Fabric Data Agents: Unlocking The Full Power Of DAX For Data Analysis

Now that Fabric Data Agents (what used to be called AI Skills) can use Power BI semantic models as a data source I’ve been spending some time playing around with them, and while I was doing that I realised something – maybe something obvious, but I think still worth writing about. It’s that there are a lot of amazing things you can do in DAX that rarely get done because of the constraints of exposing semantic models through a Power BI report, and because Data Agents generate DAX queries they unlock that hitherto untapped potential for the first time. Up until now I’ve assumed that natural language querying of data in Power BI was something only relatively low-skilled end users (the kind of people who can’t build their own Power BI reports and who struggle with Excel PivotTables) would benefit from; now I think it’s something that will also benefit highly-skilled Power BI data analysts as well. That’s a somewhat vague statement, I know, so let me explain what I mean with an example.

Consider the following semantic model:

There are two dimension tables, Customer and Product, and a fact table called Sales with one measure defined as follows:

Count Of Sales = COUNTROWS('Sales')

There’s one row in the fact table for each sale of a Product to a Customer. Here’s all the data dumped to a table:

So, very simple indeed. Even so there are some common questions that an analyst might want to ask about this data that aren’t easy to answer without some extra measures or modelling – and if you don’t have the skills or time to do this, you’re in trouble. One example is basket analysis type questions like this: which customers bought Apples and also bought Lemons? You can’t easily answer this question with the model as it is in a Power BI report; what you’d need to do is create a disconnected copy of the Product dimension table so that a user can select Apples on the original Product dimension table and select Lemons on this new dimension, and then you’d need to write some DAX to find the customers who bought Apples and Lemons. All very doable but, like I said, needing changes to the model and strong DAX skills.

I published my semantic model to the Service and created a Data Agent that used that model as a source. I added two instructions to the Data Agent:

• Always show results as a table, never as bullet points
• You can tell customers have bought a product when the Count of Sales measure is greater than 0

The first instruction I added because I got irritated by the way Data Agent shows the results with bullet points rather than as a table. The second probably wasn’t necessary because in most cases Data Agent knew that the Sales table represented a sale of a Product to a Customer, but I added it after one incorrect response just to make that completely clear.

I then asked the Data Agent the following question:

Show me customers who bought apples and who also bought lemons

And I got the correct response:

In this case it solved the problem in two steps, writing a DAX query to get the customers who bought lemons and writing another DAX query to get the customers who bought apples and finding the intersection itself:

At other times I’ve seen it solve the problem more elegantly in a single query and finding the customers who bought apples and lemons using the DAX Intersect() function.

I then asked a similar question:

For customers who bought apples, which other products did they buy?

And again, I got the correct answer:

In this case it ran five separate DAX queries, one for each customer, which I’m not thrilled about but again at other times it solved the problem in a single DAX query more elegantly.

Next I tried to do some ABC analysis:

Group customers into two categories: one that contains all the customers with just one sale, and one that contains all the customers with more than one sale. Show the total count of sales for both categories but do not show individual customer names.

And again I got the correct answer:

I could go on but this post is long enough already. I did get incorrect answers for some prompts and also there were some cases where the Data Agent asked for more details or a simpler question – but that’s what you’d expect. I was pleasantly surprised at how well it worked, especially since I don’t have any previous experience with using AI for data analysis, crafting prompts or anything like that. No complex configuration was required and I didn’t supply any example DAX queries (in fact Data Agents don’t allow you to provide example queries for semantic models yet) or anything like that. What does this all mean though?

I’m not going to argue that your average end user is going to start doing advanced data analysis with semantic models using Data Agents. The results were impressive and while I think Data Agents (and Copilot for that matter) do a pretty good job with simpler problems, I wouldn’t want anyone to blindly trust the results for more advanced problems like these. However if you’re a data analyst who is already competent with DAX and is aware that they always need to verify the results they get from Data Agent, I think this kind of DAX vibe-coding has a lot of value. Imagine you’re a data analyst and you’re asked that question about which products customers who bought apples also bought. You could search the web, probably find this article by the Italians, get scared, spend a few hours digesting it, create a new semantic model with all the extra tables and measures you need, and then finally get the answer you want. Maybe you could try to write a DAX query from scratch that you can run in DAX Studio or DAX Query View, but that requires more skill because no-one blogs about solving problems like this by writing DAX queries. Or you could ask a Data Agent, check the DAX query it spits out to make sure it does what you want, and get your answer much, much faster and easier. I know which option I’d choose.

To finish, let me answer a few likely questions:

Why are you doing this with Fabric Data Agents and not Power BI Copilot?

At the time of writing Data Agents, the Power BI Copilot that you access via the side pane in a report and Power BI Copilot in DAX Query View all have slightly different capabilities. Power BI Copilot in the side pane (what most people think of as Power BI Copilot) couldn’t answer any of these questions when I asked them but I didn’t expect it to because even though it can now create calculations it can still only answer questions that can be answered as a Power BI visual. Copilot in DAX Query View is actually very closely related to the Data Agent’s natural language-to-DAX functionality (in fact at the moment it can see and use more model metadata than Data Agent) and unsurprisingly it did a lot better but the results were still not as good as Data Agent. Expect these differences to go away over time and everything I say here about Data Agents to be equally applicable to Power BI Copilot.

This isn’t anything new or exciting – I see people posting about using AI for data analysis all the time on LinkedIn, Twitter etc. What’s different?

Fair point. I see this type of content all the time too (for example in the Microsoft data community Brian Julius and my colleague Mim always have interesting things to say on this subject) and I was excited to read the recent announcement about Analyst agent in M365 Copilot. But typically people are talking about taking raw data and analysing it in Python or generating SQL queries. What if your data is already in Power BI? If so then DAX is the natural way of analysing it. More importantly there are many advantages to using AI to analyse data via a semantic model: all the joins are predefined, there’s a lot of other rich metadata to improve results, plus all those handy DAX calculations (and one day DAX UDFs) that you’ve defined. You’re much more likely to get reliable results when using AI on top of a semantic model compared to something that generates Python or SQL because a lot more of the hard work has been done in advance.

Is this going to replace Power BI reports?

No, I don’t think this kind of conversational BI is going to replace Power BI reports, paginated reports, Analyze in Excel or any of the other existing ways of interacting with data in Power BI. I think it will be a new way of analysing data in Power BI. And to restate the point I’ve been trying to make in this post: conversational BI will not only empower low-skilled end users, it will also empower data analysts, who may not feel they are true “data scientists” but who do have strong Power BI and DAX skills, to solve more advanced problems like basket analysis or ABC analysis much more easily.