Clean The Data Estate First, Or The Agent Guesses

Transcript

Transcript of the published episode, lightly edited for readability. Timestamps match the video.

00:00 Sebastian Selman Imagine the following. A senior executive walks into her office, opens her, let's say, enterprise chatbot, asks a question, "What was our gross margin last quarter?" And the answer comes back in seconds with confidence. 46.1%. And now that same executive opens the governed BI report to double check for the same period, same company, same question. 40.5%. Same question, same company.

00:33 Estera Kot Two different answers.

00:36 Sebastian Selman The difference between them decides how she or he is going to talk to the board, let's say, on Thursday, and which one of those answers is the correct one, and which one actually matters for the discussion on Thursday is not decided by the chatbot. It's not decided by the Power BI dashboard either. It's decided probably primarily whether you have a semantic layer underneath that is unified and trustworthy.

01:04 Estera Kot So, semantic layer, something beneath, underneath the dashboard.

01:11 Sebastian Selman Exactly. Because I think the dashboard was never the point. The layer underneath it was.

01:17 Estera Kot Yeah. Hello, everyone. My name is Estera Kot. I'm the CTO at Clouds on Mars.

01:23 Sebastian Selman And I'm Sebastian, consulting lead for Switzerland.

01:27 Estera Kot And this is business mission control. So, today's episode is the continuation of the previous episode when we just highlighted that dashboards are gone. And we talk about what replaces them, who wins, who is structurally exposed, and what for Switzerland-based executive it means.

01:53 Sebastian Selman So, here's the, I think, general word out on the street that I want to discuss with you today, Estera. A BI tool exists to help people make decisions. I think everyone would agree on that. But I think that's only half the story because a BI tool does, in my opinion, two different jobs. And the second one is to actually help people in an emotional and political environment. So, it gives someone who wants to convince others a narrative and a picture to project. The BI tool or the BI dashboard might be the artifact that you take into the boardroom to convince the CFO to fund your program. Or you take it into the executive committee to defend last quarter. So, here's my hypothesis. A BI tool is sold for two jobs. One is decision support. The second is narrative generation. And agents are eating, I would argue, both. And nobody puts the second job on the pricing page.

03:06 Estera Kot Like narrative generation, it's storytelling. Yeah. But before, let's slow down, and I want all of you to slow down. So, dashboards still exist. Almost every enterprise I walk into has either Power BI estate or Tableau estate and wants to migrate to Power BI. So, we need to just be specific what is actually going to be replaced or is actually being replaced. So, not data, right? Not governance, not a model. So, what?

03:45 Sebastian Selman Exactly. So, when I say agents are eating BI, I mean they are eating the interface. The dashboard rectangle in your browser, the interface layer, that might become conversational. Practically, many move to a point where they ask a question and the answer comes back to them in chat, maybe as a chart or paragraph or a number. Sometimes rendering, sometimes just words. The underlying layer, as you said, the data, the governance, the model probably will stay, but the interface is changing.

04:20 Estera Kot And the interface is a tile in the browser. That is gone, or just that element. So, the measure, the numbers, the logic behind those KPIs measures, calculations, they are going to still be there.

04:38 Sebastian Selman Yes, I think that's correct. That is kind of the reframe I want to carry through this episode. The dashboard only was the interface and the interface probably is going to leave. But what survives is the substrate and that is where, to me, the moat, the defensibility of your business, lives now.

05:01 Estera Kot Yeah, so now jumping in to be concrete specific. So, what substrate actually look like when you are building that. So, here I'm going to reference one principal architect at Microsoft, Meinrad Weiss. He published a demo in November last year and it was a two post, one long form write up. So, the demo was about using VS Code, Visual Studio Code, using GitHub Copilot, using Fabric MCP Server. By the way, one of my friends was authoring Santosh and greetings to Santosh from Fabric Product Group. But back to the demo, so VS Code, GitHub Copilot, Fabric MCP Server and then Language Prompt in whatever the language model is supporting. So, English, just those four elements, ingredients of our soup. And what was the outcome? The outcome was a governed semantic model. So, the layer that agents grounded the answers in hours, not weeks. So, before, to have visualizations, we spent majority of the time on building semantic model. And what is semantic model here? Basically, you put the data, you structure them and on top of that, you want that data to have some relationships, some measures, some more human understanding definition. What is that? And when we, at the beginning, started to just talk with the data, raw data, how to expect AI to know what is going on. If there is a column named ABCD and it's not an abstract. I've seen that, those column names across all of the enterprises. Again, people are doing shortcuts, meaning that many times we just forget to keep the quality and then later on we expect AI to fix that quality. So, with that soup, VS Code, Fabric MCP, GitHub Copilot or Claude, coming from AI Foundry, we are getting the answers in hours, I would say minutes, sometimes even seconds. So, we have like nine KPIs, 63 time intelligence measures, the year to date, quarter to quarter, month to month, previous period comparison that every dashboard secretly runs on and do all of those either computations behind before or when we are just using that dashboard. So, all of the generated queries are based on our natural language prompts. So, we do not author DAX queries, SQL, T-SQL queries, Spark queries. It's just behind the engine and our job is just to ask a precise question. And then that soup is handling what we are looking for. And once that layer, that semantic layer is in place, the beauty is that the same model fits Power BI, lovely Excel, M365 Copilot, Fabric Data Agents, any custom applications including Fabric applications, Fabric apps. So, in November last year, Meinrad's tagline nails it. So, build once, this semantic model, semantic layer, and you can trust it from every connector, every tool that you are using. So, build once, trust it everywhere. And here is Meinrad's second tagline coming from this article from November last year that every CDO is looking for and that's fundamental for every CDO, especially in Switzerland. We care for debating and quality and we admire that. So, you want to use AI to build a solid foundation for AI, meaning that many times, yes, we are using more and more AI to build that semantic model. It must be robust. Why? Because based on your semantic model, you will get specific quality of those prompts in human language that you will get without that. And again, sounds simple, but I see that here that every single week that yes, you have a data, database, PostgreSQL, whatever, OneLake, and then you connect some agent, some language model, and you expect it will work out of the box. So, missing the solid foundation for AI, which is semantic model. Now, let's talk about specific demo, which is showing the top of funnel story. It shows what a new semantic layer looks like when you are building it from scratch. We thought that one day you will use this with AI and language models. And it doesn't tell you what you need to do about the estate you already own, because for that you need the systemic version. And here at Clouds on Mars, we published a framework for it in a deck we are calling AI Readiness. And you will get access to those resources. Just take a look. There are two tiers. Native tools are coming first. So, AI schema, verified answers, AI instructions. So, the things that Fabric already is giving you. Advanced Next is a scoring model, 0 to 100, that tells you which part of your semantic model, semantic layer, is ready for your AI grounding and which not, which will or may hallucinate from the KPI's perspective, meaning that once you will ask about the specific KPI. So, one structure methodology to express you in numbers, are you ready or not? And the workflow is scan, analyze, fix, and monitor. Again, not rocket science, nothing glamorous. It's just what should happen before you enable compile it across your data estate.

11:50 Sebastian Selman And I think the sequence matters here a lot. It's not agent first and estate second. It is estate first, agent second. What you were just describing, right? Build it once, trust it everywhere, use AI to build a solid foundation for AI. Actually, it's a good example of not simply automating the process that already is in place, because then we would still use DAX and basically use Codex or similar, more advanced maybe models these days to write the DAX queries. But we're actually shifting left the entire estate one layer down into fabric and then allowing more different engines to play on that unified data estate in OneLake, right? So, in that case, if you do estate first and then agent second, you can win. Copilot will reward the cleanup. And importantly, if you don't do the cleanup, copilot will not compensate for the absence of a cleanup having happened.

13:01 Estera Kot Yeah. So, an example is that if your model has three measures, each of them is called revenue, two of them are wrong in some different places or different ways. What will Copilot do? Copilot will happily pick one and answer with confidence. This is not a copilot problem because how can you expect copilot to check verify your CFO definition of the revenue, your as a business unit definition of the revenue and so on. It's not a model problem. It's not a copilot problem. Again, the problem is just exposed because we messed our data estate, this semantic layer.

13:42 Sebastian Selman Let's shift that angle of estate first, agent second and shift the discussion to a kind of competitive angle or a competitive read and think about who actually structurally is exposed by the shift that we just described. Right. I think every platform that doesn't own the data substrate underneath, that doesn't own where the ontology lives or the semantic layer lives, that doesn't own where the agent is grounded. That means you don't own the layer where your value is pooling. Meaning, if I look at Tableau, MicroStrategy as companies, I think they are structurally exposed. Historically, they didn't own the data platform underneath the BI tooling, which is exactly where the ontologies and the agents will pull the value from. Now, I want to be careful about how I mean this. I'm not calling their end. I'm saying they are structurally exposed and the ground beneath them is shifting. The platforms on the other hand that will survive, in my opinion, I mean, that's my hypothesis, are the ones that own the layer beneath. If you look at a Microsoft Fabric on top of OneLake, if you look at Snowflake with the data cloud plus Cortex, if you look at Databricks for sure with Lakehouse plus Mosaic, in each of those cases, the semantic layer, the governance layer and the model layer and the data layer. Are the same asset.

15:29 Estera Kot Yeah, and I've mentioned today Tableau and Power BI, which are present in every enterprise for sure, Excel as well. But I want to push on the Tableau specifically as Salesforce owns Tableau. And here one caveat when we review all of the data strategy, data architecture for other organizations is like puzzles, so many different icons expressing different tools coming from different vendors, different licensing model. And majority of those tools need integration. And because of this integration, we have a job as a company, we can help just integrate between all of them. But with Microsoft fabric, something is changing very much. It's about simplification and making sure that all of the layers from data ingestion, storing analytics and BI and AI are optimized for the same data sources. So here back to Tableau to be specific here, Salesforce is another very known, very rich company, which owns Tableau. And what Salesforce did, they acquired Informatica. And it's another tool for doing ETL. So extract transform load or extract load transform depends on the architecture patterns that the company wants to proceed. If Salesforce turns Tableau into the agent native surface on top of the substrate, they now own the CRM, the customer graph plus ETL tool, Informatica. So for the pipes that are moving the data to the end, Delta Data Lake. So the exposure here, close, doesn't it? Do you agree?

17:23 Sebastian Selman I think you're right. Salesforce is the one counter move I would watch. I would be super curious to know whether what I was discussing with you here was part of their thinking on why to actually acquire Informatica. Because that changes them from being a pure CRM player into owning the data platform and the agentic layer on top. That's a good one where probably the structural exposure is reduced. If they landed well, if they don't land it, we would still be right.

18:01 Estera Kot Yeah, so the condition is on just execution.

18:04 Sebastian Selman Exactly. I think our strategic call is watch Salesforce as the one potentially credible substrate acquiring counter move. And I think a bridge for anyone who heard the last episode in which we said fluency in a disrupted layer is not a moat. That was probably more about the individual contributor knowledge. But the same principle applies to vendors. Fluency in the disrupted layer being only the dashboard rectangle probably is not a moat for Tableau or MicroStrategy either. To come back to the two jobs from earlier. Job one, decision support. If agents answer questions from a governed model, the dashboard is rendering, not an artifact. That actually is good for decision support for the narrative question. The presentation you built to convince the board. LLMs for sure generate persuasive narratives at cost near to zero. Right. Content generation, especially written word content generation price is getting very close to zero. And that I think is a pricing attack on the BI market because the margin that supported the premium visualization tool. Collapses of what you were using that tool for primarily is narrative generation. So the BI tool spend as a category maybe doesn't even grow anymore. Whereas the semantic layer and governance spend is going to grow. The economies within the stack, I would argue, are moving in the future, which is exactly what I would tell you if you were sitting across from your CFO.

19:56 Estera Kot Yeah. And what about the humans, senior BI consultants, professionals who have built their entire career on being fluent in Power BI or Tableau or DAX. What about them?

20:15 Sebastian Selman Depending on how deep their knowledge goes in the stack, they might or might not be the most exposed. I think if you show a senior BI professional a dashboard generated by an agentic system today, you might get the answer that you can't compare the inconsistent terminology and the cobbled together visuals your LLM came up with against my beautifully tuned DAX model. That might be a defensive stance that somebody takes. But the trap is that it's potentially a correct statement for the present, but it is not for the future. Because the trap is that if you're only able to build DAX on a very well-built model, once the model is well-built, the underlying structure, the estate, an LLM will be able to do it just as well. Or putting it differently, if you are a highly accomplished BI professional only because you know how bad the data structure in your enterprise is, the defensibility of your position is not because you actually are a good data professional, but primarily because you understand the quirkiness of the data estate of your specific enterprise. Then the moment where that gets tidied up, your benefit is melting away, right?

21:41 Estera Kot So, yeah, some analogy just to explain in a very simple way. I have a feeling that we can put the analogy from, I love the iPhone analogy, meaning that the phone camera. So in 2007, professional photographers looked at the first iPhone photos and they said that you can't compare this outcome with their outcome. So to a real Olympus Nikon camera. And they were right for many years. And it has changed for the last decade. So basically they were wrong from that perspective, long-term perspective view, because people who were doing Photoshop jobs right now, switching to another job. Why? Because all of the smartphones, they are doing a pretty good job and not much is left for them. So I think that for the BI expert who built a career just on top of the DAX modeling, that person still has a choice, if that's not only their visualizations.

22:54 Sebastian Selman I think they have exactly that choice, right? They can either move up or down the stack, but anyways, into the data estate from pure dashboard building to actually engineering the semantic layer. Same person, same license, different altitude, right? And I think the people who will make the move will win the next decade. The ones who kind of defend the disrupted layer, the ones that cling to their film cameras of the BI estate, are the ones that are going to be disrupted.

23:28 Estera Kot Exactly. Let's summarize the lessons learned plus, to be specific, as is a business mission control for Swiss B2Bs, B2Cs and Swiss based leaders.

23:41 Sebastian Selman Exactly. So move one, I would say, is freeze new dashboard spend until the semantic layer is governed, right? Every hour you spend building a report on top of a bad model and a badly structured data estate is hours of future rework. Do the estate audit now, score the layer, then decide what to build. Move two is invest into the estate before the agent, right? That's kind of the flip side of the same coin. Copilot return on investment equals estate quality, right? It's not that copilot is amazing, therefore estate quality matters more than it used to. Copilot will reveal estate quality, but it does not create it. You need to do that first, right? Clean your data estate, get the fundamentals right. Ontologies will become a compounding advantage only on top of data that's actually in order. And then move three, repackage the BI team as a semantic layer team. It's not only about the visualization layer, it's about creating the semantic layer at top-notch quality. Probably will be the same people, but their job priorities will need to change slightly. It's a different altitude. The governance and the modeling are getting more important than the individual report writing. Probably the talent is already in the building, but the roles need to be updated.

25:15 Estera Kot For sure, I would add one more related to this battle and vendors' conversation. Maybe you have it. Tableau versus Power BI versus Qlik versus whatever. I believe it's not a conversation anymore about just a specific tool. It's a conversation about the ecosystem of tools integrated from the security governance and performance perspective. And also it has a caveat in the context of licensing, which is lower when you use a unified data platform. So a few services that are packaged.

25:54 Sebastian Selman Exactly. The conversation needs to move to the substrate, not just the interface and visualization layer. It's not just how you render what your substrate delivers. It's about what does your substrate actually deliver.

26:07 Estera Kot Just to do a bridge forward to the next episode, we'd like to talk about the engine. So once you have a semantic model behind it, there is specific engine, a few ones. And we got used to using CPUs for having this engine, SQL DW, Spark engine. And we want to talk more from a perspective of what will happen if I will change just one element. I will leave my semantic model, my data structure as it is, but I will just change the hardware instead of CPUs. I will use GPUs. How that will impact the performance. My bill as well. So we are going to talk about GPU acceleration that came to Microsoft Fabric Data Warehouse layer.

27:01 Sebastian Selman Exactly. So this was Business Mission Control. Show notes, sources and the public transcript are at our episode page as always. Link will be in the description. Full URL list for every post reference today is in the show notes. I'm Sebastian.

27:19 Estera Kot I'm Estera.

27:22 Sebastian Selman Next episode, we'll talk about GPU acceleration in Fabric Data Warehouse. Thank you so much for listening and see you in the next one.

27:29 Estera Kot Happy flying to Mars.