Will AI replace BI?
Consider this: The cloud didn’t kill the data warehouse. Self-service didn’t kill the dashboard. Donald Farmer, who’s watched decades of this cycle repeat, put it to me plainly on the Data-Driven podcast: every generation of data technology declares the last one dead, and the last one never really dies.
AI is going to let BI finally finish the job it started. Showing you what happened was the old job. Acting on it is the new one. An AI agent, cowork, or cocode, is only as trustworthy as the semantic model underneath it. Get your definitions straight before you let agents act on them, and you win the next decade of analytics. Get them wrong, and you’re making bad decisions at scale.
Old Tech Just Gets a Smaller Job
Donald has spent more than three decades in data and analytics. First at Microsoft and Qlik, and now as Principal at TreeHive Strategy and VP of Innovation at Nobody Studios. He’s seen this pattern before: “New technologies offer new capabilities,” Donald told me, “but the older technologies still hang around.”
Here’s a field example. Do you really want 5,000 employees asking the same question to an agent every morning? A dashboard answers that faster than any conversation could and at a significantly lower cost. Routine reporting becomes the layer that handles the questions everyone already knows to ask, which frees up the harder work for something new.
The Real Shift Is In Detective Work
Revenue drops 8% in a region. Someone opens Tableau. Was it a product? A promotion? Pricing? Inventory? For twenty years, a human has run down that tree by hand, branch by branch, hypothesis by hypothesis.
The new workflow skips the manual hunting and pecking. You ask an LLM: “Show me every product whose revenue dropped, and explain why.” Now the AI is performing the analysis rather than just answering the query.
Here’s what that means in practice: AtScale’s customer Blue Yonder connected directly to a governed semantic layer via MCP. A financial deep-dive analysis that previously required 30 hours of work was reduced to 90 seconds using an AI workflow.
Action Was Always the Hard Part
Donald walked through a four-stage model of how people use data: orientation, glimpsing, examination, and action. BI has never gotten past the fourth stage of the model.
“Business intelligence took you to the door,” he said, “and left you on your own.”
The challenge with analytics has always been with execution. Agentic AI is the first real bridge between analysis and action: the thing that walks through the door instead of just pointing at it.
Market research supports this: Gartner projects that by 2027, half of all business decisions will be augmented or automated by AI agents built for decision intelligence. IBM’s Institute for Business Value found that organizations already running agentic AI report faster speed to action and real gains in process efficiency.
An agent is only as trustworthy as the semantic model underneath it, and that’s exactly where it gets dangerous.
One Wrong Number Breaks Everything Downstream
The old idea of “semantics” was a single, locked-down number distributed everywhere. The new idea is relationships. Marketing, sales, and finance all define revenue differently, legitimately, and none of them is wrong. An AI agent needs more than the definition it’s been given. It needs the reasoning behind it and the conditions under which it applies.
That context is exactly what a semantic model carries. And the smarter AI gets, the more expensive inconsistent meaning becomes. An agent that acts autonomously on the wrong definition makes bad decisions, at scale, with nobody in the loop to catch it.
AI Changes Who Builds the Model
Building a semantic model used to require a unicorn: someone who understood the business, the schema, and SQL, all at once. Those people are rare, and that scarcity has been a real bottleneck for the entire industry.
“The people who can do that,” Donald said, “are now the people with business knowledge.”
I’d push that further: this isn’t AI taking analyst jobs, it’s AI making more of them. AI writes the SQL and builds the model scaffolding. That frees the analyst to do the part only a person can do, which is ask better questions and actually explore the data, instead of hunting through Tableau or Power BI one hypothesis at a time. Give an analyst that kind of leverage and they can do more of that work at a scale no single person could manage by hand.
One more thing Donald flagged: no enterprise runs its data on one vendor’s stack. It’s Snowflake, Databricks, Microsoft, and half a dozen others, often at once. “Monolithic platforms are really not the way to go,” he said. An agent is only as good as its ability to reach consistent, governed logic wherever that logic lives, which is exactly the argument behind open standards work like Apache Ossie.
What to Check Before You Trust the Answer
Donald closed our conversation with a line I love: “Human beings are wonderful natural analysts.” AI scales that instinct.
BI was always good at the what: what happened, what hit target. The why used to require someone who knew a BI tool well enough to go hunting for it, one hypothesis at a time. AI closes that gap and turns anyone who can ask a question into someone who can ask why. That’s the democratization we’ve talked about for our entire careers.
Before you bolt another AI feature onto your stack, ask one question: does the model underneath it know why a number means what it means? Get that right, and you’ve done more than add AI to BI. You’ve finally finished what BI started.
Donald and I covered a lot more than fits here. Listen to the full conversation on the Data-Driven podcast.
Reviewed by: Mark Palmer
SHARE
Guide: How to Choose a Semantic Layer