Latest / Machine Dreams / The Hidden Cost of AI Nobody Is Talking About | Collibra CEO Felix Van de Maele

Transcript

The Hidden Cost of AI Nobody Is Talking About | Collibra CEO Felix Van de Maele · Machine Dreams

• speaker-0: Yeah. Everyone has an opinion about AI, but very few people know how to make it actually work inside a real company with real data and real consequences when it gets things wrong. Welcome back to Machine Dreams, where we speak with the people on the inside who are actually figuring out this works, what doesn't, and how that could change our world. My name is Kalawali Samuel Adibaya. I'm a technology analyst covering AI, cybersecurity. and enterprise technology. And as always, I'm joined by my remarkable co-host. speaker-1: Hi everyone, I'm Leia Stern, global communication strategist and former venture capitalist. Hey Sam, how are you? speaker-0: Very well. Hi Leah, it's it's it's good to see you. ⁓ we're back in season two and it's shaping off to be really exciting, you know, so I'm excited about what what's gonna happen this season. speaker-1: Yeah, and it's it's great to be back at it. ⁓ today we're joined by Felix Vandemal, co-founder and CEO of Colibra. It's a data intelligence company valued at more than five billion dollars. Felix started the company back in two thousand eight and has spent more than a decade helping some of the biggest organizations in the world get their data in order, which as it turns out is the thing that decides whether all this AI actually works. Felix, welcome to Machine Dreams. speaker-2: Happy to be here. Thanks for having me. speaker-0: Thank you for joining once again, Felix. ⁓ so let's start simple. Everybody who's listening or watching now, almost everybody, has probably heard about AI making mistakes. But for someone who doesn't live in this world, who's not, you know, listening to AI or talking about AI every day, what does it actually look like day in a day to day when AI makes mistakes for a company or a business? speaker-2: Yeah, absolutely. So I think by now most of us probably have exper experimented with AI in our personal lives. I think we've probably all used some form of AI tool. Now, ⁓ often the the stakes aren't as high, right? We were kind of using it to kind of ⁓ ask some questions, but in a in an organizational setting, especially if AI use cases become more advanced, more mature, ⁓ the stakes get a lot higher, right? And if you think about when does AI truly impact ⁓ the productivity, impact the business. It's it's when we can trust it. And this is to your point, Sam, exactly where often companies are really struggling. ⁓ getting AI to answer simple questions is one thing. Relying on AI to actually take actions, make decisions that you can trust is a whole other thing. I think where where we see the most challenges today is is having to double check what AI, what an agent is actually doing. Now imagine if you have to double check everything that's happening, of course that's not going to scale. And that's a real Tax on organizations. We we think of this as almost a the hallucination tax, right? If if AI hallucinates, if we can't trust it, we have to babysit it. And that's not the intention. And that's a real struggle that ⁓ that organizations absolutely have to fix. And speaker-1: That being said, I know the firm is trying to position itself as the NERV center for AI agents with your launch of AI Command Center. Tell us more about how that works. speaker-2: Yeah, think of Qualibria as this kind of AI control plane, right? And ultimately it's all about kind of trust and confidence in AI. And there's a couple of components to that. If you think about a a an agent, an AI use case, ⁓ it's really important to ⁓ have ⁓ trust and confidence and quality of what you give to the model, and then have have really a lot of confidence in what the model does with that data to actually what it outputs, what it decides, what what what it what it provides. And so if you think about large organizations, they need to govern, they need to trust both sides. They need to trust what goes in and need to be able to trust what goes ⁓ what goes out. And so if you think about what comes in, what what typically the industry calls that is context. And we know by now that context is absolutely critical. Because the models have gotten really, really strong, right? ⁓ and I think it's it's been clear ⁓ by now that the models are no longer the constraints to drive AI adoption in organizations. The models are really strong. Think of it almost as a ⁓ A model getting to almost a PhD level intelligence. But we all know it's not a guarantee that if you hire a PhD as an employee, if you make that analogy between an agent and employee, which obviously is not perfect, but ⁓ if you hire a PhD level of qualified or educated employee, that doesn't guarantee that that person is gonna perfectly know what to do and get have it correct all the time, right? They might be really smart, but you have to train them, right? And that's exactly what companies need to do. The models are really smart. But they have to train how the model should operate in their organization. And that doesn't go away. And that's the biggest bottleneck, right? If we don't do that, either the model doesn't know and they'll hallucinate and you have that hallucination text. Or or or they do, but then it can become really costly. And that's the other big theme that we've that we started to to hear about. Like this token consumption becomes a big topic. Tokens aren't free, AI isn't free. And so it's not just about deploying a lot of AI or call it token maxing. It's all about how do we use tokens effectively and ⁓ and efficiently. And as as ⁓ as many things, ⁓ at least in my life, ⁓ it's a data problem. ⁓ this is again a ⁓ a data problem, but it's interesting again going back to the analogy with let's say the the let's say the PhD ⁓ educated person, ⁓ the trading the model requires lots of data. The more data, the better, right? And we all know OpenAir Anthropic, Google, they're training on the entire internet. And almost the more data, the better. And that's becoming a bottleneck. Now, within an organization, it's not about training the model, it's about training the model how to how to operate within the organization. It's not about more data, it's about the relevant data. It's about high quality data. Accuracy, relevance becomes very, very important. ⁓ and that's a real struggle ⁓ and and challenge that companies are are now facing. speaker-0: Great Felix, there's there's a lot to unpack right there. But let's go back to what you said about babysitting AI, which is, you know, the old the old ⁓ conversation about hallucination attacks. ⁓ on the individual level, a lot of creatives come to me and they say, Sam, we're using this AI chatbot. You've been talking about using AI chatbots for our creative processes, but now we're finding that, well, when we use these AI tools for the things that we do, we have to now go back to check out the results or outputs. And we're spending more time actually fixing these mistakes. ⁓ rather than w more time that we would have actually spent if we you know just did these things ourselves. And I'm thinking about what that could mean on an enterprise wide scale. You know, and you just ran a survey with Harris Pohl and, you know, one of the numbers that jumped at me was that in the survey it said that more than half of the decision makers that were interviewed said that they personally have to step in and to f to to to fix AI's mistakes, which is alarming when you think about it. So ⁓ can you really Tell us about this hallucination tax and how businesses, individuals even can get around it. speaker-2: Yeah, and that's exactly a a problem. If if again, if if fifty percent of the people feel like they need to babysit ⁓ what the AI is doing, it's not gonna work, not at not at scale. So we have to we have to fix that. ⁓ one is what I just talked about is providing the right context, like training your model to know how to use the data, what to do, and that's a one way to really improve ⁓ improve that that number so it doesn't hallucinate as much. It's just you need to train it. ⁓ so the hallucination goes down. The other ⁓ I think important, and that's what you said around AI governance, is you can't guarantee that it's never going to make a mistake, right? And so you have to put in place guardrails to make sure if the model doesn't know what to do or if it does something that it's not supposed to, it doesn't require a human to kind of validate everything, but you can do that at scale. And this is also why AI governance becomes really important. And this is exactly what I'm what I'm seeing at at large organizations where where they are struggling with that hallucination tax. It's not just about it requiring people. It's also about like we can't see everything. And so if there's a risk that it does something that's not supposed to do, if it's if it's handling customer interactions, we just can't take that risk. So we need to put in place guardrails that if something happens, we we are a hundred percent sure that we can ⁓ we can stop it and it aligns to OI policies and how we think the AI should behave. speaker-1: You're located in ⁓ Brussels and I'm just curious, based on how you see innovation and also regulation in this space, does it affect how the company navigates where it's going, how it's moving forward compared to let's say the United States? speaker-2: We definitely see in general a bit more regulation in in Europe. We've seen it historically with data privacy, there's more security, ⁓ data regulation, and of course on the AI side as well with the EU AI Act. But I would say that that's not the biggest driver. ⁓ I think there is going to be regulation. I think there is regulation around AI today, everywhere, right? In Europe, in Asia, a in North America, because the reality is AI is just using data and there's already data regulation. There is data privacy regulation, there's data security regulation today. With that to an algorithm to an AI model, that doesn't really change. So everybody's kind of dealing with that. We also, I think it's clear that just ⁓ the impact of AI is so big that there are, and frankly, there should be a level of regulation. Now we can argue how and and how to do it effectively. That's not that's not as simple. ⁓ but it's going to be regulation. And so one large organizations are absolutely ha planning for that. They know it's frankly irresponsible to roll out AI without visibility, without control. Without the policies, nobody, nobody is doing that. Whether you are in Europe or you are in in in the States. But I will say I've seen a significant change over the last 18 months or so with customers that let's say 18 months ago it was more of a check the box. Yes, we need to regul ⁓ we need to comply to new regulations, and obviously we'll we'll do that. But now it's it's not they're not just concerned about the regulation. They're much more concerned about concerned about is my AI use case, my model, my agent actually doing the right thing? Because the risk of it. not doing the right thing, if it's interacting with customers, if it's doing something important, is much bigger than the risk of of regulation. So a the the the best approach is again to drive that control, not just for regulation, but really to make sure it it delivers the business outcomes ⁓ you want it to. speaker-0: Yeah. ⁓ Felix, I I think that one of the big messages that many people now resonate with in the industry is this idea that good AI is all about good data. I wouldn't say everybody knows that, but a lot of people now know that. But I think that you're now pushing it even farther and saying on the enterprise level, it's not just about how good the data is, but it's also about the relevance of that data to what you're doing. I I want you to kind of unpack that. And tell us what you mean about the r the relevance of the data, because there's there's a lot of people who are listening now and there's a lot that they probably can take away from that. speaker-2: Yeah. Absolutely. And and it's data, I think we often make the mistake to think of data as a volume problem. Just more data is better and we have to find a way to deal with more more data. And of course, more data's often better. But it's not only a a volume problem. And especially that's especially true with with AI. Like I said, when you're trading a model, it often more is better. When you're ⁓ in ⁓ or supporting a model in in an organization, it's it's much more about relevancy, but also having that that broader Understanding, they call it context and knowledge of how the organization works. What when we say a customer, how do we define a customer? Typical example, right? The AI, the model doesn't know that because that's very specific to our organization. If we don't teach the orc the model what we need with customer or churn or or or any kind of metric, it's not gonna know. And so the the semantic grounding, or people call it the ontology, people call it a knowledge graph, there's a lot of terminology, people call it the your company brain, right? Again, lots of kind of terminology that people use, but it's so important to build that broader kind of ⁓ company knowledge to ground your model in your organization's reality. So it's not just about the data in the traditional sense. Obviously we need to connect our data to the to the model. It's a lot about ⁓ can knowledge, things that are typically locked away in emails, policy documents, processes, SharePoint, confluent, team, Slack. That's a broader knowledge. ⁓ becomes equally important. And the way ⁓ companies are are dealing with that is again and and again it's interesting because it almost comes full circle with what we've been doing for so long is we need to build these semantic models, these ontologies that kind of capture what our data means. And so so we can train our models and our agents so they interpret that data correctly. Otherwise again they'll hallucinate and we'll have the same problem that we just talked about. And speaker-1: You and your firm are kind of branded as, you know, seeing ahead of the future. You've been in the game a while. You you kind of envisioned what was happening in the space. You've worked with McDonald's, Credit Suisse, Adobe, major players. Tell me, what have you learned along the journey since founding the company ⁓ about the process of working with these large multinationals and and how they're using AI to further their business? speaker-2: Yes, that that actually the the the core of the problem often stays the same. It just represents in even bigger ways. And so when we started two thousand eight, you might remember financial crisis. And we started to work with all the large banks because they had to prove to the regulator that this number on this report, they could explain how that number was calculated. And that's a complex problem because it comes from hundreds of different systems, because it's combined, it's calculated to end up in that number. So traceability, lineage, quality, ownership. Very, very important. And so that's how we started working. It's a data problem, a data governance problem. Now, fast forward 15 years, I'm skipping lots of steps. But with AI, again, same problem. ⁓ AI says it does something, it says something, but why did it do that? Can we can we trust that it does the right thing? And so the same problems around trust, around traceability, around lineage, around quality kind of comes back over and over and over again. And so it's it's interestingly ⁓ the the use cases on top of data. just getting bigger and bigger and bigger and this is why it's still so exciting after eighteen years of of doing this, of ⁓ we're still kind of in the middle of I think where the industry is i is going and the the problems only get more exciting, I I would say. speaker-0: Felix, interesting question here. Two words that often come to the fore every time I've heard you or other experts in the industry speak about some of the things that you're doing. Two words ontology, semantics. What does that really mean in in simple terms? speaker-2: That's a that's a that's a complex question. ⁓ because I I I agree. If you ask ten people what they mean with ontology, you get ten different answers. I mean, Palantir obviously ⁓ educated the market a lot around kind of what they mean with an ontology. If you go back to kind of the formal academic definition, ⁓ people call it a a shared conceptualization of the world. Like how do we represent the world? Now, how you actually do that, there's lots of different levels of Formalism that you can add to that. It all all the ways to description logic, if if you will, of how to actually represent the world so you can reason on top of it. Now ultimately, I think in in the simple terms of how do we explain how not the world necessarily, but our company works? What is a customer? A customer buys products, a customer has orders, a customer pays ⁓ through invoices. How do we kind of represent that model of how our company works in a way that not just people can understand, but also machines can understand? And there's lots of different ways to do that. But that's really kind of a shared understanding of how our company works, is I would say the the easiest way to to think about it. speaker-1: I read an interview that you did where you kind of gave tips or said what you look for when you're potentially hiring new employees, how AI efficient they are. To our listeners that maybe are trying to stay on top of things, what would you say are like the most important tools that ⁓ the new job market seekers should be ⁓ available and aware of as they go in to take jobs in tech? speaker-2: It's an it's such an interesting time because the world's moving as quickly. Frankly, I I have a hard time keeping up. Like I I don't know if I'm actually the the right person to ask because we actually hire quite a lot of young people now out of school. And it's really incredible to see how native, right, ⁓ th they they work with this. And it's the first time, ⁓ and maybe it it says something about my age now, but the first time I'm I'm no longer the the the digital native. It's really like the AI native and I do see a a difference. So it's not easy to For all of us to adopt all these tools, there's a lot of change. There's so much tools. There's a lot of noise as well. You hear so many different things. Like, how do you how do you see the forest between the trees? It's really difficult right now. But ⁓ but we've had a lot of success hiring ⁓ kind of really young people that have like this this kind of completely open perspective, right? And they're almost not burdened sometimes with some of the experience that that we bring doing this for a long time. So it's almost if I can kind of for even for myself, learn from that. Like how do we s unlearn some things, right? I think that that's and that's not easy, but sometimes necessary to come as think first principles from scratch. You think something has to be done a certain way, but you almost have to unlearn some of that to think completely in a novel way and approach problems in a in a completely novel way. And it doesn't come that easy. And it's it's true. I think experimentation, testing, doing, which is hard because we all have day jobs and we all have almost so too much on our plate. So that that often comes kind of on top of everything. So it's it's it's ⁓ I I do feel and I and I hear with with our teams as well, ⁓ I think if you ask people, has AI made your job easier? Are you working less because of AI? I think very rarely are gonna say, are you gonna hear yes, right? Yes, it helps my jobs, but in some way we're all working, working more. And I think that's just a reflection of where we are in this adoption cycle. There's so much change, so much is new. And we have to all kind of get through that. And some are some are some of us are Getting through that more quickly than than others, but it's a it's not an easy process. speaker-0: One thing that I ⁓ certainly I've noticed is ⁓ almost every big company now suddenly suddenly has someone with AI in their title. It's it seems almost like everyone is rushing to have someone with that chief AI title. ⁓ why why is there such rush? ⁓ if there is a Russian date creating such roles. Yeah. speaker-2: I I think it's it's to be expected, right? I think it it's really important that companies, public companies want show that ⁓ they're kind of really doubling down on AI and doing that ⁓ appropriately. The other thing is also real because AI in an organization is not just about the technology. That's often the easiest part. It's really about change management. And having a clear leader and a real top down effort to to implement and adopt and change your organization often requires really a a top down ⁓ a top-down initiative and then hiring somebody or assigning somebody to take down that responsibility, I think is actually a good idea. ⁓ so I think this is why you probably see that happen. The one thing I will say and ⁓ it goes back to the discussion we've had, I really think data and AI are two sides of the same coin. Right. And I think the the what I would ⁓ how should I say, ⁓ drive careful that you don't make that two different teams. If you create too much distance between, let's say, the AI team and the data team, I think there's a real struggle there because you can't do AI without data. Frankly, you shouldn't do data without AI. So the the companies that that that I've seen are the most successful, the most mature, they have combined that chief data officer role, which has been existing for a long time, and a chief AI officer role and made it a chief data and AI officer role. And I think that's a great sign of a company that's approaching AI in the right way with the right level of maturity. speaker-1: Felix, last question. What what's in store for Colibra? Is there an acquisition in sight? What what's ⁓ coming up ⁓ in the horizon? speaker-2: It's it's an exciting time. There's a lot of innovation happening on the on the product and platform side. The world's moving really quickly. ⁓ we're always looking for great companies. We've done a number of acquisitions over the last years. ⁓ we'll continue to do so. It's it's always I love to bring on kind of founders in the in the company. I think it brings brings a lot of energy and entrepreneurship, which is so important, especially in this kind of AI world, which is so fast moving. So we'll continue to do so. ⁓ we'll continue to work very deeply with our partners and the ecosystem, which again is it's gonna new ecosystems with with new play. So it's just just an exciting so such an exciting time. And especially when you hear customers really doubling down and communicating how important having a strong data foundation is. It's just it's a it's a it's great to hear given that that's what we've been working on for such a long time. speaker-0: Felix, what a great conversation we've had so far. Thank you so much for making what would have otherwise been a very complicated conversation. Quite simple and easy to understand. Appreciate that. speaker-2: Absolutely happy to. speaker-1: And thank you to everyone listening. This is Machine Dreams. More honest, unfiltered conversations with the people shaping the future of AI are on the way. See you next time.