theCUBE + NYSE Wired: Data + AI: Turning Data Into Knowledge for Autonomous Systems | Felix van de Maele, Collibra | Neo4J GraphSummit NYC 2026

Felix van de Maele, Collibra | Neo4J GraphSummit NYC 2026

This interview examines Collibra product announcements and strategies to reduce the enterprise "hallucination tax" in artificial intelligence, AI. Felix van de Maele of Collibra is chief executive officer and co-founder. van de Maele outlines Collibra’s approach to accelerate production-grade AI while strengthening governance and compliance. The conversation focuses on AI governance, the hallucination tax, and new Collibra capabilities, including Live Map for context retrieval, Maestro for governance automation and Guardian Agents and Agent Contracts for runtime enforcement and policy-driven controls. van de Maele explains that Live Map reduces token consumption by roughly 60 percent while improving accuracy, they confirm that Maestro is generally available, and they describe Guardian Agents in public preview together with Agent Contracts to provide runtime controls that address legal risk and compliance requirements. Key takeaways emphasize the operational impact of the hallucination tax on enterprise AI adoption. van de Maele frames the hallucination tax as a barrier to production, citing survey evidence that many leaders re-verify agent outputs and that numerous projects stall in pilot phases. These insights underscore the importance of context platforms, knowledge graphs and robust data governance for reliable, production-ready AI. This discussion is relevant for leaders and practitioners concerned with AI governance, data governance, enterprise AI implementation, observability and telemetry. It highlights how context retrieval and governance automation reduce operational overhead and support compliance in regulated environments.

Share this session

Clips
More from theCUBE + NYSE Wired: Data + AI: Turning Data Into Knowledge for Autonomous Systems

Felix Van de Maele

Founder, CEO

Collibra

play_circle_outline The Hallucination Tax and Scaling AI: Trust, Reverification, Security, and Operational Resilience

play_circle_outline Live Map: Grounding Agents in Business Reality and Reducing Token Costs

play_circle_outline Maestro: governance automation enabling steward agents and faster policy enforcement

play_circle_outline Observability, lineage, and impact analysis for agent drift and root cause

play_circle_outline Regulatory, legal, and compliance challenges requiring runtime governance

Info

Felix van de Maele, Collibra | Neo4J GraphSummit NYC 2026

Felix Van de Maele

Founder, CEO Collibra

search

John Furrier

>> Palo Alto Studio Connection, Silicon Valley and Wall Street. I'm John Furrier, co-host here with Dave Vellante, my co-host. Well, I'm John Furrier. We are here in New York City with some breaking news here in the AI governance, agentic data layer. Breaking news around the hallucination tax in the enterprise AI world. A lot is going on with AI. Of course, hallucinations have been a big Discussion. Felix is back, CEO and co-founder of Collibra. We last talked in May. Welcome back. We got some breaking news. Yeah, thanks for having me, John. So you guys are launching new capabilities to reduce the hallucination tax on enterprise AI. That's the headline. Yep. I want to get into the 3 things that you're going to be talking about because I think they're relevant. Yes. But first, the hallucination tax. The Harris Poll came out. Explain. What they found.

Felix Van de Maele

>> Yeah. So what we found is that we asked over 300 business leaders and some interesting insights. One, over 3 out of 4 said that a lot of the AI initiatives have fallen short, which again, not a surprise. And also even a higher number have said that they really hit some roadblocks to get an AI use case from testing and development into production. And I would argue that's the biggest challenge every organization is facing. And so It's a broader challenge, but we call it the hallucination tax. It's really giving companies the confidence to move an AI use case into production. Do they have the right control to be able to do that constantly? And second is around, are they sure that the use case, the agent, is going to deliver the right level of accuracy to put that in production? And both are really interesting and big problems to solve.

John Furrier

>> Yeah, we talked last time in May when we were at the NYSE at theCUBE studio there, NYSE Wired program around ontologies, governance. This hallucination tax, the word tax is interesting because it's a tax on the company. Many, many taxable items on this one.

Felix Van de Maele

>> Yes.

John Furrier

>> we're seeing agents coming back with the same requests, different answers.

Felix Van de Maele

>> Yep.

John Furrier

>> Same prompts, different answers. You're seeing people rebuilding, stop-start. Yeah. There's a lot of operational costs. And also hallucination is also misinformation if it's not accurate.

Felix Van de Maele

>> Absolutely. There's a lot of talk around token economics, like how we use tokens, and that's absolutely important, actually linked to it. But I think an even bigger challenge is what if you don't trust the agent or the use case in the first place? And again, some of the data from the Harris Poll confirmed that we heard that over 90% of leaders feel like they have to re-verify and re-certify that they're providing the right context to the agent, right? And what's right today might not be right tomorrow. And that's an enormous, again, tax to maintain that. A lot of companies are using forward deployed engineers to do that. It's a big tax. How do we do it more in a systematized platform-way?

John Furrier

>> I saw the news last week when it came out on the Harris Poll, 90%, as you mentioned, are deploying autonomous agents. I wonder how many of those are authorized autonomous agents. We'll come back to that. 76 critical roadblocks.

Felix Van de Maele

>> Yes.

John Furrier

>> Hit roadblocks and barriers, keeping it in kind of pilot mode. That's a huge deal. Trust came up. And then of course, manual oversight kicks in. Yep. This is where it starts to break down. Production's not happening fast. The operational things aren't happening. What are you guys launching with your products? Because this ties into the news. Let's get into the news.

Felix Van de Maele

>> Exactly. These are the two big problems we are addressing. The other one is around the context, how do you make sure you're able to build context effectively and make sure that the agent is able to retrieve that effectively, continuously, right? And this is a new product we're calling Live Map, which is really the ability for agents to retrieve context that is grounded in their business reality. And that really builds upon our almost 20 years of experience in building ontologies, semantic context, or governance backbones. Another statistic was that I think over 80% said that this is really a data problem, having a strong data foundation is critical. So our new Live Map product gives the ability for agents to correctly retrieve the right context at the right time. And the benchmarks we're seeing are really, really great. There's over 60% less token consumption, which is real dollars and much higher accuracy as well. So I think that's one way that we're addressing that hallucination tax, the need for companies to no longer have to recertify or see impact to their operations because again, hallucination tax?

John Furrier

>> Better ops, more reliable agents. Yeah. What's the main

Felix Van de Maele

>> thing?

Felix Van de Maele

>> Two things. Less tokens, right? And better accuracy. And I'll give you an example, something that I'm sure is going to resonate with lots of people. You're building an agent. you have a lot of unstructured documents from your SharePoint. There's 100,000 documents there. You want to give it to the agent. The agent is smart enough. The model is smart enough to be able to parse through the data. But what you don't want is for the model to do that for every single question. Like, think of how many tokens you're wasting by doing that over and over and over again. And so building, engineering that context and making that available through Live Map, you hit it once offline, and so you have the answer ready every time. So that's a massive token efficiency.

John Furrier

>> So one of the other news pieces is that Maestro is generally available. Give some color on that.

Felix Van de Maele

>> Yeah, so we talked about governance being really, really important, governing the context, governing the AI, building on top of that data foundation, which only becomes more important. The challenge is how do you govern at the speed of AI, which is different than the way we had to govern 5 years ago, where it was regulatory compliance, it's just that the world is moving faster than ever. And so Maestro is really the ability to build agents that do the governance for you. Of course, it's human oversight, but Think about today, most of that work is done by data steward, context steward, and so on. A lot of manual work. Maestro gives our customers the ability to build these governance steward agents that actually automate up to 80% of the governance work. So you're able to govern quickly and efficiently enough to be able to keep up with the demands of new AI workloads.

John Furrier

>> Yeah, the no-code also comes in too. That's no-code, low-code was years ago. Now you got agents coding. Exactly. How does that fit in?

Felix Van de Maele

>> Yeah, so agents coding, it's really giving the right harness to the agents. And we've all read the cases where the agents escape. And so that's the other thing that we're launching that we're really excited about is we call it Guardian Agents. how do you make sure that the agents that are coding or doing other things stay within the bounds that you define? step one is how to actually define the bounds. And this is where we have a really interesting new concept that's resonating well is How do you define an Agent Contract? How do you define what the agent is allowed and not allowed to do? And that really drives a lot of policy management and things like that. So providing that context to say, okay, that agent can do these types of things and depending on the risk profile, the use case and so forth. And then how do you actually enforce that in production? Because that's one thing documenting it, which is great start.

John Furrier

>> Yeah.

Felix Van de Maele

>> But the thing is, how do you make sure the agent actually isn't allowed to do it, can't actually physically do it? And we've all read the examples where the agent escapes, right? So the way we're solving that is by encoding and compiling these Agent Contracts and deploying those into your existing AI gateways or AI orchestration tooling. And so when you do a call, before the agent actually is able to take the action, we monitor, does it align to the agent contract? If not, we block it and the agent never has the opportunity to actually take that action.

John Furrier

>> So the runtime control is the Guardian Agents and the—

Felix Van de Maele

>> The Live Map is the context. Think of governing the inputs as context governance with Live Map, governing the outputs as Guardian Agents.

John Furrier

>> Yeah. So you got the context and control.

Felix Van de Maele

>> Context and control.

John Furrier

>> And the Agent Contracts.

Felix Van de Maele

>> Yes.

John Furrier

>> And the Guardian Agents.

Felix Van de Maele

>> Yes.

John Furrier

>> We're in New York City. I can't help but reference the Guardian Angels, which ride the subway to protect everyone. Similar thing, guardrails, play on words, Guardian. That's really designed to be a guardian angel. For the context, make sure the control runtime works.

Felix Van de Maele

>> Absolutely. Runtime enforcement, which is absolutely critical.

John Furrier

>> Yeah. What's that? What does that address on that tax?

Felix Van de Maele

>> So it's a test of do we have enough confidence to put this AI use case in production? And we've seen it with a lot of customers. There's a lot of excitement, lots of prototyping works really well. And then think of the funnel of how many AI use cases actually make it in production. And it typically dies with legal and risk and compliance and all these challenges as well. How do we help the AI engineering teams make sure that the risk teams and the legal teams and the compliance teams feel comfortable to put these things in production?

John Furrier

>> It's interesting, one of the things I hear all the time on TheCUBE, I think we might address it briefly in May, was, about a year and a half ago, I just put an MCP server out there, we're good to go.

Felix Van de Maele

>> Yeah.

John Furrier

>> But not when you have third-party agents coming across domains.

Felix Van de Maele

>> Exactly.

John Furrier

>> So their guardrails or guardian agents will help. This is where the contracts I think come in. I want you to explain this because What I heard you say was it's almost like a contract with what you can do. Yep. Okay, my contract is I can hit this database. So it's really resource management. Is it like a resource management thing or?

Felix Van de Maele

>> It's a combination. Yeah, it's a great thing. It's like the way we approach it is first we build a fingerprint. Like what we do, for example, we scan GitHub because a lot of agents are just built in Python. They get pulled on a Git repository, nobody knows of them and they do whatever. How do you, one, discover what you actually have. And so we have a lot of scanning tools that discover that automatically, parse code and so on. And we're able to capture that agent's fingerprint. And then we tie that back to your policy. And then the policy might be, okay, if that agent is using unstructured data with highly sensitive information, which we know given the data background that we have, and we're using this externally, this is a high-risk agent. And so that means that legal needs to approve it before we can put it in production or it cannot create, it cannot access, it cannot have write access to certain applications. We've heard of a case where they built an agent and it had access to Workday, not just read access, write access, and they overwrote their whole employee database. You can imagine you don't want that. How do you specify what the agent is allowed to do and not allowed to do given the context? And that becomes the agent.

John Furrier

>> I mean, you're a predictive tool for resilience challenges because when you rewrite that database in Workday, you have to roll back. Yeah, it's a nightmare.

Felix Van de Maele

>> Yes.

John Furrier

>> Again, a big part of the stack.

Felix Van de Maele

>> Yes, exactly. How do you drive that resilience?

John Furrier

>> All right. So I want to get your thoughts on telemetry, observability, these discussions around explainability.

Felix Van de Maele

>> Yep.

John Furrier

>> Because agents need to track.

Felix Van de Maele

>> Yes.

John Furrier

>> Governance is a big part of that. How do you view that? How does that all fit into the announcements?

Felix Van de Maele

>> Yeah. So that's part of our AI Command Center suite. And I'd say observability applies at different levels. When I'm an engineer, I want deep technical observability. That's not what we do. There's great observability tools that actually do that with API calls and LLM calls and so forth. But you also want observability from a lineage and traceability, from an explainability perspective, right? What we're seeing is that agents work well day one, but they drift over time. What has changed that it worked yesterday, but it doesn't work today? And then you need to be able to trace back to, what data am I using? Have my source documents changed, did the quality change, is there a new system? So being able to trace back on a higher level is incredibly important. And also root cause analysis. What if something changes? Who is actually using that agent and what's the impact? And so being able to do impact analysis, because if I stop that agent, the Guardian Agent, that agent is no more, what's going to stop working? And so you need to do that on a much higher level. That lineage and impact analysis is one of the kind of observability.

John Furrier

>> This all comes back down to making it really easy to run agents, Live Map, You got the Guardian Agents and Agent Contracts there. You got Maestro to put it all together, orchestrate everything. Governance has taken on a real critical role. It's not an afterthought anymore. It's mainstream. It's in the real center of the value proposition. How do you view that? Because cross-domain and sovereignty comes up a lot. And I was talking to a friend the other day. He's like, well, in England, they drive on the left side of the road. In America, they drive on the right side of the road. You're now having different rules of engagement. You got to know this. Yes. How do you look at that? How should companies think about these policies? And that might be different based upon the environment it's in. Do I just turn on a different agent? So give me your thoughts and vision on this particular piece.

Felix Van de Maele

>> Yeah, look, I think we've been doing governance for 18 years. We've clearly cared about governance and thought it was incredibly important. But I think the stakes just have gotten so much higher, right? And the big difference today versus 2 or 3 years ago, it's not too long ago, is that we used to have people in the middle. We could count on the judgment of people that if something wasn't right, they were probably going to do the right thing. That has gone away. And so governance has moved from documentation, design time, policy setting to runtime because the agent is going to do whatever they can do. And we've seen it. It's going to goal drive to its goal, and they can go all the way to kind of lie and cheat, so to speak, to get to their goal. And so that is why—

John Furrier

>> or lazy agents, some people call them lazy agents.

Felix Van de Maele

>> Yeah, I'm not going to do the task. I'm going to say I've done the task, right? But so that's why governance has become absolutely critical to have that confidence. And it's gotten harder, right? Geopolitically, sovereignty, what models am I using? Open weights models, commercial models, who owns the IP? What do I do in my context? What are data retention policies? So it's incredibly complex. And so creating full visibility and almost control layer to do that at scale in large organizations that are subject to a lot of regulation. And I think the consensus is clearly when I talk to customers, there's going to be more regulation, not less. So we've got to be ready to do so. I think governance has just become mission critical.

John Furrier

>> I'm very bullish obviously on AI. I know you guys are too. But what's interesting from the Harris Poll findings is that 72% of tech decision makers feel AI initiatives are falling short. Does that surprise

Felix Van de Maele

>> you?

John Furrier

>> No.

Felix Van de Maele

>> Why?

Felix Van de Maele

>> It's hard. It's hard to truly drive ROI. And I think we've really seen a shift in the market as well, where there's a lot of hype, a lot of excitement, incredible promises. If you look at what's actually in production, I think there's been a lot of lessons learned, right? It's not that easy. And so I think it's also normal. I think I continue to be incredibly bullish. I think we're going to work all these things out. I think AI is going to have an incredible positive impact on everything that we do, but we're going to have to work through these problems. And it's normal. It's growing pains. It's learning lessons that we'll have to do. But this is why we have to be very intentional about how do we deal with these challenges.

John Furrier

>> I think the theoretical to actual practical shift that has happened, the rise of knowledge graphs have been big conversations we had again in May with ontologies is that, okay, every enterprise wants to have their data be a moat.

Felix Van de Maele

>> Yes.

John Furrier

>> And if you're going to have a data moat, you need to have that context, which is their domain data. It's not in the public domain. Some might, some might not be. When you start thinking about a brain, that's a graph.

Felix Van de Maele

>> Yeah.

John Furrier

>> How are graphs— why are graphs important? Why is there all this hubbub, which people like us have been in the data business love ontologies and love graphs, but why is it going mainstream?

Felix Van de Maele

>> Yeah, it's really interesting. I've used the word ontology more over the last 3 months than over the last 10 years, right? Same with the graph. And I think the reason is that there's a new platform emerging, which I call a context platform, right? And a big component of a context platform is context retrieval, right? And we've seen a number of different mechanisms to provide the best retrieval. Go back 2 years, everybody was just doing vector stores. And vector stores, RAG, that was the best thing. And it still is really good for certain use cases. For other use cases, I'm just gonna have Markdown files or YAML files and do grep and things like that. That works fine. For coding, typically that works great. I have semantic search, right, where I want to extract an entity. And semantic search becomes really a great mechanism. And of course, graph is a really important mechanism to ground the agent into the business context. And it's been very clear that we need graph approaches, typically under an ontology that kind of models your business domain to make sure all of that works. I believe it's going to be a combination of all of the above. There's not going to be one mechanism that is able to answer every potential question. The question is, how do we route, if you will? We have a lot of model routing. Let's think about context routing. How do we route the right context to the right retrieval mechanism at the right time? And Graph is key.

John Furrier

>> I remember when the cloud wave was hitting full tilt mode, it was like, okay, we went from one database rules the world to a multitude of databases. Similar here, Graph is just another layer in.

Felix Van de Maele

>> Exactly.

John Furrier

>> And this is why the context platform is emerging. It's not about ripping this out to put that in. It's really more of a compounding opportunity.

Felix Van de Maele

>> It's a combination, and it's the retrieval component, which is one component. We believe there's two other important components. One is context engineering— how to actually prepare all of my unstructured data typically to be the most effective potential context. And that's a level of modeling, extraction, filtering, and so on. That we've seen the same thing on analytics. You have analytics engineering— how do we prepare the structured data for analytics? How do we prepare our unstructured data for context? That's a key component. And then context governance. I think that's going to become a key thing because everybody today is just focused on their one agent. But as Harris Poll also said, companies don't have one agent, right? They have 10 agents, 100 agents, they're going to have 1,000 agents. Yeah, it's one thing solving it for one. How do you solve it for 1,000? How do you provide a level of consistency, standardization, reuse? Because otherwise it's just not going to work. And so the context governance again, it's gonna become.

John Furrier

>> And governance goes from being more or less of a nichey thing to more rules of the road kind of vibe.

Felix Van de Maele

>> Exactly, with agents being at runtime, there's no other option, there's no alternative.

John Furrier

>> All right, so let's get into some of the details. You got the three things, Maestro, Live Map, and Guardian Agents and Agent Contracts, three things. So the GA, you mentioned that, Live Map and Guardian Agents and Agent Contracts, Private preview? Is it available? Give us the status of the products.

Felix Van de Maele

>> Yeah, we're very excited. Maestro is GA, got great feedback from customers. We want to roll it out completely. Live Map is private preview. Again, working deeply with a lot of really early customers. Great feedback as well. More work to be done. So that's private preview. And then Guardian Agents is public preview available for everyone and will be GA shortly.

John Furrier

>> Got it. Okay, so what's next?

Felix Van de Maele

>> What's next?

John Furrier

>> Great milestone, pumping out some new products with velocity. You must be busy. What's going on?

Felix Van de Maele

>> It's a lot of depth as well, right? Making this work at enterprise scale with all these use cases. This is really what we're focused on. These are the three things we're working on: AI Command Center with Guardian Agents, Context Governance with the Ontology and Context Compiler, and then Context Engineering with Live Map. These are our three big focus areas. We want to make sure we're going to be the best in the world at doing that. And so that's our focus.

John Furrier

>> Felix, I was talking with Emil, the founder of Neo4j, at the NYSE, and there's kind of a trend emerging in the public domain, public sphere, a lot of noise. And so the general public thinks like, oh, people just woke up one day and said, hey, we're going to do AI. Oh my God, it's going to kill us. It's not safe. You guys are a great example, as is Neo4j and many others who have been doing work in this area. And it's a timing/market force. Speak to that because I think a lot of people don't know that you guys have been really doing a lot of good work here. Yeah, some of your best work. But now it's prime time. Same with Neo4j and same with a slew of others. Yeah. And then there's the newbies that are coming in.

Felix Van de Maele

>> It's prime time. And I think that depth of experience, having battle-tested our platforms, right, over the years at the most complex large organizations is important. Now we've all heard the vibe coding trend where, hey, we just use Claude Code and do this over the weekend. And I think with a lot of the startups and if you talk to these large enterprises, they can see it, right? There's a big difference between something you vibe coded, so to speak, over the weekend versus something that's been in production at scale, battle tested for years. You want that certainty nowadays, right? There's so much at stake. That foundation that you build, it's not— there's a lot of hard work and a lot of hard lessons to be learned. Yeah, I think know-how is more valuable.

John Furrier

>> Yeah, that's why I love the vibe coding trend, because it really is a promotional layer for absolutely no-code, low-code. Especially everyone knows, can relate to being a developer. Yep. Okay, but try to hook the database in, scale it out, you got to plug it into a bulletproof system with elasticity, security, governance.

Felix Van de Maele

>> Last 20% is probably the 99% now.

John Furrier

>> Yeah, anyone who's vibe coded me included, loves the first pass. And it's like, okay, I really want to turn it up a notch. I got to connect to an API. Oh shit, what's the key? oh yeah, it's a lot of stuff has to connect in. Exactly. And that's where the magic

Felix Van de Maele

>> is.

Felix Van de Maele

>> If you're a Fortune 100 company, Fortune 1000 company, especially now the level of security pressure that everybody's under, right? With these new security models. again, the stakes are higher than ever. Vibe coding doesn't

John Furrier

>> work.

>> Yeah. Felix, great to have you back on theCUBE. Congratulations on the news. We'll keep track of it and see how it goes.

Felix Van de Maele

>> Thank you, John.

John Furrier

>> Great to be here. I'm John Furrier with theCUBE. Breaking news, new products from Collibra. This is where the governance is going mainstream. It's key ingredient in the scale of AI applications. Of course, graphs are a big part of it. As the data becomes the moat for the enterprise, you got to nail all this stuff down. We're doing our part to keep up here on theCUBE. Thanks for watching.

Felix van de Maele, Collibra | Neo4J GraphSummit NYC 2026

search

John Furrier

>> Palo Alto Studio Connection, Silicon Valley and Wall Street. I'm John Furrier, co-host here with Dave Vellante, my co-host. Well, I'm John Furrier. We are here in New York City with some breaking news here in the AI governance, agentic data layer. Breaking news around the hallucination tax in the enterprise AI world. A lot is going on with AI. Of course, hallucinations have been a big Discussion. Felix is back, CEO and co-founder of Collibra. We last talked in May. Welcome back. We got some breaking news. Yeah, thanks for having me, John. So you guys are launching new capabilities to reduce the hallucination tax on enterprise AI. That's the headline. Yep. I want to get into the 3 things that you're going to be talking about because I think they're relevant. Yes. But first, the hallucination tax. The Harris Poll came out. Explain. What they found.

Felix Van de Maele

>> Yeah. So what we found is that we asked over 300 business leaders and some interesting insights. One, over 3 out of 4 said that a lot of the AI initiatives have fallen short, which again, not a surprise. And also even a higher number have said that they really hit some roadblocks to get an AI use case from testing and development into production. And I would argue that's the biggest challenge every organization is facing. And so It's a broader challenge, but we call it the hallucination tax. It's really giving companies the confidence to move an AI use case into production. Do they have the right control to be able to do that constantly? And second is around, are they sure that the use case, the agent, is going to deliver the right level of accuracy to put that in production? And both are really interesting and big problems to solve.

John Furrier

>> Yeah, we talked last time in May when we were at the NYSE at theCUBE studio there, NYSE Wired program around ontologies, governance. This hallucination tax, the word tax is interesting because it's a tax on the company. Many, many taxable items on this one.

Felix Van de Maele

>> Yes.

John Furrier

>> we're seeing agents coming back with the same requests, different answers.

Felix Van de Maele

>> Yep.

John Furrier

>> Same prompts, different answers. You're seeing people rebuilding, stop-start. Yeah. There's a lot of operational costs. And also hallucination is also misinformation if it's not accurate.

Felix Van de Maele

>> Absolutely. There's a lot of talk around token economics, like how we use tokens, and that's absolutely important, actually linked to it. But I think an even bigger challenge is what if you don't trust the agent or the use case in the first place? And again, some of the data from the Harris Poll confirmed that we heard that over 90% of leaders feel like they have to re-verify and re-certify that they're providing the right context to the agent, right? And what's right today might not be right tomorrow. And that's an enormous, again, tax to maintain that. A lot of companies are using forward deployed engineers to do that. It's a big tax. How do we do it more in a systematized platform-way?

John Furrier

>> I saw the news last week when it came out on the Harris Poll, 90%, as you mentioned, are deploying autonomous agents. I wonder how many of those are authorized autonomous agents. We'll come back to that. 76 critical roadblocks.

Felix Van de Maele

>> Yes.

John Furrier

>> Hit roadblocks and barriers, keeping it in kind of pilot mode. That's a huge deal. Trust came up. And then of course, manual oversight kicks in. Yep. This is where it starts to break down. Production's not happening fast. The operational things aren't happening. What are you guys launching with your products? Because this ties into the news. Let's get into the news.

Felix Van de Maele

>> Exactly. These are the two big problems we are addressing. The other one is around the context, how do you make sure you're able to build context effectively and make sure that the agent is able to retrieve that effectively, continuously, right? And this is a new product we're calling Live Map, which is really the ability for agents to retrieve context that is grounded in their business reality. And that really builds upon our almost 20 years of experience in building ontologies, semantic context, or governance backbones. Another statistic was that I think over 80% said that this is really a data problem, having a strong data foundation is critical. So our new Live Map product gives the ability for agents to correctly retrieve the right context at the right time. And the benchmarks we're seeing are really, really great. There's over 60% less token consumption, which is real dollars and much higher accuracy as well. So I think that's one way that we're addressing that hallucination tax, the need for companies to no longer have to recertify or see impact to their operations because again, hallucination tax?

John Furrier

>> Better ops, more reliable agents. Yeah. What's the main

Felix Van de Maele

>> thing?

Felix Van de Maele

>> Two things. Less tokens, right? And better accuracy. And I'll give you an example, something that I'm sure is going to resonate with lots of people. You're building an agent. you have a lot of unstructured documents from your SharePoint. There's 100,000 documents there. You want to give it to the agent. The agent is smart enough. The model is smart enough to be able to parse through the data. But what you don't want is for the model to do that for every single question. Like, think of how many tokens you're wasting by doing that over and over and over again. And so building, engineering that context and making that available through Live Map, you hit it once offline, and so you have the answer ready every time. So that's a massive token efficiency.

John Furrier

>> So one of the other news pieces is that Maestro is generally available. Give some color on that.

Felix Van de Maele

>> Yeah, so we talked about governance being really, really important, governing the context, governing the AI, building on top of that data foundation, which only becomes more important. The challenge is how do you govern at the speed of AI, which is different than the way we had to govern 5 years ago, where it was regulatory compliance, it's just that the world is moving faster than ever. And so Maestro is really the ability to build agents that do the governance for you. Of course, it's human oversight, but Think about today, most of that work is done by data steward, context steward, and so on. A lot of manual work. Maestro gives our customers the ability to build these governance steward agents that actually automate up to 80% of the governance work. So you're able to govern quickly and efficiently enough to be able to keep up with the demands of new AI workloads.

John Furrier

>> Yeah, the no-code also comes in too. That's no-code, low-code was years ago. Now you got agents coding. Exactly. How does that fit in?

Felix Van de Maele

>> Yeah, so agents coding, it's really giving the right harness to the agents. And we've all read the cases where the agents escape. And so that's the other thing that we're launching that we're really excited about is we call it Guardian Agents. how do you make sure that the agents that are coding or doing other things stay within the bounds that you define? step one is how to actually define the bounds. And this is where we have a really interesting new concept that's resonating well is How do you define an Agent Contract? How do you define what the agent is allowed and not allowed to do? And that really drives a lot of policy management and things like that. So providing that context to say, okay, that agent can do these types of things and depending on the risk profile, the use case and so forth. And then how do you actually enforce that in production? Because that's one thing documenting it, which is great start.

John Furrier

>> Yeah.

Felix Van de Maele

>> But the thing is, how do you make sure the agent actually isn't allowed to do it, can't actually physically do it? And we've all read the examples where the agent escapes, right? So the way we're solving that is by encoding and compiling these Agent Contracts and deploying those into your existing AI gateways or AI orchestration tooling. And so when you do a call, before the agent actually is able to take the action, we monitor, does it align to the agent contract? If not, we block it and the agent never has the opportunity to actually take that action.

John Furrier

>> So the runtime control is the Guardian Agents and the—

Felix Van de Maele

>> The Live Map is the context. Think of governing the inputs as context governance with Live Map, governing the outputs as Guardian Agents.

John Furrier

>> Yeah. So you got the context and control.

Felix Van de Maele

>> Context and control.

John Furrier

>> And the Agent Contracts.

Felix Van de Maele

>> Yes.

John Furrier

>> And the Guardian Agents.

Felix Van de Maele

>> Yes.

John Furrier

>> We're in New York City. I can't help but reference the Guardian Angels, which ride the subway to protect everyone. Similar thing, guardrails, play on words, Guardian. That's really designed to be a guardian angel. For the context, make sure the control runtime works.

Felix Van de Maele

>> Absolutely. Runtime enforcement, which is absolutely critical.

John Furrier

>> Yeah. What's that? What does that address on that tax?

Felix Van de Maele

>> So it's a test of do we have enough confidence to put this AI use case in production? And we've seen it with a lot of customers. There's a lot of excitement, lots of prototyping works really well. And then think of the funnel of how many AI use cases actually make it in production. And it typically dies with legal and risk and compliance and all these challenges as well. How do we help the AI engineering teams make sure that the risk teams and the legal teams and the compliance teams feel comfortable to put these things in production?

John Furrier

>> It's interesting, one of the things I hear all the time on TheCUBE, I think we might address it briefly in May, was, about a year and a half ago, I just put an MCP server out there, we're good to go.

Felix Van de Maele

>> Yeah.

John Furrier

>> But not when you have third-party agents coming across domains.

Felix Van de Maele

>> Exactly.

John Furrier

>> So their guardrails or guardian agents will help. This is where the contracts I think come in. I want you to explain this because What I heard you say was it's almost like a contract with what you can do. Yep. Okay, my contract is I can hit this database. So it's really resource management. Is it like a resource management thing or?

Felix Van de Maele

>> It's a combination. Yeah, it's a great thing. It's like the way we approach it is first we build a fingerprint. Like what we do, for example, we scan GitHub because a lot of agents are just built in Python. They get pulled on a Git repository, nobody knows of them and they do whatever. How do you, one, discover what you actually have. And so we have a lot of scanning tools that discover that automatically, parse code and so on. And we're able to capture that agent's fingerprint. And then we tie that back to your policy. And then the policy might be, okay, if that agent is using unstructured data with highly sensitive information, which we know given the data background that we have, and we're using this externally, this is a high-risk agent. And so that means that legal needs to approve it before we can put it in production or it cannot create, it cannot access, it cannot have write access to certain applications. We've heard of a case where they built an agent and it had access to Workday, not just read access, write access, and they overwrote their whole employee database. You can imagine you don't want that. How do you specify what the agent is allowed to do and not allowed to do given the context? And that becomes the agent.

John Furrier

>> I mean, you're a predictive tool for resilience challenges because when you rewrite that database in Workday, you have to roll back. Yeah, it's a nightmare.

Felix Van de Maele

>> Yes.

John Furrier

>> Again, a big part of the stack.

Felix Van de Maele

>> Yes, exactly. How do you drive that resilience?

John Furrier

>> All right. So I want to get your thoughts on telemetry, observability, these discussions around explainability.

Felix Van de Maele

>> Yep.

John Furrier

>> Because agents need to track.

Felix Van de Maele

>> Yes.

John Furrier

>> Governance is a big part of that. How do you view that? How does that all fit into the announcements?

Felix Van de Maele

>> Yeah. So that's part of our AI Command Center suite. And I'd say observability applies at different levels. When I'm an engineer, I want deep technical observability. That's not what we do. There's great observability tools that actually do that with API calls and LLM calls and so forth. But you also want observability from a lineage and traceability, from an explainability perspective, right? What we're seeing is that agents work well day one, but they drift over time. What has changed that it worked yesterday, but it doesn't work today? And then you need to be able to trace back to, what data am I using? Have my source documents changed, did the quality change, is there a new system? So being able to trace back on a higher level is incredibly important. And also root cause analysis. What if something changes? Who is actually using that agent and what's the impact? And so being able to do impact analysis, because if I stop that agent, the Guardian Agent, that agent is no more, what's going to stop working? And so you need to do that on a much higher level. That lineage and impact analysis is one of the kind of observability.

John Furrier

>> This all comes back down to making it really easy to run agents, Live Map, You got the Guardian Agents and Agent Contracts there. You got Maestro to put it all together, orchestrate everything. Governance has taken on a real critical role. It's not an afterthought anymore. It's mainstream. It's in the real center of the value proposition. How do you view that? Because cross-domain and sovereignty comes up a lot. And I was talking to a friend the other day. He's like, well, in England, they drive on the left side of the road. In America, they drive on the right side of the road. You're now having different rules of engagement. You got to know this. Yes. How do you look at that? How should companies think about these policies? And that might be different based upon the environment it's in. Do I just turn on a different agent? So give me your thoughts and vision on this particular piece.

Felix Van de Maele

>> Yeah, look, I think we've been doing governance for 18 years. We've clearly cared about governance and thought it was incredibly important. But I think the stakes just have gotten so much higher, right? And the big difference today versus 2 or 3 years ago, it's not too long ago, is that we used to have people in the middle. We could count on the judgment of people that if something wasn't right, they were probably going to do the right thing. That has gone away. And so governance has moved from documentation, design time, policy setting to runtime because the agent is going to do whatever they can do. And we've seen it. It's going to goal drive to its goal, and they can go all the way to kind of lie and cheat, so to speak, to get to their goal. And so that is why—

John Furrier

>> or lazy agents, some people call them lazy agents.

Felix Van de Maele

>> Yeah, I'm not going to do the task. I'm going to say I've done the task, right? But so that's why governance has become absolutely critical to have that confidence. And it's gotten harder, right? Geopolitically, sovereignty, what models am I using? Open weights models, commercial models, who owns the IP? What do I do in my context? What are data retention policies? So it's incredibly complex. And so creating full visibility and almost control layer to do that at scale in large organizations that are subject to a lot of regulation. And I think the consensus is clearly when I talk to customers, there's going to be more regulation, not less. So we've got to be ready to do so. I think governance has just become mission critical.

John Furrier

>> I'm very bullish obviously on AI. I know you guys are too. But what's interesting from the Harris Poll findings is that 72% of tech decision makers feel AI initiatives are falling short. Does that surprise

Felix Van de Maele

>> you?

John Furrier

>> No.

Felix Van de Maele

>> Why?

Felix Van de Maele

>> It's hard. It's hard to truly drive ROI. And I think we've really seen a shift in the market as well, where there's a lot of hype, a lot of excitement, incredible promises. If you look at what's actually in production, I think there's been a lot of lessons learned, right? It's not that easy. And so I think it's also normal. I think I continue to be incredibly bullish. I think we're going to work all these things out. I think AI is going to have an incredible positive impact on everything that we do, but we're going to have to work through these problems. And it's normal. It's growing pains. It's learning lessons that we'll have to do. But this is why we have to be very intentional about how do we deal with these challenges.

John Furrier

>> I think the theoretical to actual practical shift that has happened, the rise of knowledge graphs have been big conversations we had again in May with ontologies is that, okay, every enterprise wants to have their data be a moat.

Felix Van de Maele

>> Yes.

John Furrier

>> And if you're going to have a data moat, you need to have that context, which is their domain data. It's not in the public domain. Some might, some might not be. When you start thinking about a brain, that's a graph.

Felix Van de Maele

>> Yeah.

John Furrier

>> How are graphs— why are graphs important? Why is there all this hubbub, which people like us have been in the data business love ontologies and love graphs, but why is it going mainstream?

Felix Van de Maele

>> Yeah, it's really interesting. I've used the word ontology more over the last 3 months than over the last 10 years, right? Same with the graph. And I think the reason is that there's a new platform emerging, which I call a context platform, right? And a big component of a context platform is context retrieval, right? And we've seen a number of different mechanisms to provide the best retrieval. Go back 2 years, everybody was just doing vector stores. And vector stores, RAG, that was the best thing. And it still is really good for certain use cases. For other use cases, I'm just gonna have Markdown files or YAML files and do grep and things like that. That works fine. For coding, typically that works great. I have semantic search, right, where I want to extract an entity. And semantic search becomes really a great mechanism. And of course, graph is a really important mechanism to ground the agent into the business context. And it's been very clear that we need graph approaches, typically under an ontology that kind of models your business domain to make sure all of that works. I believe it's going to be a combination of all of the above. There's not going to be one mechanism that is able to answer every potential question. The question is, how do we route, if you will? We have a lot of model routing. Let's think about context routing. How do we route the right context to the right retrieval mechanism at the right time? And Graph is key.

John Furrier

>> I remember when the cloud wave was hitting full tilt mode, it was like, okay, we went from one database rules the world to a multitude of databases. Similar here, Graph is just another layer in.

Felix Van de Maele

>> Exactly.

John Furrier

>> And this is why the context platform is emerging. It's not about ripping this out to put that in. It's really more of a compounding opportunity.

Felix Van de Maele

>> It's a combination, and it's the retrieval component, which is one component. We believe there's two other important components. One is context engineering— how to actually prepare all of my unstructured data typically to be the most effective potential context. And that's a level of modeling, extraction, filtering, and so on. That we've seen the same thing on analytics. You have analytics engineering— how do we prepare the structured data for analytics? How do we prepare our unstructured data for context? That's a key component. And then context governance. I think that's going to become a key thing because everybody today is just focused on their one agent. But as Harris Poll also said, companies don't have one agent, right? They have 10 agents, 100 agents, they're going to have 1,000 agents. Yeah, it's one thing solving it for one. How do you solve it for 1,000? How do you provide a level of consistency, standardization, reuse? Because otherwise it's just not going to work. And so the context governance again, it's gonna become.

John Furrier

>> And governance goes from being more or less of a nichey thing to more rules of the road kind of vibe.

Felix Van de Maele

>> Exactly, with agents being at runtime, there's no other option, there's no alternative.

John Furrier

>> All right, so let's get into some of the details. You got the three things, Maestro, Live Map, and Guardian Agents and Agent Contracts, three things. So the GA, you mentioned that, Live Map and Guardian Agents and Agent Contracts, Private preview? Is it available? Give us the status of the products.

Felix Van de Maele

>> Yeah, we're very excited. Maestro is GA, got great feedback from customers. We want to roll it out completely. Live Map is private preview. Again, working deeply with a lot of really early customers. Great feedback as well. More work to be done. So that's private preview. And then Guardian Agents is public preview available for everyone and will be GA shortly.

John Furrier

>> Got it. Okay, so what's next?

Felix Van de Maele

>> What's next?

John Furrier

>> Great milestone, pumping out some new products with velocity. You must be busy. What's going on?

Felix Van de Maele

>> It's a lot of depth as well, right? Making this work at enterprise scale with all these use cases. This is really what we're focused on. These are the three things we're working on: AI Command Center with Guardian Agents, Context Governance with the Ontology and Context Compiler, and then Context Engineering with Live Map. These are our three big focus areas. We want to make sure we're going to be the best in the world at doing that. And so that's our focus.

John Furrier

>> Felix, I was talking with Emil, the founder of Neo4j, at the NYSE, and there's kind of a trend emerging in the public domain, public sphere, a lot of noise. And so the general public thinks like, oh, people just woke up one day and said, hey, we're going to do AI. Oh my God, it's going to kill us. It's not safe. You guys are a great example, as is Neo4j and many others who have been doing work in this area. And it's a timing/market force. Speak to that because I think a lot of people don't know that you guys have been really doing a lot of good work here. Yeah, some of your best work. But now it's prime time. Same with Neo4j and same with a slew of others. Yeah. And then there's the newbies that are coming in.

Felix Van de Maele

>> It's prime time. And I think that depth of experience, having battle-tested our platforms, right, over the years at the most complex large organizations is important. Now we've all heard the vibe coding trend where, hey, we just use Claude Code and do this over the weekend. And I think with a lot of the startups and if you talk to these large enterprises, they can see it, right? There's a big difference between something you vibe coded, so to speak, over the weekend versus something that's been in production at scale, battle tested for years. You want that certainty nowadays, right? There's so much at stake. That foundation that you build, it's not— there's a lot of hard work and a lot of hard lessons to be learned. Yeah, I think know-how is more valuable.

John Furrier

>> Yeah, that's why I love the vibe coding trend, because it really is a promotional layer for absolutely no-code, low-code. Especially everyone knows, can relate to being a developer. Yep. Okay, but try to hook the database in, scale it out, you got to plug it into a bulletproof system with elasticity, security, governance.

Felix Van de Maele

>> Last 20% is probably the 99% now.

John Furrier

>> Yeah, anyone who's vibe coded me included, loves the first pass. And it's like, okay, I really want to turn it up a notch. I got to connect to an API. Oh shit, what's the key? oh yeah, it's a lot of stuff has to connect in. Exactly. And that's where the magic

Felix Van de Maele

>> is.

Felix Van de Maele

>> If you're a Fortune 100 company, Fortune 1000 company, especially now the level of security pressure that everybody's under, right? With these new security models. again, the stakes are higher than ever. Vibe coding doesn't

John Furrier

>> work.

>> Yeah. Felix, great to have you back on theCUBE. Congratulations on the news. We'll keep track of it and see how it goes.

Felix Van de Maele

>> Thank you, John.

John Furrier

>> Great to be here. I'm John Furrier with theCUBE. Breaking news, new products from Collibra. This is where the governance is going mainstream. It's key ingredient in the scale of AI applications. Of course, graphs are a big part of it. As the data becomes the moat for the enterprise, you got to nail all this stuff down. We're doing our part to keep up here on theCUBE. Thanks for watching.