BI's original sin

📬 This newsletter has moved. The Vibe Signal is now After the Dashboard — a newsletter about what comes after the era of drag-and-drop BI. This article and all future posts now live on Substack.

👉 Read this article on After the Dashboard →

Subscribe on Substack to follow what's next.

Here's a thing that happens every day, in every company, that nobody finds strange anymore:

Someone has a question about the business. Maybe churn is up. Maybe a campaign is underperforming. The question is clear. The instinct is good.

And then they open a BI tool.

And the first thing the tool asks is not what do you want to know? It's what do you want it to look like?

Pick a chart type. Choose your dimensions. Drag this here. Drop that there. Set your filters. Configure the axes. Name the dashboard.

All of this happens before the person has seen a single row of data.

This is BI's original sin: it forces you to make design decisions before business decisions.

You walked in with a question about churn — a business question, with real stakes — and the tool immediately converted it into a series of design questions. Not "which customers are churning?" but "do you want a horizontal bar chart or a vertical one?" Not "what's driving the drop?" but "which field goes on the x-axis?"

This isn't a bug. It's the architecture. Every traditional BI tool is built around the same assumption: that the human knows what the answer should look like before they know what the answer is. The tool is a canvas, not a collaborator. It waits for instructions. It doesn't think.

When a VP of Marketing wakes up and sees that last week's campaign underperformed, her first thought is not "I need a stacked bar chart filtered by channel, grouped by cohort." Her first thought is: What happened?

That's a vague, open-ended, exploratory instinct. And it's the right instinct. The best analytical thinking starts in ambiguity — you don't know what you're looking for, and you certainly don't know what chart type will be most appropriate. You find those things out as you go.

But BI tools don't let you start in ambiguity. They demand specificity upfront. It's like asking someone to write an outline before they've read the book.

The whole "self-service BI" movement was supposed to fix this. But self-service BI didn't fix the original sin. It just moved it. Instead of making the data team decide what the dashboard looks like before understanding the question, self-service BI made the business user decide. The person asking changed. The backwards sequence didn't.

Now, the industry knows AI is coming for this problem. Salesforce just launched Tableau Next, an "agentic analytics" platform. Microsoft has Power BI Copilot. Chat boxes are everywhere. Hasn't this already been solved?

Not quite. Because the sin is structural, and bolting AI onto a structurally flawed system doesn't fix the structure.

The copilots inside traditional BI tools work by navigating the existing UI on your behalf. Power BI Copilot generates DAX formulas, configures semantic models, and manipulates the same drag-and-drop canvas that was already there. Tableau Next's agents — Concierge, Data Pro, Inspector — automate pieces of the analytics workflow, but still operate within Tableau's existing visualization and data modeling framework. It's clicking the buttons you used to click, picking the chart types you used to pick. It's a faster version of the old workflow. But it's still the old workflow.

Tableau's own FAQ is telling: even with AI, Tableau Agent "can't do data modeling, build dashboards, or create interactivity with elements like filter controls and parameters." The AI lives inside the tool, but the tool still won't let it do the things that matter most. The copilot is trapped inside the sin.

There's a deeper problem that gets less attention.

Traditional BI tools don't just constrain how you ask questions. They constrain what questions you can ask.

Every BI tool comes with a fixed vocabulary of chart types, aggregation methods, and analytical operations. Tableau has its Show Me panel. Power BI has its Visualizations pane. These are finite sets — useful ones, well-designed ones — but finite. When you drag and drop, you are selecting from a menu. And the menu has edges.

This matters more than it used to. Can you do cohort retention with dynamic window sizes? Can you run a simple regression to separate trend from seasonality? Can you build a visualization that doesn't exist in the gallery because your specific business needs a custom way to see the data? In traditional BI, the answer is usually: not without a workaround, a plugin, or an engineer.

AI should blow this open. An LLM can write arbitrary code — it can compute anything, visualize anything. But when AI is trapped inside a traditional BI tool, it inherits the tool's limitations. The copilot can only do what the canvas can render. You've given it a brain, but locked it in a room with a fixed set of crayons.

Meanwhile, something interesting is happening outside the BI industry entirely.

Developers have started using AI coding tools — Claude Code, OpenAI's Codex, Cursor — to build dashboards from scratch. You describe what you want in natural language, the tool generates a working application. No chart gallery. No drag-and-drop. No pre-defined aggregation menu. Just code that does exactly what you asked for, including analytical operations that no traditional BI tool supports out of the box.

Platforms like Squadbase and Replit are making this accessible beyond developers — turning "describe what you want, get a working dashboard" into something a business user can do without a terminal. The dashboards persist. They update when data refreshes. Teams can drill down and filter without re-asking the question from scratch.

This isn't one product or one company. It's a pattern — what you might call vibe coding for BI. People are routing around traditional BI the same way they once routed around traditional media: not because someone declared the old thing dead, but because a new path appeared that was faster and had fewer walls.

Tools like Cortex Analyst and BigQuery's Conversational Analytics are solving a different — and genuinely important — problem: letting analysts and data teams query warehouses in natural language without writing SQL. That's valuable work, and it belongs at the data layer. But it's not BI. Business teams don't just need answers to one-off questions. They need dashboards they can revisit, share, and monitor over time. Investigations need a home. Vibe-coded dashboards give them one — and they do it without inheriting the structural constraints of the traditional BI tools they're replacing.

The thing about original sins is that they're not dramatic. They're structural. Nobody at Tableau or Power BI sat down and said, "Let's make people design dashboards before they think about their data." It just happened. The technology demanded it. And once the interaction model crystallized around it, everyone accepted it as normal.

But the constraint that created the sin — that machines couldn't reason, that someone had to translate a business question into a visual structure manually — is gone now. And when a foundational constraint disappears, the structures built on top of it don't automatically adapt. They have to be rebuilt.

Adding an AI copilot to a traditional BI tool is like adding a GPS to a horse-drawn carriage. It tells you where to go, but it doesn't change how you get there. The tool still thinks in chart types and drag targets and fixed analytical menus. It still asks you to design before you think.

The question for the BI industry isn't whether AI will change analytics. It's whether the tools that defined the last twenty years of BI can let go of the architecture that made them what they are. Whether the next generation starts with a question — just a question — and builds everything else from there.

If the answer is no, the people building dashboards with Claude Code and Squadbase won't wait around.