The Fabric Data Agent Goes GA

Microsoft is folding AI into nearly every part of Fabric, and the Fabric Data Agent is now generally available after a long run in preview. Brick and Landon talk through what they've found in testing: where natural language querying holds up, what it takes to get consistent answers, and why we've been adding a platinum stage to the medallion architecture with context files that tell the model how the business actually works.

Landon also covers Microsoft's open-source Fabric skills, the PBIP and PBIR report format that finally puts Power BI into source control, and the validation loop where the AI generates a report, screenshots it, catches its own broken visuals, and corrects them. Both come back to the same conclusion: the results depend on a strong data foundation and someone who knows their way around Power BI.

The Dashboard Effect is hosted by Blue Margin. We build the data foundation that mid-market and PE-backed companies run their reporting and AI on. 

Learn more: https://bluemargin.com

Transcript

Copy Transcript Copied

SPEAKER_01 0:05

Welcome to the Dashboard Effect Podcast. I'm Brick Thompson. And I'm Landon Oaks. Landon, I wanted this morning to talk a little bit about fabric and some of the AI enablement that's happening there, especially in the world of BI. We use Fabric to build our data lake houses. Obviously, our Power BI semantic layers live there. There's all sorts of interesting things happening there. Maybe we can just talk through some of the things that have been coming out over the last few months.

SPEAKER_00 0:32

Yeah, for sure. So you know, Fabric's been as we've said before, it's changing a lot, right? Microsoft's putting all of its all of its time for the analytics piece, data engineering side, into fabric. That's where they're pushing people. So yeah, they're they're churning out new features pretty often. Um some of the big ones that they're they're really pushing now is really just AI. So they're trying to fold AI into almost anything in Fabric with their co-pilot kind of agent there. Trevor Burrus, Jr.

SPEAKER_01 1:04

And are you talking about when you say agent, one of the first things we saw was that fabric data agent.

SPEAKER_00 1:09

Yeah.

SPEAKER_01 1:10

I think. And that first came out several months ago.

SPEAKER_00 1:12

Aaron Powell Yeah, I think more than that even.

SPEAKER_01 1:14

Many months, or a year ago, maybe. Yeah.

SPEAKER_00 1:15

Yeah, it was in preview at the time. So we played with it quite a bit in preview. Um it's now generally available, which is a good thing, because preview, you know, you always have that asterisk. Will it work? Will it break? It's in preview, right? Um so now that it's GA, it's it's a good thing. Um But yeah, it's essentially a way to connect directly to your data in Power BI, whether that be like your model behind a Power BI report, a lake house, a warehouse, et cetera.

SPEAKER_01 1:42

It can kind of go connect to it and then you can ask national language questions, you can you can build agents on top of it, that type of thing. We I think we found in our testing over the last year that um it struggled a bit to consistently give good answers until you build a really solid foundation. So you can't just have sort of a messy uh data model and hope that it'll go figure it out. It's pretty smart, but it'll also give you very confident bad answers or incomplete answers as well. And so that brings us back again to this idea of you need a strong data foundation to have um uh LLMs and agents do a good job with it. So we've been building something called this uh platinum layer, um, staying in that medallion architecture, you know, the bronze, silver, gold, uh now platinum, where we do a bunch of things that we've talked about in other episodes. Maybe one of the most important things, though, is adding that context, text files, markdown files, to the data lake house so that when the LLMs are working on uh getting you an answer, they have as much to go on as you can give them about the business and about the users and different things like that.

SPEAKER_00 2:57

Yeah. Yeah, it's uh pretty critical to have good, good data. It's the same as like our MCP server, fabric data agent, you know, other other companies have their own kind of flavor of that that will go out and do NLQ. Um you just throw it at the data, it's you know, it's gonna get confused. Like an example for us was we have utilization, and sometimes our data can get a little bit confusing in terms of like how much is this engineer working on a client versus not? Um I threw at the question, right? It came back and it counted all kinds of random time entries that shouldn't necessarily be marked as client work, you know. Um but as soon as I gave it the context, I actually baked some of that into the to the back-end data. You know, it started returning good results.

SPEAKER_01 3:41

Yeah, yeah. So I I think everybody's discovering this now. This uh this idea of connecting to sort of six or eight different systems through their MCP servers and hoping that the LLM will stitch it together well just gives very um inconsistent results. You can definitely connect it to one system, and if the MCP server is built well and the LLM knows a lot about that system, you can get really good answers. But then trying to pull different um sources of data together can get challenging. And that's where fabric can be really useful. Yeah, building that one lake house and then adding those context files and that type of thing. There was another thing that we talked about right before we started recording, which is the open source fabric skills. Yeah. So things like uh things that will help you build Power BI reports and so on. Um obviously we do a ton of Power BI creation, and you've been doing some testing on that. Maybe you can tell us just a minute or two about that.

SPEAKER_00 4:39

Yeah, for sure. I want to do more. So hopefully we'll have more to talk about in a bit. But um essentially it's it's a way if you save Power BI in this kind of new format they're pushing for those developers out there you'd know is PBIX was the file for Power BI. Now you want a PBIP and a PBIR or something. I can't I don't know if I got that one wrong. Yeah, sorry. But um the exact the exact letter, I don't remember for sure. But um it's kind of two different sections, a model and a visual layer for your your report, and that allows AI to actually work with it, edit it, it allows you to put it into source control. So a lot of the things that uh Power BI struggled with in the past were fixed with that. Um and so yeah, they're they're skills that Microsoft release, you just throw it directly into Cloud or ask whatever AI you're using to reference those skills, they are able to read it. It gives gotchas and strategies on how to actually model and visualize different data pieces. It provides a uh feedback like validation loop. So it'll make its change, it'll generate the report and take a screenshot of it, look at those screenshots and see like, oh, I messed up here. This visual has a broken, you know, reference on it, it'll go fix itself automatically, um, which has been really cool. That's really cool.

SPEAKER_01 5:58

Yeah. I think our I know our engineers are using it also to help just with the semantic model, um, going in and helping to debug if there's something that doesn't seem to be working right or performance isn't where we want it to be. And I think we're getting good results with that now as well.

unknown 6:13

Yeah.

SPEAKER_00 6:13

It's helping a lot. Helping a ton. Yeah. And that that's you know, that's time consuming sifting through DAX and models, especially for optimization for like performance. That can, you know, that can take a day, a week. Yeah. But uh Cloud's really good at it.

SPEAKER_01 6:27

Yeah. You still need the strong data foundation. And you also need someone who understands BI to get a really good result, uh, at least at this point. I mean, uh in in years to come, I'm sure that it'll be able to do all of that. Um, but it is pretty amazing what you can do for someone who understands their way around Power BI and what's going on there and how to build a good, good report, those types of things. Yeah.

SPEAKER_00 6:50

Yeah. And I think, you know, that's kind of a key point there that you just raised. Like you have to kind of shepherd it and, you know, point it in the right direction. It's going to go off track. You got to push it back on track. Um so if you have the experience and you're able to do that, it's insane what you can do. Um, but you do need to need to know where when it's going off track and how to get it back on track, just with your knowledge of having done it yourself before.

SPEAKER_01 7:13

Yeah, there's that. And you need to have a good data platform team providing a good foundation for you to build on. Anyway, I I feel like I'm kind of beating that drum to death, but I'm just seeing it more and more when people have AI projects that are failing that are connecting to data. Very often the data foundation is is the proximate or root cause. So still need to be building that piece. And the AIs can help with that a lot, uh, but you still need to be smart and bring it into the context of your business. Make sure you're uh educating the AI about how you view certain metrics, how you use certain systems, um, how different uh cohorts in your company, when they're asking about something, what do they actually mean? So that you're giving them the right answers.

SPEAKER_00 7:55

Exactly. Yep.

SPEAKER_01 7:56

All right. Well, anything else you want to cover there?

SPEAKER_00 7:59

No, I think that's it.

SPEAKER_01 8:00

Okay.

SPEAKER_00 8:01

Thanks. Awesome. Thank you.