_processor, data_engine, ai_integration), its multitable foreign-key inference workflow, and the engineering principles — especially “AI is an assistant, not a source of truth” — that keep AI output safe to put in front of business users. We then generalize the pattern into a multi-agent topology suitable for production self-service BI, and discuss governance, security, and the open research questions that remain.