Why Context Alone Won't Make AI Trustworthy | CDO Matters Ep. 109

Speaker A: Hi, I'm Malcolm Hawker and this is the CDO Matters podcast. The show where I dig deep into the strategic insights, best practices and practical recommendations that modern data leaders need to help their organizations become truly data driven. Tune in for thought provoking discussions with data IT and business leaders to learn about the CDO matters that are top of mind for today's challenge. Chief Data Officers, Good morning, good afternoon, good evening. Wow, my voice is crackling. Um, welcome to another episode of the CDO Matters podcast. I'm Malcolm. I'm your Sherpa, your guide, your Host of episode 109. 109. Welcome to episode 109. I can't believe I've been doing this consistently for like four years. Think about that. Started, uh, in June, so three, just a little bit. Three. And at not quite a half. It'll. We'll be, we'll be, we'll turn four next June. June 2027, will turn four. Uh, my voice is crackly. I sound, uh, a bit, uh, stressed because I was, uh, in conference mode yesterday. I was speaking, uh, at the fourth annual Prophecy Data Hero Summit. Uh, this episode should publish late October. Be the 29th or 30th of October. Hey, Happy Halloween, boo. Um, should publish then. Uh, and because of the focus of this podcast, which is going to be my keynote from the Data Hero Summit, I'm going to share my insights from the keynote of the Data Hero Summit that is the topic of today. Uh, and that keynote is focused on the issue of trust, the unifying factor, the through line, the focus of the presentations Yesterday at Prophecy's 4th Annual Data Hero Summit, of which 2,500 people registered for 60 something countries around the globe. Crazy. Um, we've been doing it, obviously four years in a row. It gets bigger, it gets better. It's free virtual half day Data Summit. What more could you ask for? Uh, world class speakers. I'm not just talking about myself. We had, uh, Jordan Morrow come, uh, and join us yesterday in our sound studio in Buckhead in Atlanta. Jordan is such a powerful speaker, um, and such a well spoken, thoughtful and clearly knowledgeable speaker. We were thrilled to have him. Um, yeah, it gets bigger and better every year. That was yesterday. And I did a lot of talking and then I fly. Um, and it's, it's weird. It seems like anytime I do a lot of talking and then I fly, I think obviously it has something to do with the dryness of the air. I don't, I don't know, I'm guessing, but I sound a Little rough. So forgive that. Um, or not. Maybe you like it. I kind of like it. Um, but that does suggest that my voice maybe needs a, uh, maybe needs a rest. But hey, no rest for the, uh, the, the whatever they say, the weary, the wicked. Uh, I could be both, I don't know. But no rest for me because I wanted to regale, uh, the insights from the Data Hero Summit. Now, I know we're going to probably publish most of the sessions, uh, and you'll be able to check out most of the sessions, I, I believe, within the next couple ofweeks on prophecy.com. if you weren't able to attend the Data Hero Summit yesterday, maybe you were, um, and maybe you did, and maybe I'm about to tell you things that you already know because this will be two weeks after the fact. This will be around Halloween, and it would have been Yesterday, which was October 8th. So, uh, if you did hear the keynote already, well, maybe there's some more to take away from what I'm about to share with you today. I suspect most who listen to the poke for the podcast, some. I suspect there's going to be some overlap, but we don't need to belabor that point. It's still important enough that I think that I, I wanted to share a podcast episode about it, um, because I'm, I'm. It's rare that I kind of get excited about my presentations because I tend to be rather clinical about them. You know, I was a Gartner analyst, and it's, it's kind of like just the facts. Right? Give, Give me the, Give me the, Give me the info. Do it in a polished, solid story arc. Right? Uh, and, and, and share the insights, share the best practices and you know, uh, uh, and then, and then move on and do it again. That's kind of. I, I tend to take that approach in my presentations. Not because I don't care, not because I'm not passionate. Obviously I am. If you've heard me speak anywhere, clearly I'm passionate, maybe even excessively so. Um, so it's not a lack of passion, it's not a lack of those things, but, uh, that, that give me maybe sometimes too mechanistic or, or clinical approach to my presentations. But when you're Gartner analysts, they teach you how to give presentations and there's very much a formula, there's, uh, very much an approach, there's very much a cadence. There's all sorts of things that you need to do well to be a good public speaker. I feel like I'M reasonably good at it. Um, there's certainly things that I can improve on. Um, but I do take them very seriously and I approach them very professionally. And I tend not to get that excited. I figure out what I need to say. I think that there's value in what I need to say, and then I give the presentation and that it's kind of like a, maybe this is maybe a bad metaphor because I'm not nearly this important or this talented, but like a doctor going through a surgery, right? You, you, you do the thing, you do it, or, or, or, or any experienced trades person or any experienced, anybody doing what they do well, and they do it and then they do it again, and then they move on and they do it again. And it's not like, you know, a doctor is going to get excited about one gallbladder surgery versus another gallbladder surgery. Right. Um, and maybe that's kind of how I tend to look at things. But when it comes to the Data Here Summit keynote, I'm really kind of excited about it. And I think it's partially because I think it's a necessary part of the conversation we need to be having right now as data leaders. And that's the importance of meaning. Uh, you could substitute the word context for meaning. Meaning is critical, right? That's, that's the context problem that we are all trying to solve right now. What do words mean? One of the best ways to convey meaning is through the lens of context. You know, that something means something in relation to something else, right? This is, this is a knowledge graph. Uh, any, any sort of data that is hierarchical in nature or insight that is hierarchical in nature, including a knowledge graph. This is related to this, right? That provides context, right. That'll help you understand the broader intent of what you're trying to convey or the broader environment of the thing that you're trying to describe in which it exists. That's the context problem. I would argue that is synonymous with what I'm just loosely calling meaning. And we're trying to solve for meaning right now. It's important, it's critical. Uh, obviously, uh, Gartner, in its keynote sessions this year, ah. At its data and analytics summits around the globe, had a 1 slide. It said context is king. And that has sent context to the top of the Gartner hype cycle. Everybody's trying to figure it out because it's important. Right? But I would argue the underlying thing here is, um, is the concept of meaning. What do things mean? The. Well, I'm just going to get into it. I'm done. My preamble. Here we go. We're going to talk about my keynote, which was the importance of trust. What are we ultimately trying to do as data leaders when it comes to the infrastructure, systems, processes, data that we are providing to AI? Right. What are we trying to do? We are trying to enable AI to act in a trusted way. Right. I think, ultimately, I think that is the end goal. Right. Will we ever get to 100 trust? Of course not. Humans don't have 100 trust. Right. If you're interacting with somebody, do you have 100% trust in somebody? Well, maybe, maybe your significant other, maybe some of your family members. But I think we're not aiming for perfection here, but we are aiming for trust. Ultimately. At the end of the day, that's our business. And I know that sounds a little pithy, but if you're creating insights or analytics or dashboards or data and your consumers can't trust it, you're done. I mean, you're done. The most beautiful dashboard, the most eloquent insight, the most transformational insight. If your customers just can't trust it, well, uh, it doesn't matter. You're done. So I would argue we are in the business of providing trust and what is needed to do that. Right. So the title of my keynote was what is Trust Made Of? And I, I think that's, I maybe over pithified it. Little too pithy, uh, a little too sound bitey, a little too clickbaity maybe. I don't know. But that's what we're trying to solve for. And that was my keynote and I will tell you my, the theory, the through line, the talk, the primary talk track of my keynote, and then we'll kind of back into that through, through a little bit deeper discussion. Oh, and by the way, it's just going to be yours truly today. I'm not having a guest. This is my prerogative as the host. Not, uh, because I don't want to have a guest, but I'm just. Everything is so fresh of mind because I did this yesterday in a sound studio in Buckhead in Atlanta, Georgia, and I just wanted to share the insights today. So I, I thought it was appropriate. Uh, that's why I'm, I'm visually not prepared. I'm not in my prophecy uniform, um, and you don't see my prophecy LCD sign behind me, um, for some other reason. I don't know why my recording platform didn't add. It usually does. I don't know what happened. My background, who knows? A streamyard update more than likely. Um, so, yeah, I, I'm, I'm flying a little bit by the seat of my britches here, but let's talk more about trust. The through line of my keynote is that there are two key ingredients to trusting anything an enterprise system tells you. Whether that is AI, whether that is an API, whether that's a dashboard, whether that's an insight, whether that's a CRM, ERP doesn't matter. I would argue there are kind of two key things, right? And if, let's let. But, but the example that I used was analytics. You know, uh, insights through analytics. And my preposition, my assertion is that you need two things. You need, meaning the system needs to know what you're asking, right? The system needs to know the con you're operating in. So meaning is one thing. And when you're talking about analytics, if you're asking a question or looking for an insight, the system needs to know, uh, what you mean. It needs to know what you mean by active customer. It needs to know what you mean by termed customer. It needs to know what you mean by supplier and the difference between a supplier and a vendor, and on and on. That's the issue of meaning. The other thing that you need in the trust equation, 50% of the trust equation is truth. So you need to have meaning and you need to have truth. And they are equal co conspirators in, in this equation you must have both. And one without the other isn't going to cut it. In fact, the phrase that I had in my presentation is that meaning without truth is still confusion. You can have the world's greatest semantic layer that has defined every object KPI attribute in gory detail, right? Every nugget of metadata defined, cataloged, lineaged, even, um, you can have all that stuff well defined. But if you put, uh, create a dashboard where you show five different versions of Acme Incorporated as active customers, how is the customer of that ever going to trust the data? Is this one Acme 4, Acme 3 Acmes? The phrase I use in my keynote is if you pull a report and it shows 200 customers, is that 200 real customers or, you know, 80 customers entered 200 different ways? That's the core question. That's the issue of truth. And we've all been chasing after context for the better part of three years now. It's the top of the hype cycle. And every vendor somehow is in the context business. Interestingly, go figure. Uh, keynote session at Big Data London. My Friend Mike Ferguson shared the idea of what he was calling the Context Wars. Makes it sound very Star Warsy. The Context Wars. Good. Good grief. Right? Everybody's, everybody's got a, uh, a stake their claim in, in, in the context. Land grab data catalogs, semantic layers, even, even increasingly now, um, you know, ontologies being provided by lakehouse environments, uh, knowledge management infrastructures, content management systems, and on and on. Everybody, every. The semantic layer baked into your bi. Right, Everybody's got a claim now on the context, uh, wars. But I would argue, and I do strongly, that that's only half of the trust equation. It's only half. It's only half. You can have the greatest medical area in the world that's not going to solve for inaccurate underlying data. So the metaphor that I. It's not a metaphor. The actual story that I use throughout my Data Hero Summit keynote is the story that Anthropic itself told in June of this year, in June of 2026, about their journey to enable Claude to become their interface for their BI natural language interface. Ask your data anything through Claude. Hey Claude, I want to report or hey Chatbot. Hey, whatever. Doesn't matter. The example I gave in my keynote is I created a fictitious chatbot named Hugo Harmonize, Unify, govern, Orchestrate acronym. Hugo. Hugo just happens to be one of my dogs and he's the most adorable dog ever. Um, so it doesn't matter. Hugo, your analytics bot, Claude. OpenAI, Kimmy, who cares? Doesn't matter. Uh, but in the case of Anthropic Claude, of course, because that's their product and Anthropic shared the story, um, of how they actually enabled Claude to be their front end for all their analytics. How they went about creating porting Gen AI as the interface to their BI layer. I didn't think it was possible. Uh, I'll be honest, I did not think that was possible. I didn't think it was possible to take this inherently probabilistic system and make it more deterministic. But Anthropic regaled their story in detail in a blog post in early June of 2026 where they said exactly what they had to do in order to get to a level of trust where they could actually use Gen AI as their burst tableau Qlik Power, bi whatever. By the way, separate issue about the long term viability of BI tools. About a year ago I was saying, ah, uh, they're probably not going to go away. I don't know. I don't know. Um, I, yeah, I, I see. Natural language queries being the future. Now, will there be a BI tool sitting underneath that based on what Anthropic did to solve the problem? I'm not entirely sure. So what did Anthropic do to solve this problem? How am I so certain that meaning and truth are both equal partners in the trust equation? Well, Anthropic certainly created a semantic layer. They created some idea of a semantic layer. They provided the ability to inject through a complex retrieval pattern, uh, to inject context into every single query. So yeah, they built a semantic layer, some idea of a semantic layer, and they built a repeatable process around it. But what they learned was exactly what I just was talking about. What they learned is that at a certain point, once you had provided sufficient context so that the LLM knows what you're talking about, that the LLM knows the difference between an active customer and a previous customer, or an active customer and a prospect, whatever, right? Once sufficient context has been provided so that the system knows exactly what you're asking, more context doesn't help move the truth equation. Doesn't help address the issue of truth. Anthropic proved this themselves. What they did was they added insights from thousands of previous queries for any one query, and it didn't move the needle when it came to the trust equation or the truth side of the equation. They had great meaning, they had great context. The system knew exactly what an active customer was. But it couldn't reconcile Acme versus Acme Inc. Because it hadn't provided what Anthropic calls a canonical reference source. That is the truth side of the, uh, trust equation. So, meaning yes, but, uh, no truth. So what did they do to build the truth layer? Well, they did all the things that data governance and data management people have been doing for the last 30 years. They built data quality rules, they built data stewardship processes, they defined data stewardship process, they built a canonical reference source, AKA Master data management. So to solve for the truth side of the equation, they built infrastructures, processes, repeated repeatable scalable processes. This wasn't a one and done. This is, this is a program, right? Repeatable scalable processes to enable their LLM, um, to understand if Acme and Acme Inc. Were one thing or two things. And this is beyond just entity resolution, my friends. This is a single source of truth. That's what they built. A canonical reference source, a list of known customers, a list of known suppliers, a list of known assets for all the domains that would be a part of these queries. And it's in essence all of them all the most critical ones, or put another way, all their master data. Right? So that was my keynote. My keynote was, hey, how do we deliver trust? Well, we have an example and the example just happens. How do we get deliver trust in an agentic world? That, that, that's, but, but arguably how, how do we deliver trust? It's, it's how we've always delivered trust even even without agendic AI, even without AI in our BI, in our dashboards, in our analytics, our APIs, everything. The formula that I describe is, that is still accurate whether it's AI or it's not AI, right? You need, you need to have some idea of meaning and you need to have some idea of truth. I need to know what you're asking for, right? And I need to provide you insights that are, uh, accurate. Simple stuff, right? So we had an example, anthropic, solved it, and lo, uh, and behold, what they did to solve it was replicate everything we've been doing as data professionals for the last 30 years, literally lineage, quality, governance, stewardship, all of it. And they did it from a first principles perspective, meaning from the ground up. No prior assumptions, no ax to grind, no consultant sitting in the room saying, hey, you need X or you need Y. No vendor sitting in the room saying, hey, you need X or you need Y or you need my product or you need my services. The smartest AI engineers on the planet set out to solve for the problem of trustworthy insights using an LLM on top of uh, a legacy analytics infrastructure on top of all of their raw data. And what did they do? They ended up building everything that we've been doing. And what do I mean by that? Data management. The broader wheel of data management, which includes data governance, which includes some idea of metadata management, master data management, reference data management, data quality integration. They built an integration semantic layer. Meaning of things, they rebuilt everything that we've been doing, I would argue, without even knowing that it was always there. This is a phenomenon I'm seeing more and more and more. I'm seeing it in Snowflake, I'm seeing it in databricks. We're really super smart. Data engineers are given the problem of make this data trustworthy and what do they do? They go hand roll some Python that does everything that I just said or tries to do everything that I just said. It can't necessarily do all the things that I just said. Stewardship, metadata management, reference data management, master data management, data quality integration. So I'm seeing this more and more and more where you have people that are trying to solve these problems without actually acknowledging that these tools exist, these disciplines exist, that these teams exist and oh well, I'll just go hand roll it in Snowflake or hand roll it in data bricks from the ground up. Um, in the case of Anthropic, they hand rolled it from the ground up. But I happen to think that that is a massive validation, right? In fact all of these are validations, right? If you're going to like, custom code, all sorts of transformations, um, bake in data quality rules, use Python for entity resolution, like do all the things that you could turn to tools to do that you could turn to existing people to do that you could turn to existing processes to do. But if you end up building everything from scratch, which I don't recommend, but separate issue, and you build it all from scratch, and where you end up landing after you've built something from scratch, is it the exact same place that you would have been had you just done, had you just used the data quality tool, the MDM tool, the integration platform, had you just done all that stuff at the beginning? That's a validation of that other stuff, right? Like if Anthropic says you can't do AI, you can't use Genai as a natural language query on top of your analytics. You can't do that without a canonical reference source. Which is exactly what they said. Which is exactly what they said. That's a validation of the need for a canonical reference source, AKA Master Data Management. You use mdm. MDM is the discipline, the output of that discipline is a reference source. I know this thing to be true because it is governed with specific rules that are repeatable and explainable. It's managed, it's wrapped in a workflow to make sure that the data doesn't decay and it stays current over time. We apply human beings to it, sometimes, not all the time, but sometimes. In the case of Anthropic, what they did was that any analytics, any output that was headed to the C suite, they had humans reviewed before it went to the C suite. So yes, even stewardship, even human in the loop. Smartest AI engineers on the planet, reverse engineer data management and validated that it was important all along, including mdm. So yeah, a little bit of a victory lap. Hip hip hooray for those of us amongst us like me, who have been doing this a long. Time. So that's a validation. So a little scary. And why? Well, because the people at Anthropic, and I'm increasingly seeing people, particularly in the databricks environment, redoing everything from scratch, from the ground up, without ever knowing we were there all along. Building what in essence isn't like MDM is exactly mdm, okay? A canonical reference source defined by governance policies, defined by business rules. The canonical reference source that is created by a known set of business rules, right, that is managed by stewards, that is wrapped in some idea of a workflow to make sure that that data stays current over time. That's not like mdm, and that is mdm. So the concern is, is you have a growing number of data engineers, data scientists, AI engineers, whatever the titles are, that are increasingly oblivious to the fact that the data management function exists. Or at least there's evidence to suggest they're oblivious. Uh, that exists because if they weren't, maybe they would have just done that from the beginning or gone and had a conversation with me. It's like, hey, wise guy, how do we go create trustworthy outputs here? Well, I got some ideas because it's what we've been doing literally for 30 years. So yeah, it does sting a little bit. It's a validation, but it also stings a little bit. It's like, uh, how did you not know this? Or why, why would you, why would you, super smart data engineer, spend eight months and a lot of tokens rebuilding an MDM from scratch using Python to get 60 or 70% of the capabilities you could have got from, uh, a platform that would give you 100% of the capabilities out of the box on day one. See it all the time. I see it all the time. Uh, this mindset, I just, I just, I, I've never understood it. I've never understood it. Right? It's the build versus buy, but without any due diligence, right? Without any due diligence, it's just build. My God, what do I have to do? Oh, uh, man, I gotta, I gotta make sure that the data is consistent over time. I gotta, I gotta kind of understand the, the lineage. I've, I've gotta define some rules to make sure that it is, is actually accurate. And I'm not just talking about, you know, uh, is it an integer versus a varchar? I need to make sure that it's accurate. Yeah, uh, I need to layer in maybe some idea of uh, human review here for high stakes. Like if you wrote the requirements for solving this problem, if you wrote out the requirements for solving this problem, they would literally be a carbon copy of the DM bok, the Data Management Book of Knowledge, which is like the The Anyway, if you, I assume you know what the DM BOK is, maybe you don't. It's like, you know, it's the reference source for the discipline of data management. So that part is flummoxing to me. I'll be honest. That. That part is flummoxing to me and it always has been. Because it's not new. It's not. It's not new. Right. This idea that the only way possible that I could solve such a complex problem is to engineer it myself from the ground up. Yeah. Okay. Anywho, back to the issue at hand. My keynote. So anthropic. Anthropic. Okay. Like Frontier AI Lab said, hey, to get trustworthy behaviors from these probabilistic systems, here's what you need to do. And the here's what you need to do is data management, data governance, data quality, MDM lineage and yeah, semantic layer. Yep, that too, but it's both. So it's not just a semantic layer. So I'm about to burst some bubbles and I've already said it three times in this podcast, but I'm going to say it again and if I'm bursting your bubble, forgive me, but the semantic layer alone will not deliver trustworthy insights. It will not, because it needs the truth. It needs some idea of the truth. And how does truth get into the machine? Well, historically for most enterprises beyond a certain point and the cutoff varies. But once you get to the point of having multiple databases, multiple silos, complex workflows, conflict processes, MDM is how truth gets into the analytics machine. Now we can have an argument about whether you need an MDM platform to do that. Do I need MDM software like Prophecy's amazing software to do that, or can I do it some other way? We can have a conversation about that. Most certainly. There are plenty of small companies out there that are doing MDM right now and are doing it without MDM software. I'm talking about the discipline here, my friends. I'm talking about truth gets into the machine through the idea of mdm, through the discipline of mdm. And that's what anthropic reverse engineered from the ground up. Uh, so more context isn't going to solve that problem. You still need to solve. And ah, you can call it identity. It's more than that. But you can call the identity problem if you want. You can call it a data quality problem if you want. Do. How do I know that Acme Inc. And Acme Co are one thing or two things? I m do need to Say, because this is true, this is accurate, the insights that are created through ontologies and knowledge graphs most certainly, most certainly can be valuable insights into assessing the quality of something. There's a lot we can learn from an ontology. Relationships can start to infer the accuracy or inaccuracy of things. They can allow you to start inferring it, most certainly. I mean this, this is, this is like Philosophy 101, right? If ducks have feathers and all ducks have beaks, then, or, uh. Wait, or I'm failing, right? If all, if, if all, if all people. If, if, if, if two things share two attributes, then maybe they're related, right? My brain is much. I had a long day yesterday. I had a, I had a long day yesterday. So, so forgive me, but I, I hope you understand where I was going with that. Right? If two things share common attributes, that's, that's, that's an indicator where you could infer that maybe they're the same thing. Right? That if, if all ducks have beaks and you know, uh, all ducks quacked in, anything with a beacon that quacks is a duck, right? I think I got there. Maybe, um, you, you could, you could infer that anything with a beak and anything that quacks is, is a duck. You could, um, and if you did that enough at scale, uh, I think, I think you could, I think you could reasonably deduce sometimes that they're the same thing, but that's not the same. Uh, uh, so, so that's entity resolution, that's identity management. And I think increasingly ontologies and particularly knowledge graphs can certainly help with that and maybe even be the engine, maybe even be the engine for complex identity resolution. Maybe. All right, most certainly at a minute, at a minimum, they can, they can most certainly help inform the decisions. Most certainly. But it's all the things you need to wrap around that a lot of people who are just so high on, um, to be kind of missing, which is the idea of business rules, which is the idea of consistency and predictability and determinism, right? Yes. No rules and the ability to infer something is not necessarily the ability to say yes, it is or is, it isn't. And Even if you're 99% confidence, confident in your inference, that's great. But how do I put rules around that that are manageable, configurable, and provide the explainability that we need out of these systems? That's key. So it's not necessarily that MDM is like the be all and end all. It's the rules that are created through a governance process. Right. And three explainable. You can explain them. Here's the rules. And when your auditor comes knocking, uh, why is ACME Inc. And ACME Co. Now one record. You can explain them. Here are the rules that I used. Now, we can argue about whether those are good rules or bad rules, but those are the rules. So that's, you know, I, uh, I don't. It's not like I'm going to sit here and say, oh, if you're using an ontology or knowledge graph for entity resolution, that means it's not mdm. I'm, uh, not going to say that because it wouldn't be true. The real value of data management, including data governance. The real value here is those things that I just talked about, rules consistent over time, that are explainable. That's, that's, that goes into the concept of truth, what I'm calling truth in my keynote. Meaning, how do I define things? Truth. Is it accurate or not? So much of the truth equation is consistency. Right. And so much of the trust equation is consistency. Because you've got different answers over time. We're all human. We're going to be like, wait a minute, yesterday it was red, now it's green. What? How do we go from red to green? And if you've been in this game long enough, like me, you've been in the situation where you've had to defend that. Why am I seeing what I'm seeing? Here's your report. Thanks for the report. Huh. That doesn't look right. Why am I seeing what I'm seeing? I mean, anybody who's been doing this more than like a few weeks knows what I'm talking about. And can you stand behind it? The worst thing possible. The worst thing possible would be to say, well, the worst thing possible is to say, I don't know. Right. Because that's not confidence building. That's not going to build trust with your customer. Customers. The second worst thing you could say is the robot did it. LLM decided. Uh, what? It's this idea, I think, that is killing so many POCs. Right. It's, it's not, it's, it's not. It's not that these things are producing inconsistent results per se. Right. It's not that they're producing inaccurate results per se, is that we just can't explain it. Yes. This touches on things like lineage. Yes. Even maybe the idea of observability. I'm less and less fond of that word, by the way. Um, observability. Because, because I think, I think, I think it just comes, it just comes down to explainability. Why are things how they are? Do I care that, uh, I'm explaining a dashboard or why a pipeline broke? Observability. I don't know, I just, I'm less and less enthused with that word because it just, to me it's just explainability whether I observe it or not. Can I explain it? Yeah. Rant. Anyway, so, yeah, that was my keynote at the Deity Hero Summit yesterday was talking about the improve, uh, the importance of truth and meaning. Truth and meaning together are the core pillars, the two core pillars of trust. No truth, no trust. No meaning, no trust. Must have both. I know why I'm doing weird things with my hands. If you're just listening to me. I don't know, I'm, I'm doing little Trumpian things with my hands and I don't know why. So that was the keynote and it was all couched through the story that Anthropic told in their journey to enable trustworthy BI using their AI infrastructure. So yeah, I mean this is, this does this. We're, we're back to our dashboards. Dead. Well, dashboards certainly aren't dead. We'll just use LLM to create them. We'll just use LLMs to create them. Um, and if you're a BI provider, is there still value to be had as that layer between a raw, between the raw data and the consumer of the data? Well, everything that I just said and everything that Anthropic experienced and documented in their journey to use CLAUDE as their BI tool certainly suggests that, that the BI providers need to move closer to data management and data governance, need to pivot away from the front end and towards the back end. But here's the problem. There's already no vacancy. There's already people doing that. Right, Like Prophecy doing mdm, like others doing data integration, like others doing data quality, like others doing actual semantic layers. Yeah, the BI providers are building some of those semantic layers. They always have. You've always needed to know what active customer means, what's the definition there? But there are others. Even data catalogs, for heaven's sakes, are arguably could be used as some facsimile of a semantic layer. And I think they should, frankly, I think they should be the foundation of a semantic layer. Not the only thing. I see semantic layers as being multi layer cakes and kind of at the baseline of that cake, the bottom layer of the cake definitions, what do things mean? Right. Um, so if you are a provider of a BI tool. If you're a vendor of a BI tool, I think you're going to need to find a way to go, um, downstream, right, to get closer to data management, to that neighborhood. But that's a crowded space already because you're not going to go upstream, you're not going to get closer to the end user, you're not going to go into visualizations. I mean that's LLMs. I mean they're good at that stuff. They're really good at that stuff. Turns out they're actually pretty good at writing SQL too. Um, that's what Anthropic found is that increasingly the LLMs are really good at writing SQL, but you just can't let them loose on raw data because they can write SQL all day long. They just don't know the difference between an active customer and an account. And they don't know that when they go query that. And if they pull back 200 rows of customers, they don't know that Acme Inc. And Acme Co are two different things. So our dashboards dead? No, we're going to continue to use dashboards. Of course we're going to continue to use dashboards. We're going to have to. But will they exist from a BI tool provider or as projects in Claude? Or as skills in Claude? I think it's going to be more. The latter certainly seems that way. And it certainly seems like Anthropic solved that problem. I don't know. I don't work at Anthropic, um, which increasingly seems to be maybe a good thing. Uh, I don't know. If you do, let me know. Maybe I'm wrong. So I don't know how successful they are, but I mean they, their, their blog seems pretty, pretty solid. Right? About what they had to do in order to do this, in order to enable CLAUDE to be their interface to their analytics layer. So that's it. Uh, my friends, that was my keynote and I ended on a high note because basically what I said was, which I believe to be true, which I believe to be true is a massive validation of the things that you and I have been doing a long time. The data governors, data managers, data stewards, data owners have been doing a long time. So this is the stuff that Anthropic did, is a massive validation of that Data governance, data management are not going away, of course. They're increasingly being augmented. So instead of humans building the data model, AI will recommend it. Instead of humans defining the data quality rules, AI will recommend them humans will ultimately remain accountable for these governance policies. And that's what gives me hope for the short term, for at least the next five years that I don't see that going anywhere. In fact, I see the demand going up. Why? Because according to Gartner, we're only really managing and governing about 20% of the data state. The other 80% is unstructured data. So once we start figuring out how to govern, steward, do data quality for that, whoo. That's just even more demand, my friends. So I I'm bullish. I am bullish. Data management, data governance. Is not not something you can hand wave away. Tropic Think you're in a good space. So with that, I will bid you ado what's coming up on the podcast calendar here if I stick to this as episode number 109? Oh, I'm going to be talking to, uh, Gabby Steele from prequel AI. That's cool. She does semantic layer things. That's pretty cool. I'm going to be talking to Sean Rankin from Huntington bank about their governance and MDM journey. Very, very cool. They've done some really cool stuff at Huntington. Looking forward to sharing that story. So, hey, if you haven't subscribed to this podcast, please go ahead and subscribe. I do this every two weeks and have been for nearly four years now. Sharing insights, data management, data governance, data and analytics, AI, all things related to data. Subscribe like tell a friend all the things I hope to see you on the next episode of CDO Matter sometime very soon. With that, I will bid you farewell. Bye for now.