Three years ago, virtually every conversation was about data and dashboards. Today, they’re all about AI. Almost every company wants to apply AI to their data, and preferably yesterday. That’s understandable – there’s something magical about being able to ask a question in plain English and get an answer right away.
But there’s one question very few people ask themselves before pressing the button: What exactly does the AI have to work with when it’s integrated into your business?
Because an AI assistant always answers. That’s the whole point of it. You ask about the contribution margin for your three largest customers last quarter, and it gives you a number. Quickly, articulately, with absolute certainty. Whether the number is correct is an entirely different matter – and this is where things get serious. Because while an incorrect formula in an Excel sheet was at least confined to one department’s spreadsheet, AI can now produce the wrong answer in seconds – neatly packaged – for anyone who asks. Without anyone being able to see how the number was arrived at.
Before we go any further, let's be clear: this isn’t AI skepticism. The technology is capable of a great deal and the possibilities are both real and vast. But they’ll only come to fruition if the foundation is in place. And before we even talk about the foundation, we need to agree on what AI actually is in a data context in 2026 – because the term has become so broad and vague that it has almost ceased to mean anything.
Five Things We Call by the Same Name
So let’s clear things up. Today, “AI on your data” covers at least five quite different things:
1. Chatting with your data. You type a question in plain language and get an answer: a number, a table, a graph. This refers to Copilot in Power BI and similar tools, and it’s the use case most people think of when they say “AI.” It feels like talking to a colleague who knows the numbers. The question is just whether that colleague actually does.
2. Augmented analytics. Here, the AI identifies the patterns on its own. You don’t just ask, “What was the DB in Q2?” but “Why did it drop?” and the machine points out the factors that dragged it down – a customer, a segment, a discount that got out of hand. This shifts the analysis from showing what happened to suggesting why, and that’s a bigger difference than it sounds.
3. Forecasts and “what-if” scenarios. The AI projects forward: sales forecasts, liquidity, demand. And it lets you play around with the assumptions – what happens to the bottom line if raw material prices rise by 10%? From an economist’s perspective, this is where things get really interesting, because this is exactly the kind of exercise that used to take weeks to complete in a spreadsheet (and was outdated the moment it was finished).
4. Generative AI for builders. The AI writes DAX, SQL, documentation, and model suggestions. It’s not aimed at the decision-maker, but at the developer – and it makes the work of building the solution significantly faster and cheaper. Much of the manual work that used to take days is now being reduced to minutes.
5. Agents. The latest step: AI that doesn’t just respond, but takes action. Think of it as a macro – but whereas the classic macro was confined to a single spreadsheet, the agent works across your systems. It retrieves data from one place, enriches it in another, generates the report and places it in the executive team’s inbox on Monday morning. All on its own, week after week. We’re still in the early stages of this development, but that’s the direction things are heading.
Five categories with vastly different levels of maturity, but with one common trait. And here, I’m afraid, I’ll have to dampen the mood a bit.
What Very Few People Talk About
Each of these five categories stands or falls on something no one thinks about when they talk about AI: whether the machine understands what your data means or not.
Let’s be specific. Most companies’ “data platform” is, in reality, a data warehouse. The tables from the ERP system, the CRM, the online store and the payroll system have been copied more or less as-is into a shared layer, and then they’ve called it a platform. But there’s no translation. “Customer” appears in five tables with five different definitions. Revenue is a journal entry, not a business figure – and nowhere is it written that the three accounts must be deducted, or that intercompany transactions must be eliminated, or that returns belong in a different period. It’s all in the controller’s head. Or in a formula in his/her Excel spreadsheet.
This kind of thing deserves a name: a spaghetti platform. It looks impressive in an architecture diagram and it has cost a great deal of money, but it contains no business logic. It’s storage, not intelligence.
And now we’re putting AI on top of it. What happens then? It reads the tables exactly as they are – without knowing that “DB” in one and “contribution margin” in the other are the same thing. Without knowing the three accounts that need to be excluded, without having a clue that one region does its bookkeeping differently from the others. It guesses. And it guesses with complete conviction, because that’s how language models are built: they don’t hesitate.
The result isn’t that the AI says, “I don’t know.” The result is a number. A nice, confident, incorrect number – delivered to an executive who makes a decision based on it.
AI Doesn't Reduce the Problem. It Amplifies It.
Here’s the point that every leadership team should take to heart: We’ve always had multiple versions of the truth. That’s why a management meeting might start with twenty minutes spent agreeing on which number to use. That’s the price we pay because each department has built its own logic into its own spreadsheet – since the platform never gave them the answer.
AI doesn’t solve that problem. Without a foundation, it makes it worse. It produces versions of the truth faster, formulates them more convincingly and makes the process invisible – because whereas you could previously open the Controller’s spreadsheet and see the formula, no one can open the language model and see how it arrived at the number. We’re moving from a problem we could at least review to one we can’t. And worst of all: an agent built on a flawed foundation doesn’t just make a mistake once – it automates the error and sends it out, week after week.
What’s needed is what the industry calls a semantic layer: the model that sits between the raw tables and the user, translating data into business concepts. Where “customer” has a single definition. Where revenue is calculated the way finance and sales agree to calculate it. Where the rules for what counts and what doesn’t are written in one place – and thus apply to everyone, including AI. It’s not just a technical embellishment. It’s where your company’s language lives.
And that’s exactly what a spaghetti platform lacks.
Why It's Not an IT Issue
This leads to the point that this article is really about – and which the next articles will elaborate on.
IT can build the platform. IT can integrate AI. But whether AI gives you answers you can trust depends on whether someone has done the tedious, time-consuming work of embedding the business logic into the model. No IT department can do that work alone, because they don’t know the business well enough. And the business can’t do it alone either, because they don’t know the data well enough. It lies in the tension between the two. That’s the whole reason why BI has never been, and never will be, an IT project.
So when the next vendor promises you "AI for your data", don’t ask them what AI can do. It can do a lot. Ask them what it has to work with. And if the answer is a spaghetti-like platform, then AI isn’t a shortcut to better decisions. It’s a faster path to the wrong ones.
Written by Lars Taagaard Christiansen
This is the first in a series about why so many data platforms don’t deliver what the business needs – and what it takes to make that happen. In the next articles, I’ll dig a little deeper into the spaghetti platform, the tension between IT, data, and business, and the semantic layer that ties it all together.
