The promise: describe your data in natural language and receive a complete dashboard. Upload a dataset, tell the tool what you care about, and it designs the layout, selects the chart types, and populates everything automatically. Five tools now claim to do this. The results are mixed at best.
The test was straightforward. A single dataset — twelve months of e-commerce data with revenue, orders, average order value, return rate, and customer acquisition cost across four product categories and three geographic regions. A moderately complex dataset that any business analyst would recognize. Each tool received the same CSV and the same prompt: "Create a dashboard showing business performance by category and region."
The Results
Tableau (AI-assisted layout). Tableau's AI features suggested a layout with KPI tiles at top, a line chart for monthly revenue, and bar charts for category comparison. The suggestions were reasonable but generic. The tool did not identify that the return rate in one category was anomalous — a finding that a human analyst would flag immediately. The resulting dashboard was functional but uninspired. It looked like every other Tableau dashboard.
Power BI (Copilot). Microsoft's Copilot integration generated a dashboard with a mix of card visuals and bar charts. The natural language interface worked well for simple queries: "show revenue by region" produced the right chart. But the layout was a collection of disconnected visuals rather than a coherent narrative. No visual hierarchy. No indication of what the reader should look at first.
Thoughtspot Sage. Thoughtspot's AI-first approach produced the most interactive results. Natural language queries returned focused answers rather than full dashboards. "Which region has the highest return rate?" produced a bar chart with the answer highlighted. This query-response model is different from dashboard design, and it works well for what it is. But it does not produce a persistent, scannable dashboard.
Observable (AI Assist). Observable's approach was the most technically sophisticated. The AI generated JavaScript code that could be edited and refined. The initial output was a set of small multiples showing each metric by category — a genuinely good design choice. But the tool required JavaScript literacy to modify. This is a tool for practitioners, not for business users.
ChatGPT (Code Interpreter). The weakest dashboard designer. ChatGPT produced individual matplotlib charts in sequence — not a dashboard layout. The charts were correctly plotted but used default styling, default colors, and no annotations. Assembling these into a coherent dashboard would require significant manual work.
Common Failures
All five tools shared three weaknesses.
First, none produced meaningful annotations. No tool identified that the return rate for one category was twice the average. No tool flagged the seasonal pattern in orders. No tool added a single piece of contextual text beyond axis labels and titles.
Second, none established a visual hierarchy. A good dashboard has a clear reading order: the most important metric is largest and most prominent; supporting details are smaller and subordinate. Every AI-generated dashboard treated all metrics as equally important, creating a flat grid with no emphasis.
Third, none asked clarifying questions. A human dashboard designer would ask: "Who is the audience? What decisions will this support? Which metrics are most important?" These questions determine the dashboard's structure. The AI tools skipped them entirely, producing generic layouts that served no specific purpose.
What AI Dashboards Need
The technology is not the bottleneck. The bottleneck is intent. A dashboard is not a random arrangement of charts. It is a designed experience that guides the reader through data in a specific sequence to support specific decisions. AI tools cannot yet infer this intent from a CSV file and a one-sentence prompt.
The tools that came closest — Observable and Tableau — did so by providing strong defaults that a skilled user could refine. This is the right model: AI as assistant, not architect. The design decisions — what to emphasize, what to subordinate, what to annotate, what to omit — remain human decisions.