MCP Your Fabric Data – Ep. 494

Mike and Tommy unpack Model Context Protocol (MCP) servers for Power BI and Fabric — what they are, how to use them, and whether organizations should be turning them on today. Plus, a deep dive on what makes AI agents actually useful versus just impressive.

News & Announcements

100x a Business with AI — Lessons Learned

Mike shares an article from X on the realities of making AI agents effective in business. The key lessons: context is everything, agents multiply outcomes but still need human guidance, memory and state matter, and you need to catch exceptions early. The standout quote: “Agents don’t eliminate the need for human judgment — they eliminate the friction around human judgment.”

Tommy draws a parallel to Power BI adoption: organizations expect instant results, but there’s real organizational build time required — approvals, onboarding, validation. AI agents are in that same stage: people see ChatGPT work effortlessly and wonder why their enterprise agent can’t be productive in two months.

Mike’s hot take on dashboards: “Dashboards are useless” — unless they drive resolution, not just visibility. Every dashboard should have a means to an end. An agent tied to your data should identify broken things, surface next steps, and propose solutions — not just show you’re down 10%.

Google NotebookLM for Podcast Knowledge

Mike experimented with Google NotebookLM by uploading the last five podcast episodes as audio. The standout feature: every AI response includes annotations that link directly to highlighted sections of the transcript source material. Mike envisions this as a center of excellence tool — bundle Microsoft documentation, internal training videos, and organizational reports into an agent that can answer questions with sourced citations.

Tommy highlights the podcast generation feature, where NotebookLM creates a 5-10 minute AI-generated conversation from any uploaded content. Both agree Microsoft needs to build equivalent capabilities into their M365 ecosystem rather than forcing users to third-party tools.

Main Discussion: MCP Servers for Power BI and Fabric

What Is an MCP?

Model Context Protocol — an integration into an AI tool (VS Code, Claude, any IDE or chatbot) that provides additional commands and context. Tommy breaks it down: instead of explaining to an AI what TMDL is or how to execute a DAX query, the MCP server provides all that knowledge, commands, and context automatically.

Microsoft has released three MCP servers in public preview:

Power BI MCP Server (remote and local versions) — Manages semantic models, executes DAX queries, updates columns/measures/relationships
Fabric MCP Server — Broader Fabric workspace and item management
Real-Time Intelligence MCP Server — Eventhouse and KQL query support

How Mike and Tommy Are Using Them

Mike’s workflow: Connecting to semantic models to organize measures, create calculation groups, and manage properties. Example: “Hey agent, what property do I set to turn off all implicit measures and only use explicit measures?” — the agent knows the property and sets it directly.

Tommy’s approach: Using the LLM layer for semantic analysis, not just execution. Examples: “Look at my descriptions — which ones don’t fit with the rest of the model?” or “Are all my dimensions distinct? What am I missing? Give me suggestions.” The key insight: you still have a full LLM underneath — it’s not just an automation tool, it reasons about your model.

Both see huge potential for model auditing: which measures are used across reports, which columns are unnecessary, which tables should be dims vs. facts. People have built entire paid tools for this — MCP servers could do it conversationally.

Who Is This For?

Mike’s position: Medium to senior developers. You need to know what you’re trying to build — the MCP makes the tedious work disappear, but you still need the right words and concepts. Getting started requires VS Code, installing the MCP server, and configuring connections — that’s a cliff for beginners.

Tommy’s counterpoint: The chat interface actually screams beginner-friendly. An intern or data steward could say “organize all my models, help me with descriptions and measures” without knowing the underlying complexity. The barrier isn’t the MCP itself — it’s the setup.

Where they agree: The setup friction is the real problem. Mike’s vision: MCP should be built directly into Power BI Desktop — a new pane where you pick your LLM (Copilot, GitHub Copilot, AI Foundry, or any API endpoint) and just start chatting with your model. No VS Code required, no manual installation. “At that point, you don’t even need to call it an MCP server. It’s just ‘chat with my model.’”

Is It Ready for Organizations?

The verdict: Not yet. Both agree this is early-stage, public preview, and not production-ready for enterprise deployment.

Security considerations Mike flags:

When using Claude or ChatGPT directly, your prompts go to their servers — anything sent could be used for training
GitHub Copilot and Azure AI Foundry keep prompts within Microsoft’s infrastructure and don’t use them for model training
Mike only uses VS Code with GitHub Copilot for any client-related MCP work — never sends client data to third-party LLM APIs directly
Microsoft recommends organizations do their own security review before any MCP integration

Tommy’s take for CTOs: Keep one eye on it, set up alerts for updates, but don’t plan a 2026 rollout. It’s still a novelty — exciting and powerful, but the security, governance, and ease-of-use aren’t enterprise-grade yet.

Kill It or Keep It?

Unanimous: Keep it. Both want Microsoft to keep evolving MCP servers. The potential is clear — it’s the friction that needs to disappear. Mike predicts adoption will stagnate if the learning curve doesn’t flatten dramatically. The technology is sound; the packaging needs work.

Resources & Links

Introducing Fabric MCP (Public Preview) — Microsoft’s announcement of the Fabric MCP server, an open-source framework that gives AI agents context about Fabric APIs, item definitions, and best practices. Runs locally on your machine with you in control of when generated code executes.
Power BI MCP Servers Overview — Microsoft Learn documentation covering the Power BI MCP servers (remote and local), how they enable AI assistants to interact with semantic models, and setup instructions.
Real-Time Intelligence MCP Overview — Documentation for the RTI MCP server, enabling AI agents to interact with Eventhouse and execute optimized KQL queries through natural language.

Episode Transcript

Full verbatim transcript — click any timestamp to jump to that moment:

0:21 [music] Good morning and welcome Back to the Explicit Measures podcast with Tommy and Mike. Hello everyone and welcome back to the show. Good morning, Mike. How you doing? It’s a cold day here in the Midwest. It is. We We had had some It’s been a very weird start of the year or new year for us, I guess. It’s been hot, cold, hot, cold. We’re getting down to the cold side again. So, I’ll take it. We’ve been taking our family skiing a couple times. we have three snowboarders and two skiers in our

0:53 Family. So, we’ve been doing a bit more outside activities. That’s been super fun. I enjoyed that. definitely did something to my back. was jumping off a lift and fell down and like definitely tweaked my back. So, we’re back and recovered again and let’s do it again. Let’s go punish ourselves for another weekend. it’s funny, Mike. This is the time I You understand what you would say to your kids or what your parents used to say to me. Today, I walked my kids up to the bus stop, 10 degrees, there’s ice everywhere, so they got to walk and they’re all, walking like it builds character. So, , there’s a there’s the Wisconsin

1:28 Northwest comedian who who does all those talk about Charlie something. His name is Charlie something. , he did a really funny sketch the other day. He goes, “People in like I guess it was people in Wisconsin or something like that.” And he was pretty true. At this time when everything gets like that sheen of like it was warm and now it gets really cold and then everything gets that sheen of ice. You see everyone doing like a very weird walk. They do like the flat foot scoot walk, right? It’s just the you you don’t really move your feet very far. You just do this little shuffle and you keep your feet as flat as possible so you don’t

1:59 Slip on the ice. And so like you learn how to like walk on ice to some degree and you do this little flat walk shuffle wherever you go. Well, dare I argue that tells you if someone’s from the area or not. That’s true. Yeah. If you see someone trying to walk like normal near ICE, like, okay, these guys, they’re not from here. They’re not from around here. Every year though, it takes you like a couple really good falls to feel like, “Okay, all right.” [laughter] Like, just one mids slip though. Yep. You’re like, “Oh, yeah. It’s that time of year.” Okay. Yeah. I I can’t stand those stuff rolls across my shorts feed all the

2:33 Time. And every so often there’ll be a feed of, hey, watch all these people walk out their front door and they’re slipping on their front door steps and they’ve got things in their hands. They’re trying to leave their house or whatever. They’re trying going to work and they just slip on the step. I can’t watch it, Tommy. It gives it sends chills down my back. I’m like, this purpose, man, because it brings back so many bad memories of me maybe falling on [laughter] stuff. I just can’t stand it. Absolutely hate it. YouTube shorts, people who are recording themselves. First off, you think about this. I’m glad you’re doing YouTube shorts for the podcast because

3:05 I’m sure people have seen that because I’m getting to the point now, Mike, where I go through three shorts. I go, I’m done. Brain waste. I’m out going to think, “Wow, who’s recording that?” Somebody There’s no way that someone’s randomly recording the guy slipping. Oh, they are. It’s the Ring doorbell stuff now. So, it’s all it’s all the doorbell cameras that are catching all this stuff happening at your front door. It’s it’s now always there. Every home has a camera. You slip. It’s on Ring. Are you putting on YouTube? somebody is. Someone thinks it’s funny, I guess.

3:36 Are you going to you yourself? I wouldn’t. I I don’t have Exactly. Exactly. This tells me who I want. Yeah. If you do that, fine. I don’t know if we’re going to be good friends, but I I just don’t understand. I I see your point now. Like like who’s the person who’s actually got to spend the time to go through their Ring doorbell and say, “Hey, that was funny, right?” And then post it. Maybe people post all kinds of crazy things, Tommy. I know. That’s why. Good. You’re feeding the internet with good things with the podcast. That’s what I like. If you want to call it a good thing, I sure I’ll I’ll take it. That’s

4:08 [laughter] If you want to call it that, great. Better than people slipping on the ice. That’s true. That’s true. Let’s let’s add some good information. Let’s add some education. I will find though, if I had to look at my social media feeds, Tommy, they have 100% been inundated with anything that’s AI learning automation. Like right now, that’s all that my feed is. Is your feed do the same thing or is yours still like just a bunch of Yankee stuff that just keeps coming across the If I watch something for more than five seconds, whatever it is, that’s what

4:41 YouTube’s going to feed me. So I have to keep like thumbs down and a bunch of stuff. Yeah, it’s true. I’ll find like that’s interest. No, no, no, it’s not. YouTube thinks I like it. Do I want a whole bunch of, destruction derby videos? No, I don’t. Just down some that. Make sure even though you may really want to watch it, you you have to not do that. Ammani Python one came up like that’s funny. And that’s literally been like eight out of 10, which not complaining that much about that one. Anyways, we got a lot to cover today, Mike. I don’t even know what the topic is. So our main I didn’t even even talk about the main topic. So the main topic today is around

5:15 MCP servers. I you’ve heard this term a little bit. Let’s unpack what this thing called an MCP is and just understand like okay what is it? How does it work? How does this work with existing tooling that you have? Let’s just unpack the idea a little bit. And Microsoft has released a number there’s probably more coming but right now I believe there is three that are out in preview. We have a Microsoft Fabric MCP server. We have a PowerBI MCP server. And I think we have a real time intelligence MCP server.

5:48 What is this? What does this mean? How do we turn these things on? Where do they live? how do they do they make your life more effective or more complicated? Let’s let’s unpack that concept here. So, if you haven’t played with these tools before, this episode’s for you. We’re going to unpack that a little bit in lie of this whole AI conversation. I’m going to transition over here back to some news. Tommy, I’ve got a news article here and I’m going to go post this one here in the chat as well. 100x a business with AI. This is this is maybe misconceptions of business leaders and how they look at AI. And this fits, I

6:22 Think, very well with the MCP conversation. I’ll post this out here, Tommy. The gist of the article is, it takes effort to make agents effective in your business. Mhm. And I think there’s some expectations you just and I hear it. People just think I’m just going to throw an agent at something and it just works. It just goes. And I think the reality is there’s some lessons to be learned about how to implement agents well. And so this article goes through I don’t know if the Vas I don’t know his name particularly that’s his name

6:54 On online goes through three kinds of lessons. context is everything. , agents can multiply your outcomes, but you still need someone thinking with the agent, memory and state, how do you have the agents remember what they need to know? And then it talks through catching exceptions, , your economics of AI agents versus a generic software as a service and then deploying time. And then he gives the TLDDR at the very end of this of the the of the article wrap-up here. What are your thoughts on this article, Tommy? I think this is spot on where people are at with AI

7:29 Agent or the buzz term now I don’t think is no longer AI it’s agents in terms of everyone’s like oh we need an agent agents are the way to go the and the problem is to me the same that people have when they are adopting PowerBI for the first time let’s just build reports we’ll build semantic models and you realize to do that effectively can I build a semantic model in a day first time sure but if I’m going to deploy something that’s going to be utilized trusted in the organization and have other use cases be certified. We know it’s going to take more effort.

8:00 There is the actual I think human build time and called like organizational build time to get the approvals to get people on board to make sure that it’s actually doing what it’s supposed to and you you jump into an organization right now with PowerBI for the first time they’re going to be frustrated initially why things are taking longer than the concept the perception is going to be. And I think agents are right now in that same stage where it’s an agent, it’s AI. Well, I’ve seen how AI works as easily as it does for me in chat GPT. Why can’t we get our agent to

8:35 Be productive, efficient, and working in two months or less than? That seems like a terribly long time. But the problem the problem is not finding the solution. And I’ll break that down. I’ve said this before, but right now organizations are looking at AI agents both in I think fabric and generically as a solution in need of a problem and they’re trying to find problems for AI agents rather than the other way around. No one’s identifying the already made

9:10 Problems in their organization [music] that they have which they are there which they are there. They’re just saying, “Let’s get the AI agent and let’s figure out what it can solve after that.” And I think that’s the wrong way to look at it. Yeah. There’s a one statement here in the middle of the article. It’s in lesson number two, I believe, talking about agents multiply outcomes. And I feel like this is a representative as I’ve been exploring what agents can do, how they aid people in what they do. It said, “Let’s think about this the right way. Agents are going to replace human labor, bar

9:44 None.” And I would agree with that 100%. I see agents can do specific things that I’m not as good at, but I can maybe I can describe and and maybe to some degree, Tommy, like this is like the ability for me to program a computer using my native language, right? English. I can just talk to it and it can understand my context and do things. I can make images, I can make graphics, I can do a lot of things that would take me a lot of time to learn how to do. And I I know some things. I’m in Figma a lot. I build backgrounds and images and scrims for reports. So that’s stuff I I’m familiar with. But even using tools

10:16 To aid with that, I can outproduce. I can produce more faster. But the point he was making here was the positive is agents don’t need don’t eliminate the need for human judgment. They eliminate the friction around human judgment. So here’s all the details. Let me collect all the things for you and then from there you can then leverage that to become more effective. I think there’s a good analogy there, Mike, with I think the human need and what jobs it takes away or what work. Let’s say there was

10:50 An AI agent to build a dashboard. Well, that eliminates my work, but that doesn’t mean me as an analyst can’t doesn’t have to speak to it or understand what the information is still conveying. If I just had the AI agent build it, deploy it, and people ask questions, and I have no idea what it did, that’s a problem. Yes. So there is work that it eliminates, but that doesn’t necessarily mean it’s going to eliminate everything you need to do. So let let me go to So that’s a great lead in Tommy that you’re making there. Let’s lead into lesson number four in

11:23 This article, which I thought also was extremely relevant, which is catch the exceptions, right? So catching exceptions is not about catching obctions exceptions in code. It’s looking at what you do on a day-to-day workflow. Like where where are the things that you’re spending a lot of like manual effort doing stuff? And so he goes here, instinctively we look at AI systems to build and create dashboards. Well, dashboards are reporting or surfacing information. They don’t do anything. And he he makes a very clear point, and I would agree with this one. Dashboards are useless. You’re right. So having a bunch of dashboards that are

11:56 Out there, unless you can look at some information and know what to do next, there’s no really good point to put the dashboard in front of people. And he goes on to explain here, your finance team already knows there are missing receipts. Your sales team already knows what deals are stuck in the pipeline. What your agent should be doing is giving ideas, identifying what should be worked on next. , how do you get unstuck in that situation, right? This deal is stuck. we recommend based on your prior history, maybe you could try

12:29 This, this, and this or maybe if you pitch it this way or, that’s where the agent can surface additional next steps or actions to help them fix whatever it is as data comes into your system. Don’t just let data passively flow into your reporting. Catch the error as far up a pro as possible. And not only just catch the error and say, “Hey, there’s a problem. You’ve entered the data wrong in the in the system, right?” But say you’ve entered the data wrong. I’ve looked at the analysis here. I based on what I know and the agent, I recommend you may

13:02 Get this here. I’ve already prefilled it out for you. Click yes to accept those changes. That’s immensely valuable. Yeah. I am going to push back a little. I don’t know if it was you who said it or in the article. If it was good, it was me. If it was not, it was the article. Okay. Just kidding. I don’t know what you’re going to say. [laughter] that dashboards are useless. Dashboards are useless if they are the end. Not and I think what you said sums it up. Every dashboard should have a means to an end. An identical means to an end where the goal for a

13:34 Dashboard is to force resolution, not necessarily just to create visibility on the problems. Correct. And I think a lot of times we we we disillusion ourselves by just creating things that show data. Right. And and there’s no call to action at the end of it. Right. If your dashboard says you’re down 10% and you go okay now what that’s a problem you have should have a dashboard and theoretically an agent assigned to that tied to those problems. So this is where the surface area of this piece, this is where I think AI agents struggle. Right

14:07 Now when I look at what agents can do, I find when I look at the landscape of like things I’m seeing on social media, things that I’m playing with, how am I building software like I look at agents in and I don’t know if this is the right way to this probably won’t stay this way for long. It’s probably going to evolve into something new. But agents are really good at helping you create. The creation side of things is where agents do a good job. Let me give you a couple examples, right? Okay. I’ve seen a lot of people on Twitter, I

14:40 Guess it’s I keep saying Twitter, it’s I’m it’s stuck in my head. I’ve seen a lot of people on X and LinkedIn. They’ve been saying, “Look, I’m building SVG visuals. Look at me build this complex measure. Let me build this. of giving an image to something and it’s doing a good job of getting through the prompting and building out some visuals with AI and agents. So when I look at that I go wow that’s pretty impressive that they’re doing that and that’s aiding them building the product. What’s happening though is the agent is used on the creation side. Agent is expensive. It it costs money to run it. Once you have the

15:14 Thing that’s made it’s making things in systems that are already efficient, right? It’s using a PowerBI report. It’s using SVG. It’s using whatever code language you’re running, right? If you’re building a web application or whatever, it’s writing the code that you then run on the efficient system. And the reason I’m bringing this point up is I don’t think it’s valuable for agents to be continually hit over and over again to produce like the agent use of effort every single time, right? it makes more sense to have the agent create something that can be built and

15:49 Deployed and then that is what’s run efficiently in your system so that way you’re not paying the extra cost for the agent because Tommy there’s a race everyone’s building bigger data centers everyone’s building all these agent things the big three Microsoft Google and Amazon are all racing to build better bigger data centers and acquire chips it’s it’s energy and resource resource intensive to run the agents. So, what can we do to leverage the agents where they’re good at and I think that’s in the creation experience side.

16:22 I’m going to backtrack to our previous episode and one thing I I mentioned as homework that I think part solves that Mike is really this concept of agent skills because right now that creation side you’re feeding all the initial context to the agent. You’re introducing an idea or a topic to the agent. But if I actually have what we are getting to and we’re seeing this become a much more closer and or a sooner sooner reality is this idea of skills like an SVG skill a skill around the model context for

16:55 My organization where the context is already there and more importantly not just the context but what skills can do is provide the tools an agent can do like hey you’re an you’re a visual designer for powerbi you create custom visuals you can create SVG skills, you can use D3, you can use DANB. Oh, by the way, if you want to use DANB, literally, you can feed it a Python script in that skill folder that says all the things you need to do. So, to me, I think to your point of not necessarily right now, it’s all about

17:29 The creation. Yep. That’s minimal effort. I think if you’re just starting an agent right now, you’re doing minimal effort. If you set it up the right way, it can do the modification. It can do the review, it can do the quality review and even deployment or or enhancements I think but you have to you have to work on it and I think that goes back to the article there is an effort involved to make a agent full scope not just from creation but to help you in each capacity on a particular item or workflow.

18:00 I I would definitely agree with that. There there’s definitely effort there and again the thing is I’m you’re talking about like the creation and the skills and like how it has memory and things. So those are interesting. But where I think where I think the value from these agents will come from will be, using agents to help me create things. I’m on that again, I still feel like I’m not ready to throw a bunch of agent things at very beginner users, right? , the one thing that I would probably say that I’m really interested in here is there’s this there’s this whole concept now around like knowledge graphs and

18:32 Searching for things. There’s a lot of impressive things. So, Notebook LM from Google. I’ve just started playing around with this thing. Yeah, I haven’t played with Notion. I heard Notion is another really good one that’s also out there as well. But when I look at Notebook LM from Google, I I just put a I just trying it out, Tommy. I took the last five episodes of our podcast. I suck it in there as audio. I literally just uploaded the audio. Awesome. And then I gave it a URL of the podcast and called the whole notebook, like part of my podcast and said, “All right, let’s see what it does.” And it was really neat because I could ask

19:06 Questions about individual episodes. I could ask about specific fabric terms and ask about what the guest talked about. U I could ask questions about just general things. And so it started forming this knowledge and the part okay that was interesting. Like a Google search would return basically the same results. What really did it for me? Like what turned it into more of a really useful tool? It was the idea that every time it responded with an answer, it would have one or two or three annotations on every single statement it made. And when you clicked on those

19:40 Annotations, it went to multiple highlighted places in the transcript of the video of the audio and said, “Here’s where they said this. Here’s where they talked about this.” And so that was a lot a really awakening moment for me like there’s there’s a need here for like having catered, groomed, collected information in in your organization. And I I stood back and said, “Okay, this is interesting.” And you’re dealing with soft data there, too. It’s pretty soft. it’s just audio

20:11 Audio, right? That’s what I’m saying. A website URL, right? With tabular data, correct? It figured most of this stuff out, right? It’s this semistructured space. Mhm. And so I I thought to myself, okay, so I I like this experience. Well, how can I overlay that experience on top of what Microsoft is doing? Like what what does my fabric capacity need to do? How does that work? So I needed I stepped back and thought this is really interesting to me. but that that area that lens that information right that is what I wanted

20:45 To see as for like a center of excellence like a center of excellence felt like to me they should be able to build or Microsoft should provide one look here’s all the Microsoft documentation bundled into an agent that you can go deploy in your environment for all this stuff. I think companies would actually pay for that. Or if Microsoft provided the agent, here’s an agent trained on all the specific product pieces that we have. Oh, and by the way, Mr. company, you can supplement that with your, hey, Tommy’s done some instructional videos about

21:17 This for , for this particular company. Why not add those in? , here’s the reports that we’ve we’ve developed. Here’s let’s record a video on each one of those. Let’s throw those videos in the pile as well. And so that way we have more of this agent that can use and generate all this information. It then can source and provide answers back to users. Here’s the video. Here’s the section of the video where Tommy talks about this bar chart or this line chart or whatever. That is where I think you could really leverage a agents to really assist you with what you’re already building in your center of excellence. I really feel

21:49 This is a a very directly transferable skill into that other space or with notebooks LLM. if you’ve not taken a look at it. Mike, did you try the podcast feature with it? I I I saw it. I didn’t have a chance. I was on a podcast, [laughter] which would be great. Yeah, I built I was building slide decks and infographics based on the podcast and seeing like, okay, based on these episodes here, what are they talking about? Like, what’s an infographic about these episodes that that I should know and share with people? For those listening, if you’re not aware of Google’s Notebook LM, it’s it’s a really amazing tool that simply allows you to feed it PDFs, white

22:23 Papers, URLs, audio, , books, whatever you want it, whatever you have that’s usually long form, and to your Mike’s point, you can ask a questions, but also has like six distinct tools out of it where you say, “Hey, what? I have this PDF on, the maybe it’s a philosophical paper like the intricacies of technological design since 1903. Well, you don’t want to read all that. There’s a thing called podcast and it will actually generate a 5 to 10 minute podcast AI generated of that PDF or people talking back and forth real.

22:59 Yeah. Yeah. or there’s a video or there’s slide decks or there’s a mind map. But the podcast feature is great. I’d be interested what AI would say about us in a podcast like hey these guys don’t know what they’re talking about. We we are on episode 494 right now. So that’s so we’re we’re clipping along now here on the number of episodes. And when I looked at the note the notebook LM features you could have if you paid for the premium thing whatever their highest paid services which was only like 20 bucks a month something like

23:31 That. It wasn’t too crazy. But if you paid for the most expensive one you could up you could add up to 500 items or references. So, we’re fastly going to run out of items to throw at the the agent to go talk about things. But how cool would that be to be able to take again I’m just you that’s one example here, but I’m just thinking about what is your knowledge like what can you bundle up in a way that you can reference and create and bring more of that together such that when you need to ask questions about something, you can discuss it, right?

24:03 Should we pay for that for the five episode 500? I don’t I don’t think so, Tommy. I [laughter] think I don’t think that’s that’s the we’re probably not going to do that one. Maybe maybe like every year you you you build one for like one year at a time. I don’t know how it would work. But anyways, it it’s interesting to me and I feel like that would be something that would be interesting for people to interact with and would enhance whatever you’re doing like building your company, building documentation for your organization. Yeah, I saw a cool idea of this. Someone basically went to Perplexity and AI and asked to search all the information about themselves. They got their resume. So they fed all

24:36 This information, bios, everything that they’ve done into notebook LM. Wow. That was a resume. Nice. So then people like if any job person wanted to find out about a person, say, “Hey, does this person do this?” Yeah. Da da da does this. Here’s all the information [clears throat] about that. Here’s where they did this stuff. Here’s when they worked on this project. Right. Wow. So anyway, more and more though, I feel like to that point, Tommy, wanting to get hired, do things with fabric. , I every day, every day I get multiple people reaching out to me, hey, we we sell XYZ services or hey, I’m looking for a job in consulting or doing something with

25:08 PowerBI or fabric. Do you have any openings? And I’ve been giving a bit of recommendation back to them. And so when people hit you up out of the blue, just, hey, I’m looking for some work. I need to help. I want to help build something. Awesome. Thanks for reaching out. I, I definitely want to, , I appreciate you taking some effort to do that. But I look at it going I ask first I like how much what’s your pay rates like what I want to understand like what what’s your scale of like what how valuable do you think you really are it’s one of them but then right behind that I say what publicly

25:42 Available works do you have what can I look at that shows or highlights some of your skills because anymore I don’t go through interviews anymore like when companies hire me they it’s there’s no I’m not going through like the who are you what do you do what acumen do you bring to the table here. I point them to like, you want to know what I’m about? Go listen to my 490 hours of content that around this topic. Like that’s my resume, right? Go look at the website that I’ve developed. Go look at the training material that I’ve put out for

26:13 Free. Here’s my YouTube channel. Look at the number of subscribers I’ve got. Right? To me, when I look at So, if you need to go get work or job or efforts, if for those of you who are looking for opportunities, start participating in the community. Go to fabric.com. publish stuff. You don’t even need now with the community stuff that’s out there, you don’t even need to have a fabric tenant to deploy stuff. You can go build there’s there’s challenges, there’s or there’s items that are currently out on the Microsoft community. You can just download their sample PBX file, build a sample report and just upload the PBX

26:47 File back to the community. It automatically publishes it and serves it for you. It it it hosts your own report for you. So you could have a whole resume or series of resume of just different reports if you experimented with something that’s interesting to you. And I don’t care what it is. It doesn’t matter to me like what the topic is. Inform me, right? Give me some interesting teach me something. Right. Here’s a report that I built. One thing I’ve been looking at, Tommy, that I think would be a lot of fun. Mike will hire you if it’s on Legos, by the way. So that’s exactly what I was going to

27:18 [laughter] go to. That’s exactly You knew me so well, Tommy. I was going to say this before I’ve been looking at data sets and data sets are interesting but they’re always difficult to work with and so one of them would I thought would be really interesting would be is I really start I I really do need to start grooming or curating or collecting like a Lego data set like I think that would be really interesting. I know there’s a couple people on Microsoft I know Will Thompson is is big into Legos as well which is absolutely amazing additionally. So that being said, I really want to get my hands on some more

27:52 Lego set, Lego data set. So build build I’m interested in that. You may not be, but I could build a fun I find interesting. I could find a fun report or build some fun report around that topic or those things. You like watches, build a little dashboard around watches that you’re interested. You like supercars, great, do that. You like football or as the rest of the world calls it or US calls it soccer, the rest of the world calls football. like go build something about a sport that you’re really into. those are interesting dashboards and you can build there’s a lot of data at your disposal. Spend some

28:25 Time doing that. Make it interesting. Show me your skills. That will for me that will invoke a conversation immediately. I will at least engage with your content and go click on some of the stuff that you’ve got going on because I just want to see what you’re doing and never know that would I think that will push the needle further than people who don’t do that. 100%. Awesome. Well, that being said, I think we’ve talked about these episodes enough. One final point, Tommy, I just want to point out here, and this is just maybe for the Microsofties who are listening. When I am talking about all

29:00 These AI things, and I look at like notion, I look at notebook LM, I look at Perplexity and the stuff that you’re talking about, Tommy, I look at all these tools that are not Microsoft and then I say, “These are cool ideas. I could see how these things could be effective and useful. And then I immediately go back over to Microsoft products and say, where’s the equivalent of this? Where [snorts] does this live? And this is the one thing that I feel like I’m really struggling right now is I don’t know. I I don’t know how to I

29:35 Don’t know how to get that going. Like Microsoft, you’re you’re behind in this space. There’s a lot of good ideas coming out. There’s a lot of tools that are being developed. Microsoft, you’ve got my attention. You’ve got a lot of my text and great data already there. Why can’t I bolt on a notebook or notion like experience onto my loop documents? Yeah, I’ve got hundreds of rich documents. Tommy, being an open- source company, you do not allow a lot of other external tools out there or models. Well, I think I think I like things that

30:08 Are out there. I I’m I’m of the opinion of like I’m already paying for Loop. I’m already paying for these products. So, I’m already invested in that ecosystem. I’m not going to go into Loop and then suck out all the documents and try and put them somewhere else. I’m just not going to do it. I I’ve set my ways on those tools. So, two things either need to happen. You either need to let me continue doing what I’m doing in my existing tools and make it really easy to get me that data out into these other experiences. Mhm. Notion notebook LMS like make it easy or to the opposite effect which I think is

30:40 The right approach is hey Microsoft when I pay for co-pilot for M365 build this stuff build notion build notebooks LM like why can’t I select a number of episodes in my loop document and get the same experience chat with the information chat with the data I don’t want a co-pilot button to show up and rewrite my paragraph Like it’s that has been minimal to no use to me at all ever. Right. I don’t like frustrating than helpful. Absolutely. More frustrating than helpful. Even in

31:11 The emails, I don’t even use like a lot of time you have co-pilot in your Outlook and it says you should rewrite this. There’s all these little messages that keep popping up. Hey, I want to rewrite it. I like what I wrote. You’ve got six bullet points here. Like do you want me to help you rewrite it? I’m like no I don’t. This is what I want to say. Like shut up. Like instead like write me the email to begin with. like just have the like fully fully bake a response for me. Hey, email comes in. Hey, this is this is back to the article, Tommy. I’m going to get a little heated here. This is back to the article. Dude,

31:42 Emails are coming in the door. The AI should be smart enough to figure out what the context of that thing is. Are they asking me about a product? Are they looking for consulting help? are and the the AI should draft a full response based on the other conversations you’ve had from your email about getting clients in the door. Here’s what I think you should say and I would rather have a I approve that looks good. Go ahead and send it or change it by saying no your Outlook dashboard I like love the idea of that.

32:14 That that’s the stuff like so that’s where I think AI would be effective. So again it goes back to that the comment on this this thing here is AI should be used to identify broken things weaknesses and and solve it come with an answer or come with multiple answers back to the table here and give me what I want here. And so this is where I’m very like disappointed in what Microsoft is producing because a lot of the products they’re building right now don’t do what I want and I’m seeing other market view of the application you’re in. Yeah. I’m looking at other applications

32:46 And they feel like they’re doing the right thing. I’m looking at them going, “That’s what I want.” Every co-pilot feature feels like it’s siloed compared to the actual application you’re in. It’s not doesn’t feel integrated to Outlook or integrated to PowerPoint. It’s like a chatbot. But to your point, I should already see when I open Outlook, hey, here’s the four emails that are most important. Here’s the response that we have. Do you want to edit it? Rather than I have to go through each individual one, there’s a button. It has no context of the previous conversations. I completely agree with you. Yeah, I I have other things to complain about, but dude, I think we got

33:19 To get to the main topic. All right, let’s let’s keep going. I love it. I love it. But I think Microsoft has enough notes to take away from this. Yeah, we’ve been we’ve been jamming on this a lot, but this is just general AI stuff. So, let’s talk about MCP things. And Mike in the chat here, you are 100% correct. I think Microsoft is going to push 100% a lot more AI. That’s going to be especially at FabCon in Atlanta this year. That’s going to be a lot of conversation. And then the MCP side of things, MCC servers, they’re also going to push that heavily. And I know because I’m talking on it. So if you’re coming

33:51 To Microsoft Fabric Conference and you want one of the early sessions, you can actually pay for precon sessions. Nice. I’m doing a full day deep dive on MCPs. Nice. Getting started hands-on keyboard doing a workshop. It’s a full eight hours prior to the event. We’re on Monday. So on Monday, Matias Turbach, , who’s amazing, by the way. He’s he’s way better at AI than I ever will be. , he’s such a genius when it comes to that. So I’m I’m very privileged to be able to even just speak with him on a stage. Super happy about it. But we’re going to do a full day around MCP, MCP

34:24 Servers, how to get them started, how to use them, and how to use them effectively in your workflows. So that being said, let’s get into our main topic today. Tommy, kick us off here. What’s going on with these MCPs? So, first off, if you have not heard of the other episodes, what is an MCP? I think is a good place to start. Model context protocol. Context protocol. Let’s start with the definition of what it really is. And I think this is really important for us to understand what do they do? What is a model context protocol? And from the from here on out, we’ll say MCP.

34:58 It’s an integration into an AI tool, IDE, chatbot, whatever it may be that acts as if it’s a skill or gives additional commands to your AI tool that you’re in. So, let’s take a I’ll take a brief example. There’s an MCP out there that’s called it’s called Task Builder. And what you would do if you integrate this MCP into your let’s say claude or VS code the task builder would have multiple commands like build a build out my project plan and it would

35:32 Say like project plan your chatbot IDE understands as part of the model context protocol the MCP and begins to basically do what that command does. So there’s a ton of MCPS out there. I would say there’s definitely over 400 out there that do general tasks. There are ones that connect to other tools like notion obviously outlook say check email or whatever it may be. Yep. Now Microsoft has finally developed and these are built by Microsoft a few MCPs

36:07 For us in the fabric world. There is the remote PowerBI MCP server. There is the local one that runs on your computer. There’s a fabric MCP and all these again have commands that help it understand context around timle around building semantic models promatically. enables AI agents to update and manage table columns measures and relationships. It can execute DAX queries for validation. So they have all these additional commands rather than me having to explain to the tool that I’m

36:41 In basically hey what Tim is and then you have to execute a DAX query DAX query is that the MCP provides all of those commands context and basically understanding into that MCP where does an MCP work an MCP works most ideally in VS code cursor any IDE tool but they also work in cloud cloud is actually Anthropic is the company who developed MCPs. They are on a roll by the way because they also develop skills. So you can integrate this not just in an IDE

37:15 Like VS code that you have to have GitHub but you can also integrate this into your chatbot if you’re just chatting with Claude and say hey Claude let’s take a look at my semantic model over here and just have a conversation without any code actually up. It can still execute and provide that those commands. So, right off the bat, Mike, MCP, I think just an explanation. It seems it’s very powerful and it is very powerful. Obviously, you’re doing a training, something that I’ve tested out a ton and using now. Yep.

37:47 They’re incredibly powerful. They have amazing use cases here, but let’s start here with why do you think Microsoft even built them in the first place? Was it just to get on the on the trend or or get on the train? or was it more because the the need for it? Yeah, let’s I think I’m going to maybe take this question as more of like a general this is a a footnote. This is a a step a stop on the journey of things with large language models, right? I think there’s

38:20 A lot of research coming out about different spaces. I think a you could build these MCP like experiences and talk directly with the agent but I think a lot of times what we do is these agents arrived you could ask interesting questions it would give you very good results the evolution is now let’s start moving these agents and this is day one when when chat GPT came out I said look yeah it’s giving you useful information you can ask it questions I do like that that’s fun I enjoy that but for [snorts] me the power of an

38:52 Agent really or any large language model was I wanted to start doing things initially very day one I was like until they start doing things I’m not able like to really stand behind them and so what I see happening now is we’re moving away from this world of like agents sourcing information and doing lookups and searching to let’s start talking to a computer and actually having executing things and doing actions and creating things and so that’s the step where we’re at right Now I think which is the first pass of this is well let’s build

39:25 These MCP like things that can communicate and provide skills and knowledge and tools directly to the agent. So giving the agent more capability. the other thing too Tommy is what a progression that we I I feel like I’m seeing here as well. You see a lot of agents in a browser right? So you go to the browser you talk to the agent in the web browser and it it does certain things. I can make images. I can do but it it has this context window. It has like it can only do or action on things that are inside the browser window. Right. I couldn’t have the agent come down to my computer and organize my

39:59 Files. I couldn’t have the agent come down and create stuff on my computer. Yeah. Now you’re installing these agents directly like servers, MCP things. Now you’re giving through VS Code and now through the command line interface, the CLI, you can actually run agents there as well. And again, we’re seeing I feel like there’s another wave of MCP servers, command line interfaces. That’s where agents are getting more control directly on top of your computer. And so this is this is a wave of now they’re becoming directly more actionable that

40:32 Is allowing me to build things, create stuff, and and I can speak to it, and it’s still doing what I want. It’s doing what I need it to execute. Build this model, create these folders, add these measures. add this description to stuff, right? I can talk in my normal language to it and it does all the DAXs and all the all the writing for me behind the scenes. And that’s the powerful part, Mike, rather. First, I think that first iteration was the glorified Google. It was a glorified search engine, probably. Right. Yeah. Yeah. I figured they were figuring out the technology, but it was also it would wow you like there’s things that it

41:04 Could do. It was doing incredible things. Yeah. Even in that knowledge generation and pulling stuff together, it was impressive and that’s why you wanted to use it. Yeah. Yeah, but it was just more like to your point retrieval rather than actionable. Correct. And then I think the second iteration was still more developer based where it had those specific commands. And now we got to the point, Mike, where with the MCP server with PowerBI, I can be in cloud and go, hey, let’s take a look at something. I’m not sure. And I can ask it like I’m talking to an assistant or a colleague like I’m not sure if semantic

41:37 Model A is if all the DAX measures are organized. can we take a look at that rather than execute DAX query organization tool and then like we’re talking to a robot it’s it still has that LLM conversation it’s still not just natural language like we heard in PowerBI that’s now getting deprecated but you can have that conversation piece go hey are all my DAX measures organized ask it a question it’s not and it’ll basically provide the tool on the back end but provide you with the natural language just like you’re having the

42:11 Conversation. This is an interesting point you bring up, Tommy, and I like what you’re what you’re saying here. And as you were saying this, another thought popped into my head. People, so let’s let’s talk about the interaction between like computers and and humans and like what’s going on here. Like a lot of a lot of what we’ve done again even for modeling. Let’s just talk specifically about modeling. I’m writing in a user interface. I’m writing M code in Power Query. We love it. So good. Love it. And we’re trying to build tables and data transformations on top of stuff. And as we do that, we’re

42:44 Clicking buttons. And the reason we like Power Query so much is because I’m not writing code. It’s it’s it’s doing a very precise thing with things that are easy to interact with. And the same analogy, I think, can be brought forward with these these agents and AI based things. It’s the same concept. I’m able to interact with these agents and what they’re doing, but it doesn’t have to be so precise. Like if I go back to again thinking through like PowerShell or when you’re writing code, you have to be extremely

43:16 Precise when you write the code because if you don’t write it the exactly right way, it will not work. You can make a function or define a function. If you don’t call it off with the right capitalization, the right everything, it doesn’t work. It won’t run. It’ll say I can’t find the function. It doesn’t it doesn’t do what I want. So I feel like when people humans work on stuff there’s this little bit of like friction that is removed when you work with other people like oh I can’t remember the name exactly the function like I know what I want to do I know conceptually I want to make more measures put them in a folders and organize my model

43:48 Right I that’s what I want but and unless I’m using very precise like okay going in click these exact things to make everything work you’re able to pull back a little it and make all that stuff just happen naturally. And so my feeling here is the AIs and the agents is becoming how more like how people work with other people. Yes. Yes. It’s it’s a little less precise. It’s a little bit more unstructured. It’s a little bit more free form. I guess maybe I’ll call it I’ll call it a

44:20 Free form, right? It’s a little bit more a free form. I don’t need to be as precise with what I’m writing and how I’m doing things. Yeah. So in the same way Power Query generalized the code that I needed to write to build the data transformations, I get the same thing for agents and now MCP servers do the same thing, right? I don’t have to be as precise, but I get all the full capability of doing, editing, changing, manipulating the particular tool that the NPC talks to. I hope that never becomes just matterof fact what you just said. And I want to

44:52 Emphasize how how important that is to I think the world that we live in now where to your point like we’re no longer function based and for the first time in the world when it comes to computers. and I know at some point we’re going to go yeah that’s just the how the computers are but this still incredible to me like I I still hope I have the wonder of this. For example, Mike, one of the things that I’ve been asking my MCP is rather than having to go through the model like, “Hey, can you make sure that all all my dimensions are distinct here? Is this distinct here? Can you just take a look at this and

45:25 What am I missing and take a look at the model? Can you give me some suggestions?” Yeah. And so rather than going look at distinct, evaluate query DAX query. Okay, it’s not distinct. Okay, now I have to take a look and do that, , , that whole engineering myself. I’ve just asked it a question and provide suggestions because you’re still dealing with the an LLM. it’s not just the a tool. So, Mike, we are at I think a really cool point here, but I think it begs the question and unless you had anything else to say there, I was going to ask you about who

45:58 Are we dealing with for because I have two fundamental things to talk about with the MCP. I think it’s who’s this for and is it ready? Is it ready for production and organization? So I don’t know where do you want to start with that? I guess we should just maybe talk about where we see it evolving like so let’s give a little bit of background on both these two topics. Right? So Tommy you’ve talked about what what is the MTP? We’ve defined it’s it’s this layer. It’s information that the agent whatever the agent you pick is right you can use

46:31 GitHub copilot. You can use claude you can use all these different agents right the large language models. Sonnet chatgpt it gives them information to let them be effective and make changes and actually do things make API calls on behalf of the user without actually having to write the individual API call just figures it out to some degree. It feels a little bit magical in that way. So that’s the concept there. How are you using it? Like right now I’ve been playing with the MCP server for PowerBI and I’ve been working a little bit more with like having it look at a model and

47:04 Understand what’s going on in the model and like creating measures or organizing measures. Just some of those standard things that I would normally have spent a lot of time doing, right? I don’t remember the property that I need to do to turn off all the implicit measures. Hey agent, what property do I set the property on this model to turn off all implicit measures and only use explicit measures? Oh yeah, I know what you’re talking about. Here, let me add the property here. I’ll just set it for you. Right. It’s it’s that stuff, right? I know what I want to accomplish. I’m using the MCP server there. I have not played extensively. I’ve just barely touched the one around

47:36 The fabric MPC server. And I got to be honest, I have not touched the real-time intelligence MCP server at this point, but that’s another one that’s out there as well. So, that’s what I’ve experienced. And again, I I would consider both Tommy and myself like the the pro developer, right? We’re in that we like all the details and all the prof all the professional stuff over we overthink everything right so we’re we’re into the positive we like GitHub we like all the code like that’s that’s what I’m doing and I’m finding these tools effective

48:10 Yeah no to your point Mike I’m asking you questions that I wouldn’t be able to do before example even one like hey look at my descriptions which one of them don’t seem to fit with the rest of the model because a lot of times I’ll have Yeah. So stuff like that where it’s not just execute something that I I could have done myself. It’s things that I’m also utilizing that again remember you still have a LLM underneath this. It’s not just an automation tool of executing something in PowerBI for you.

48:43 It is still the same LLM, the same amazing agent that you have but with the tools that you have. So okay my other model created all the descriptions what doesn’t fit and then it will find like oh that one doesn’t make sense what else can we do here hey actually these measures actually have to do with this background so can we update all the measures that way the organization is a big point of that but then again it’s things that you would almost assign a human to do rather than just an automation agent where it’s like are we missing any measures here what

49:17 Measures am I not using let’s let’s clear those out or organize them a little better. Can you give me suggestions? And I always ask it rather than just executing for me or doing something. Yeah. Give me some suggestions about a better way to organize my measures. What is a better way based on how I’m using them in the report or based how the models set up? Hey, what else can we do here? And honestly using the reasoning feature that’s already available in all these tools which I use extensively reason and help me suggest other ideas for other data a better way to

49:52 Optimize this and and again like what columns do I not need? I think that’s a great example there. Tommy what I haven’t done on the large language model and I’m not sure if it it works really well yet on it. I know it has context to the model pieces. It also has context of the report pieces as well I believe. , so this would be one thing that would be a a hard question that people have built deliberate tools around which is Yeah. Hey agent, of all the visuals on this page, which

50:25 Measures are used or not used? Right. Right. Oh, and where is across this whole report which measures are used or not used? Right. So that’s the stuff like people have built full-on tools to go scrape and parse and all these things. Th this we’re at a place right now where the tools or the MCP servers could be adapted to build something like that, right? And now you don’t need to go buy a product or have other things. You can just use these things off the box. Now I’ll also say this too. These MCP server tools that we’re talking about here, these are free tools.

50:58 You can go download them today. Microsoft is just providing them as part of the product. You can go use them. They are all extensions though. So Tommy, you asked a question, who should be using the MCP servers, right? I don’t think this is new users. I don’t think this is a user who’s brand new. I’m just starting with PowerBI. I’m going to jump right in and immediately start using MCP servers. I think it’s I think it’s I think it takes people who already understand what they’re trying to build, some seasoned, medium to senior

51:32 Developers. it’s going to make them more effective because what you need to do but it’s just all the tedious work of these little steps over and over like I want to go through every column that’s a number and I want to make a measure that’s called sum of the column name and I want to have the original column hidden I want to have the measure presented in the model blah blah blah done build a calculation group that does this right here’s the base measure and here’s the time patterns I want to implement on this calculation group Those are the things that

52:04 I know as a developer I have the right words to say to get it to do and build the right things. I might play a devil’s advocate here because to me, if I’m using Claude to look at a model, this almost screams what a great introduction for a beginner to get used to a semantic model for my organization to get the idea of what the semantic model’s about, but also provide and do these actions that I may not know how to do. To your point, I agree there is that point of you have to know what to ask or what to talk what to have

52:37 The conversation about. But at the same time, you are literally providing a chat interface to execute complex functions, actions, and automation around a semantic model that I don’t necessarily need how to do. And it’s in a nice friendly user interface. And that almost on one side of the the other side of the coin, Mike, that almost screams beginner. That almost screams introduction, right? Because I could say, “Hey, I know you’re new. Here’s an MCP set up on our development workspace. I want you to take a look at this, you

53:11 Know, see what we have and just help organize everything.” A be an intern could do that. A data steward could do that. Why couldn’t a data steward say, “Hey, organize all my models. Help me with the descriptions and the measures.” Okay. So, I I have a I have a point of friction about your comment. While in concept, I understand what you’re trying to produce and the MCP server does exactly what you describe, right? It it takes chatting experiences and makes it simple and and lets you do complex

53:44 Things with it. What I think is the missing part here is in order for me to turn on the MCP thing, this goes Yep. Okay. So, so there’s there’s a there’s a cliff. There’s a learning or educational cliff that I think is causing some mental pause for me. That stuff needs to be in VS Code. So, you got to have VS Code installing your machine and you got to be running it. Then you got to install the MCP server and then you’ve got to spend a little bit of time figuring out how to connect

54:16 To the model or get it up and running on on desktop. So, it’s not built into the product. Now Tommy, had you told me had you told me PowerBI desktop is now shipping directly with MCP server and now even though it’s still the MPC server now if you go into PowerBI desktop open up a new window pane and you can attach window like so there’s another pane on the so call it MCP pane more pain from the MCP right from desktop you just let’s we already have a thousand stupid window panes anyways let’s add more pain to it

54:49 So Drop in a new one. Make the MCP server run inside desktop. How is that not a thing yet? Right. And so enter enter here. Right. What large language model do you want to use? Right here. By default, we’re going to use the co-pilot from fabric. You can just use that one. That’s fine. And then I again me I’m in desktop. I don’t want to only use co-pilot from fabric. Give me AI foundry agents like so turn on that thing. Let me have let me pick from any agent from AI foundry. Let me pick hey,

55:22 I pay for GitHub copilot. Let me pick from any co-pilot there. here’s all the the the GitHub co-pilot experiences and I can pick whatever model I want. So I or even better yet, give me an input field. Here’s all the different MCP servers that all the different large language models are out there. Enter your solution. Enter the URL of the API you’re going to hit and then just put in the key that you need, the token, right? let me enter that directly into desktop and then I can then to your point Tommy. So that’s the idea. The idea is make that part of the desktop product so that

55:57 Just works with it and I can add my own agents directly into desktop. Then I don’t need VS Code, then I need all these other things. And now now we’ve leveled up. Like to me that’s that’s the that’s when I would want to give this to basic users, right? It’s part of the product. It’s in the desktop application and there’s minimal to no setup. I click on something, sign into something, and it starts using agents. That’s Yeah, that’s the level we need to get to to really push MPC stuff to the next level. And who cares? Like at that point, you don’t even need to call it an MCP

56:28 Server. It’s just chat with my model. Yeah. Or chat with desktop, right? And that going back to the article earlier, right? Agents aren’t going to do a good job of building the insights on top of your data and pulling those out. Even now with co-pilot adding stuff, it just feels like a novelty at most. It it doesn’t really feel I don’t hear organizations clamoring for leveraging chat with your data like experiences everywhere. However, yeah, I am hearing a lot of organizations clamoring for I want to create with

57:01 Agents like that feels like where people need help. Well, let’s lean into this because you the novelty thing and I think this goes to work, dude. We’re already in your time. , is this a novelty in itself? Are you using this with client work right now or in in in terms of I always go back to the workflow? It’s one thing to test something out to trigger it. It’s another thing where I know when I’m going to build modify something, I go through the same usual patterns in terms of the tools that I use, the Q&A that I do. And there are some AI tools that

57:36 Are now part of my workflow. They’re integral to what I do to the point where if something were to be removed from the internet or removed from the world, sure, I would have a moment of, loss. And right now for an organization, if I’m an organization looking at this or I’m the, lead BI, , director of CTO or a direct director of BI or CTO. Yep. Am I looking at MCPs for PowerBI and saying 2026 is the year that we’re going to implement MCPs for the BI team and it needs to be part of their

58:10 Workflow or are we still at the novelty space? Is this still more individual to individual? I may use it but my colleague may not. That’s a great question, Tommy. I I think right I think we’re in the early stages. I think it’s starting to gain more momentum. I also think here too, Tommy, I think the MPC server experience just in general, like when you look at everything that’s MPC server, I think this is like the easy win that organizations could be like, look, let’s

58:42 Take a developer, let’s have them spend two months building out whatever the MCP server thing is. And it’s easy to build the skills, it’s easy to build the integration. It’s not super difficult. And you can pretty much pop out MPC server on anything. Anything you look at now, you literally Google what product you’re using and say MPC server and almost all of them have it. Oh yeah. Which it’s nice that this thing evolved, was built, and now every product supports it. Right. So that’s close down your machine. Correct. That’s for sure. But I’m just saying like does the fact that you can just turn them on and they

59:15 Just work and there there’s lots of them makes them effective. Is this the final form of what this will look like? I don’t think so. I think this is one step in a series of evolutions building tooling that’s going to help support agents to be more effective inside whatever you’re doing. So I don’t think MCPS will be the final form. again back to like I really just want if I’m if I’m doing all this stuff why can’t why do I need to put an MCP server anywhere on my machine right? Why doesn’t why isn’t there an easy

59:51 Interaction directly with Azure and the MCP or Fabric and the MCP? Why can’t I deploy the MCP thing inside Fabric today currently? You right? What what does that do? Right? I could imagine there’s a whole bunch of MCP admin stuff that I want to do as an MTPC admin on top of my fabric tenant. Why can’t why doesn’t that just exist? Why can’t I make a workspace and deploy an MPC server directly into the workspace? That that feels a bit more natural to me. So, I don’t have to run it on my machine. I can just talk to it directly. I can just

1:00:23 Wire some some things in, sign in, and just use it. So, I don’t know if I really answered your question there, Tommy, about what you’re looking at there. I think there’s just it’s still to me, it just feels like it’s not stable yet. And I think that’s where I’m at, too. If I’m a CTO or director of BI, I’m looking at this with one eye. I’m setting up an agent to or a Google alert to take a look at the updates here because it’s still very novel. It’s very cutting edge, but as with any new technology, Microsoft both is said this is in public preview. It’s not in GA. Correct.

1:00:55 A big part. Microsoft also recommends each organization do their own security review before any integration. I think this is point. So, this is a great point, Tommy. Good point. Yeah. Yeah. So, I want to definitely double down on this one. I wanted to bring I actually wrote down security considerations for this one because dude, we are on the same page today. We are, man. This rarely happens that we’re so in align [laughter] in alignment here on stuff. So, I guess it’s the end of the podcast. I I guess we’re done. Like, there’s no more conflict happening here. I will argue with this one too, Tommy. So, when you pay for So, again, we talk about

1:01:27 These large language models. They’re given information about your semantic model or details there. I don’t use any open like there’s two ways you can use a large language model. Yeah, you can pay for it directly and you get the API keys and you get the tokens or you basically send data directly to open AI. So where the AI runs on matters I think. Mhm. And so if you buy claude you’re sending data to the cloud servers wherever they are AWS as your it’s their stuff. When you sign off on those agreements in

1:02:02 Those fine prints and what you see it’s basically all their data like anything you send them could be used may show up in a model and it’s it’s I wouldn’t Yeah, I’ll say it. It’s not secure. like you’re you’re g you’re sending prompts to them. No. Yeah. You don’t know what the a agent is sending or not and how much it’s getting back and forth and you don’t know what they’re going to do with it. So, I’m of the opinion that I want to be able to trust where the models are running on. And I feel like the messaging that I’ve gotten from Microsoft and particularly GitHub agents is yes, we’re going to use these various

1:02:37 Large language models, but we’re going to host them for you and the prompting you’re using to the agent will not be used to go back and train the next version of the agent. Right? So if I’m using Claude inside VS Code, I know I’m not going back to the Claude servers. I know I’m going to something that Microsoft has stood up where I’m sending my prompts to Claude. It’s doing the work and then it stays it stays there. It ends at the end of that conversation. So let’s go forward here a little bit. If you’re talking models MCP stuff, I only feel comfortable using Foundry my

1:03:12 Azure Foundry as agents and I only feel comfortable using GitHub co-pilot agents to do any of this stuff. So I’m doing exclusive development of my MCP stuff directly inside VS Code. That’s all that I’m doing. I don’t feel comfortable sending it out to other APIs, ChatgPT, Claude, , Gemini. I don’t pay for them directly and I don’t send stuff to them that I know is any client work at all. So I think inam companies and organizations are starting to become aware of this. , but I think really

1:03:45 There needs to be a stronger stance around this and like at the end of the day, how do you control this? What can you really do? This is the big part. This is the big part. Yeah. Right. We have it’s been it’s too MCPs have not been around nearly long enough for organizations or even the people creating them to do the security review. Mike, I can’t wait for the day that Open AI or one tool is going to get hacked one day. it’s just a matter of time. If Target can get hacked, I’m sure Open AI can. But I think that’s this is a big point. If you’re an organization, it’s going to be a non-starter to tell your IT department to install this on

1:04:20 Your BI departments or the BI computers, right? Like for me, yeah, like I have my own security and my own sandboxes that I play in when it comes to client data or my own, but that’s fine. It’s my computer. But this is not something I don’t think an organization and if you do, you better tell your boss if you’re going to install this, I think, because there’s a lot of ND or what is it? NDA and all the IP information that I don’t think has been fully vetted out yet. That being said, at least from the local the local MCP server.

1:04:52 Yep. If you are someone who is more developer base and I think already pretty pro, have the I would still have the conversation and maybe use as an individual to pilot to talk about it. , it is a cool feature, but no, I I think where we’re at is I have my eye on it. I’m but I’m not there installing it yet on on my enterprise or my production side computers. No, I don’t think aren there yet. Again, it’s in preview. I don’t think I would go to production with it. I definitely think it’s a very interesting novelty. I definitely want to try it out. I would highly recommend if you’re on that

1:05:25 Medium to senior level developer, building semantic models, building reports, your your team, you’re constantly given models from the business and you’re asked to evaluate and figure out why it’s slow and what it’s not slow about. That’s the stuff that I think you would really want to use this for. Like it’s it helps you get through the discovery phase a lot faster. Seth on the podcast a long time ago would always say he he would be given a SQL statement and he was a SQL developer and his first thing he told me he said every time I got a SQL statement or I was looking at someone else’s code the first

1:05:59 Thing I would do is I’d go through and organize it so I understood what it was doing. And so if I step back and look at someone else’s model, to me, I could make the analogy that what Seth was doing with the SQL code, which is I you’re giving me a bunch of SQL. I don’t want to just straight up use it. I want to just at least reformat it in a way that makes sense to me. And so I think someone in the business coming to me, here’s my model, here’s what I’ve built. I would want to go through and say, okay, agent, organize this. go through here and figure out which tables should

1:06:32 Be dims, which ones which ones should be facts. Let’s just figure out how it works. Put all the measures that are calculating things from fact tables. Put the measures in the tables are calculated in like that would that would make a lot of sense. So, I really do think like that would be useful and but again, I think it’s still a tool for that senior developer. Can can we have that chatbot the way you’re asking it also with Seth’s tone? So whenever you ask it to do something, it’ll go, I guess I can. The snarkiness like we need to start the sassiness. Bring back the sass the sass level.

1:07:06 [laughter] Yeah, exactly. Oh, man. All right. Well, dude, I think we are way past time. , but Mike, I love what you said today and I and I I’m glad we’re on the same page because we both, if I’m hearing you correctly, we both have the excitement of what the MCP potential is can do. And it’s something that we’re not just saying kill it. We’re saying no, keep building it. Keep in maturing this product, this feature because this is absolutely something I want a part of my workflow. That being said, we’re still in the early stages. We’re

1:07:38 Still in incredibly early stages here. So, but no, my vote, let’s keep moving forward with it. I I agree. I think this is a we talked about in the past, Tommy, we’ve done a lot of episodes around kill it or keep it, right? This is and this is a good opportunity to say, would we kill this feature or would we keep it? I would say keep it. This is definitely a keep it feature. Keep evolving it. Keep building it. I would say though as we’re ingesting this information, right? Make it easy for me to use it. Honestly, people like it’s too difficult right now to use anything AI related anywhere in

1:08:12 Fabric. And that’s why you got eight hours for your precon. Dude, I’m telling you, like there’s just too much friction. So, I’m I’m going to be willing to bet you’re going to get to a point where adoption starts stagnating because you’ve made it too hard to get going with it. It doesn’t it doesn’t give enough wow factor. It’s not easy enough to get started. That learning curve, you’ve got to make it super flat. It’s got to be really easy to get started or else it doesn’t adopt. So, that’s where we’re at. Anyways, very fun topic. Tommy, thank you for bringing this great topic to us. This is great.

1:08:44 , thank you chat for participating and giving great comments here as We really appreciate you. You’re you guys make this fun and enjoyable. So, we really appreciate you being live here on the episode. , we’ve had a good number of people show up today. So, thank you very much for chatting to us and giving us your information about what you feel about this topic. That being said, if you want to help support this, if you want more content like this, if you want more exclusive stuff, we’re looking at trying to invest more, but we need help from you. So, if you wouldn’t mind u consider becoming a member of our channel, you can subscribe. That’s

1:09:15 Always for free. We’d love you for to do that as well. But consider becoming a member, support us. , see if we can get more stuff out the door here and continue to ramp up educational parts around all this fabric and PowerBI stuff. Tommy, with that said, where else can you find the podcast? You can find us on Apple, Spotify, wherever you get your podcast. Make sure to subscribe and leave a rating. It helps us out a ton. If you have a question, idea, or topic that you want us to talk about on a future episode, head over to powerbi.tv/mpodcast. Leave your name and a great question. And finally, join us live every Tuesday

1:09:48 And Thursday, 7:30 a.m. Central, and join the conversation on all PowerB tips social media channels. Thank you very much. We’ll see you next time. [music] [music] out.

Thank You

Thanks for tuning in! If you’re experimenting with MCP servers, we’d love to hear what you’re building with them.

Want to catch us live? Join every Tuesday and Thursday at 7:30 AM Central on YouTube and LinkedIn.

Got a question? Head to powerbi.tips/empodcast and submit your topic ideas.

Listen on Spotify, Apple Podcasts, or wherever you get your podcasts.