Why Business Intelligence Is Finally Getting Interesting

Here's something nobody wants to admit: we've been staring at dashboards for decades, and most of them have told us exactly nothing useful.

IBM's Josh Spurgin isn't wrong when he says decision-makers are "drowning in reports and dashboards, but they're starving for insights." That phrase—drowning but starving—captures something real about the weird paralysis that happens when you have infinite data but zero clarity about what to actually do.

The pitch for what comes next sounds simple enough: instead of clicking through seventeen different dashboards to figure out why Q3 sales tanked, you just... ask. "What's driving our drop in Q3 sales?" And an AI system—powered by large language models and something called retrieval augmented generation—gives you an actual answer. In English. With context.

It's called conversational BI, and if it works the way Spurgin describes, it would be genuinely transformative. But there's a lot happening under that "if."

The Promise: Data That Actually Talks Back

The core idea is straightforward. Traditional business intelligence gave us pretty visualizations of what already happened. Conversational BI, at least in theory, gives us a dialogue partner that can explain not just what happened, but why it happened and what to do about it.

The technology stack involves two main pieces working together. First, there's the large language model—the thing that understands natural language and can generate human-readable responses. Think ChatGPT, but trained on your company's specific data patterns.

But here's where it gets interesting: LLMs alone are famously unreliable with facts. They're great at sounding confident while being completely wrong. That's where RAG (retrieval augmented generation) comes in. Spurgin describes it as "kind of like a librarian for your LLM."

When you ask a question, the system converts your query into what's called a vector embedding—basically a mathematical representation of what you're asking. It searches through a database of your actual company data to find relevant information, then feeds that context to the LLM. The result is supposed to be both conversational and grounded in truth.

"The LLM helps with reasoning and context while the RAG component is grounding us using real data," Spurgin explains. "And in the end we get accurate insights."

The use cases he outlines sound legitimately useful. A sales leader asks about Q4 forecast variance by region, and instead of spending an afternoon in Excel, gets back: "Your Q4 forecast is 8% below target, mainly due to delayed renewals in the Southeast. Competitor discounts are impacting your top three accounts."

Or a manager asks for customer sentiment from last week's social media, and the system responds: "Overall sentiment is 76% positive. Customers report they are loving the new loyalty program, but there is a major negative theme around shipping delays."

These aren't hypotheticals plucked from a marketing deck—they're the kind of questions people actually need answered, right now, without having to become data scientists first.

The Complications Nobody Wants to Talk About

But here's where my reporter brain starts asking uncomfortable questions.

First: data access. Spurgin acknowledges that "conversational BI only works as well as the data that it can and should reach." Cool. But who decides what data the AI can access? And crucially, who decides who can ask what questions?

If there's a database with salary information, yeah, obviously not everyone should be able to query that. But what about the gray areas? What about competitive intelligence? Customer complaints? Internal communications about strategy? The question of who gets to ask what reveals power structures that organizations might not want to make explicit.

Then there's governance and security. Spurgin uses this car analogy: brakes aren't for stopping, they're for going fast safely. Fine. But that analogy kind of assumes we all agree on what "fast" means and where we're trying to go. In reality, different stakeholders have different risk tolerances and different definitions of "safe."

And then—maybe most importantly—there's bias. "LLMs are very powerful, but they reflect the data that they're trained on," Spurgin notes. "And that means they can pick up on our biases."

This is the part that deserves way more than the thirty seconds it gets in the video. If your historical data shows that your company has consistently promoted certain types of people over others, and your conversational BI learns from that data, what happens when someone asks it to recommend candidates for promotion? Does it perpetuate those patterns? Almost certainly, unless there's active intervention.

Spurgin suggests "transparency, diverse data, and humans in the loop to review." All good ideas. Also: incredibly vague about implementation. Who are these humans? What authority do they have? What happens when the AI's recommendation conflicts with what the human reviewer thinks is right?

What Actually Changes

One thing I find genuinely interesting: Spurgin emphasizes that you don't need to rebuild your entire BI infrastructure to do this. Conversational BI can integrate with existing data warehouses and visualization tools. That's important because it means adoption doesn't require a total organizational overhaul—always the thing that kills promising technology.

But integration with existing systems also means integration with existing problems. If your current data is siloed, inconsistent, or just plain wrong, conversational BI doesn't fix that. It just makes it easier to ask questions that surface bad data faster.

The real shift here isn't just technical—it's about who gets to interact with organizational knowledge. When insights require specialized skills in SQL or data visualization, you create information gatekeepers. When anyone can ask natural language questions and get real answers, you potentially democratize access to knowledge.

Potentially. If the system is designed with genuine accessibility in mind. If the governance structures don't just recreate old hierarchies in new forms. If the bias mitigation actually happens instead of being a checkbox on a compliance form.

The technology exists. The question is whether organizations can handle what happens when everyone can ask anything.

---

Zara Chen is a Tech & Politics Correspondent for Buzzrag.