DATA MODELS AND AI

FabCon is the European conference on Microsoft Fabric. It is organized by ESPC in close collaboration with Microsoft, and a large part of the content comes from the Fabric and SQL community. Much was of course about AI, Copilot and agents. 

With so many sessions, my payoff is colored by what I myself chose to go for. Others with a different focus have probably taken a slightly different angle home. For me, the point about the data base was the one that went again. 

Quick response, great confidence 

“What was the turnover last quarter?” 

It was one of the simplest examples in the sessions I attended, which is precisely why it hits so well. To answer correctly, AI must know your definition of revenue. It needs to understand which calendar you are using, choose the right date and know if data is even up to date. 

If one of those things is missing, you will get the wrong answer. Delivered quickly and with great confidence. 

When AI answers incorrectly, it is tempting to turn the prompt or change the model. The problem may as well lie in unclear definitions, in the date AI chooses, or in data that is not updated. And a better prompt does not correct them. 

The semantic model is gaining new users 

How many different figures for revenue are there in your reports today? 

The semantic model is the layer between your data and your reports. This is where calculations, relationships and names are defined so that all reports calculate in the same way. 

Until now, we have primarily built semantic models for Power BI reports and Excel. Now the same models must also be used by Copilot, by Fabric Data Agents, who are agents you can ask questions about your own data in plain language, and by applications and other AI solutions. 

This makes measures, i.e. the fixed calculations in the model, even more important. The same applies to relationships, naming, descriptions and security. If you have already decided how revenue and contribution margin are to be calculated, AI must use the calculations you have. It must refrain from piecing together a calculation of some columns that look relevant. 

When the model lacks the context 

A semantic model contains calculations and relationships. It rarely contains the entire company's context. 

It can tell how revenue is calculated. What you mean by an active customer, what date applies when someone writes “last quarter”, and what business rules need to be taken into account, it doesn't necessarily say. 

Who among you would be able to answer clearly what an active customer is? 

This is where Fabric IQ and ontologies come into the picture. In short, an ontology is a common description of what things in the company are and how they are connected. The idea is to bring together your concepts, relationships, rules and measurements in one layer, which can be reused across Power BI, Copilot and various AI agents. The ontology in Fabric is still in preview. 

In other words, AI must be able to do more than find data. It must also understand what data means in your particular company. 

Read also: AI is management, about why AI requires management to decide how the organization works. 

The report will be joined 

Power BI continues to be an important channel for reporting and analysis. But the direction is that the semantic model also becomes the foundation for dialogue, applications and actions. 

Users do not necessarily have to open a report and look for the answer themselves. They can ask in plain language, work with data in an application, or have an agent respond when a particular situation occurs. 

As one of several ways of using your data and your business logic, the report consists. And all the ways draw on the same model. 

The star chart still holds 

I was also at a session on practical data modeling beyond the classic star scheme. The star chart is the classic structure with one central table with numbers, e.g. sales, and a number of lookup tables around it with customers, goods and dates. 

It is still a good starting point. Reality doesn't always fit nicely into it. This applies, for example, when data is available at different levels, in the case of historical changes, in the case of snapshots and in the case of budgets that are not at the same level as the actual sales figures. 

The point was to stick to the good modeling principles and understand the problems well enough to choose the right solution. It is at least as relevant when AI later has to use the model. 

Who remembers the latest change? 

Who among you knows exactly what was last changed in your most important data model? 

In several of my sessions, there was also a focus on Git, i.e. a system that keeps track of each change, and on choosing the right way to integrate the solutions into Fabric. 

When Power BI and Fabric become the foundation for both reporting, AI and operational solutions, it is no longer sufficient that one developer has the latest version lying around locally and can roughly remember what was changed. The solutions must be able to be maintained, tested and put into production safely. 

Where can you start? 

My most important message is that you don't have to use all the new options right away. Some are still new or in preview. 

But you can start to get the foundation ready, regardless of whether your data today lives in Power BI alone or on one data platform in Fabric: 

• Review the semantic models and the measures you have 

• Ensure common and clear business definitions 

• Give tables, columns and measures good names and descriptions 

• Check data quality and security 

• Get more structure on development, documentation and version control 

• Select specific areas where AI or automation can actually solve a problem 

So you can't skip all the hard work and go straight to the intelligent part. 

So maybe your first real AI project will look like a thorough cleanup in your semantic model. 

Jannik Raun Christensen is Senior Consultant in KOLLAB and works with Power BI, Microsoft Fabric and data modeling.