AI Agents Are Failing and It's Almost Never the Model's Fault | Alberto Pan, Denodo

0:00The adoption by enterprise of both generative AI and agentic AI has been much slower than a lot of people anticipated. They don't trust the AI. Most companies, most organizations, as you know, during the last two years, they have been doing pilots around AI, right? How long is it going to be before all large enterprises are using agentic AI in all relevant processes? I think it will take significant time because at the end of the day, there are bottlenecks that in many cases are related to organization or even legal reasons in some cases. Can you talk about the trust problem, what you guys, I think, call the trust gap?

0:43Why don't you start by introducing yourself, how you came to Donodo, what Donodo does? Yeah, absolutely. Well, first of all, thank you. Thank you for having me, Craig. So, yeah, I am Alberto Pan. I am chief technology officer of Denodo. I'm also a member of the founding team. Denodo is a global company today. We are headquartered in Palo Alto and we have offices in more than 25 countries. But our story actually began in Acorunia, which is a small city in the northwest of Spain. And actually, I'm still based there today. I am talking from Acorunia today. And here, basically, I'm leading our R &D team.

1:29The majority of our R &D team is based here. And about my journey, my journey actually started in academia. I was for many years doing research and data management. and actually the core technology that powers Denodo today grew out from that original research that me and other members of the founding team of Denodo were doing. And Denodo, describe the problem that you guys set out to solve and then the solution you've come up with. Denodo, first of all, is a data management company that enables organizations, typically big organizations, to create a unified, real-time access layer, data access layer across all their data sources.

2:20We typically call this a universal semantic layer because it provides the data in the language of the business, so it makes it easy for people and, of course, also today for AI agents not only to get access to the data that they need, but also to understand how this data should be used in different business contexts. And a big difference between Denodo and traditional data management architectures is that Denodo does not force you to consolidate everything up front in a central system, like in traditional data warehouse or layhouse architectures. With Denodo, you can query the data where it lives.

3:05And this has a number of benefits. It removes the bottlenecks that are typically associated to centralization, also allows accessing the data in real time where it lives, as I mentioned, and also allows users and agents to get access to all the data. Because typically in many organizations, what you have in the central systems and data warehouses or really houses is only a small percentage of the data. So with the nodal, you can actually get access to all the data. And we call this typically a logical data management because, well, you don't need total physical data replication, right? And the underlying technology that allows this is called in the market typically data virtualization.

3:56and the semantic layer that allows both humans and ai agents to make queries in natural language or or how does that work well um natural language is one of the of the interfaces supported and obviously we are using gen ai for that but actually you can also consume the data with more traditional interfaces. For instance, actually, what we try to do is to deliver the data so it can be consumed with any tool. So, for instance, if you are accessing the data products that you create with Denodo with tools like Power BI or Tableau, you will probably access them using technologies like JDBC or ODBC, right?

4:45Technologies oriented to SQL. But if you are a data scientist and you want to access Denodo data with a notebook, maybe you will use a technology like Arrow. Or if you are creating a web application on top of the data products exposed by Denodo, maybe you will use REST APIs, right? Or if you are an agent, you will probably use MCP. All those are examples of interfaces that are available and that you can use to query these data products that you create with Denodo across your data sources. But the semantic layer, in order for the interface to know what data you're trying to access in disparate locations, does that unify not only access, but the language that you need to access?

5:40I mean, I've spoken to a lot of people about data ontologies, and maybe one data store calls a customer a customer, and another data store calls it a client, you know, something like that. And you need to link the two. You mentioned graph databases, is that what you're saying? That problem of solving inconsistencies across data sources is very, very important in practice. and it's crucial in Denodo. And yes, Denodo provides ways to explicitly specify the meaning of each term and each piece of data in each data source and also even to consolidate different formats in a unified format, right? Right.

6:28So you can go either way. You can either expose in a consolidated data product the data in a single format Or you can also say, no, I have these different data products, and the way, for instance, that this KPI is computed in this data product is this one, and the way it is computed in this other data product is this one, right? So the agents or the users accessing those data products have the full context to understand, okay, for instance, what data product should I use in this particular business context? You describe it as a semantic layer. Is that software that sits in the cloud? Is it software that sits on-premise at the organization?

7:16Well, actually, both options are available. I would say that today, most of our customers use Denodo in the cloud. But actually, Denodo can be deployed at any location, on-premises, private cloud, public cloud. If you are running Denodo in the cloud, you can also choose between a SaaS model where basically we manage the infrastructure for you, or you can decide to manage the infrastructure yourself. so we provide a lot of deployment options. We typically work with very big organizations so I think for this type of organizations giving them flexibility in the type of deployment is very important, right?

7:59Because for instance, sometimes they are accessing very sensitive data so maybe they want that for this particular use case they want another to sit on-premises data because that data is not in the cloud. So we really try to give that flexibility to our customers. How then has evolved with agentic AI? Probably the main change between agentic AI and the initial iterations of gen AI applications is that, well, we have just mentioned a minute ago, right? If you think about it, the first gen AI applications were for data analytics. So they were about helping humans to get information, to make better decisions.

8:46And now we are seeing a new phase that maybe we could call them agentic AI or operational AI, where basically agents are being embedded directly into business workflows. I don't know, to manage, I don't know, customer incidents or insurance claims or logistics. So this is just, it's no longer only analyzing data. It's real-time decision-making. And I think that's where traditional data management architecture start to break down. Because if you think about it, traditional data management architectures, data warehouse, data warehouse, and so on, they were built for analytics. So while AI remained confined to analytics, I think this was mostly fine.

9:37Obviously, there are also limitations there, but it was mostly fine. But when we are starting to extend the scope, I think that changes things significantly. And that is having a big impact also, I think, in the nodo and in the perception of the nodo by the market. Adoption by enterprise of both generative AI and agentic AI has been much slower than a lot of people anticipated. and it's primarily a trust problem because enterprises are reluctant to integrate generative AI or agentic AI into their systems beyond maybe a customer-facing chatbot, but even that can be problematic. They don't trust the AI, And this is particularly a problem with generative AI, which is probabilistic.

10:39And, you know, older systems, certainly from the expert system era of AI, were deterministic. And a lot of enterprises feel much more comfortable knowing what the answer is going to be, or the answer is going to be deterministic. Can you talk about the trust problem, what you guys, I think, call the trust gap? And then we'll talk about the study that you just came out with. This study basically was a survey on more than 800, more than 850 data leaders in big organizations with representation of companies from the U.S., from EMEA, from APAC. And all main sectors, including financial services, the public sector, healthcare, you name it, focuses specifically on big organizations, organizations with more than 1 ,000 employees at least.

11:44And yeah, what we found is consistent across industries and also consistent with other industry surveys from the likes of McKinsey or Gartner. and it's that most failure modes of AI agents today are related to data. Most companies, most organizations, as you know, during the last two years, they have been doing pilots around AI, right? And now I would say in the last few months, they have been analyzing the results and I think most of them have reached this surprising conclusion that is typically, it's not that the model is not smart enough. It is that it does not have access to the right data, or maybe the data is not up to date, or the agent is not understanding how the data should be used in a particular business context, right?

12:37And that is what creates hallucinations. That is what it creates when the agents overreach what they should do. And this is basically what is creating this trust problem, right? We drill down in the report. There are some numbers that I would say, and we try to understand exactly what are the root causes of the problems. We get some striking numbers. For instance, almost 70 % say that the lack of real-time data for the agents has been a problem. And the percentage is even bigger if you take organizations that actually have AI in production. So that suggests that, you know, as organizations go into production, they realize that this is a real problem.

13:25Also, another problem that is mentioned in the study is that in these organizations, AI agents need data coming from an average of 400 data sources. And more than 85 % of organizations require more than 100 data sources. So these numbers are amazing. And almost 70 % also report problems, probably because of this data distribution, report problems getting agents to comply with consistent governance policies and consistent work rates. So I think that's the main conclusion that you can get from the report, that the reason why there is this trust gap, the reason why the agent sometimes overreach or providing consistent results are mostly data problems.

14:23Can you break some of those down? Live data, for example, is that a data architecture problem? I would say it's an architecture problem because, as you were mentioning before, initially AI was focused on analytics. So it was based in these data warehouses, data lay house systems that forced you to first copy all the data there. So when you have a data replication process, you always have some latency. So the data will never be 100 % real time. That's fine for analytics, typically, when you are trying to find trends or things like that. But it's not good enough for many business workflows, right?

15:10So that is why, let's say, the whole foundation for AI is not good enough to get real time data. As I said before, 70 % of the organization said that even data that is one minute stale is not valid for some AI agents. How does the Denodo semantic layer correct that or enhance that? Yeah, because with Denodo, you don't need to centralize the data up front. you don't need to replicate the data. The Nodo is able to query the data where it lives, right? So you can think about this data product that you create with the Nodo as virtual data product. So maybe one data product is created with your lighthouse and is pointing to the lighthouse, but maybe this other data product is pointing to your CRM or an operational database, right?

16:11Where data is updated live. So actually the data product you create with the Nodo can provide this data up to date in real time. Well, for instance, if you are using a centralized architecture, you first need to copy the data from that operational database to the central system. And that will always introduce some latency. So data will not be in real time. The right data is obvious that you need the right data for the right action or answer. But what's the issue there? What creates the trust gap in traditional architectures? There are several issues here, right? Because having the right data means different things.

16:55First, you need access to all the data. If you are confining your AI applications to a single system, for instance, to a data warehouse or a layhouse, then you will probably be missing part of the data that is needed. So that is one part of the problem. And the second part of the problem is these semantics that we were talking about, right? First, that the semantics need to be described. And second, the semantics need to be consistent. To explain the problem of semantics, I sometimes use an example, right? Imagine that you have a very brilliant DBA, a new hire with perfect SQL knowledge, but it's a new hire.

17:39So it lacks the business context, right? If you ask them for, I don't know, for instance, in the healthcare space, the readmission rate of cardiac patients, and then you give them access to the raw tables that contain the data to answer that, but you only give them that, they will fail. They will fail because they won't know, for instance, which specific diagnostic codes define all the different cardiac diseases. Or they will not know what is the exact formula for calculating a readmission in this case. For instance, if a patient is readmitted for a different diagnostic, is that a readmission or not, right?

18:20They might even have problems knowing what data is in each column table, right? So without a semantic layer to define these business rules and also in a consistent way across all data sources, even the smartest expert is forced to guess. And in the world of AI, guesses are basically hallucinations. So that's what we mean with the right data. First, having access to all the data. Second, the data with the right level of freshness. If you need this piece of data in real time, it should be in real time. And third, with consistent semantics. From the user's point of view, is it a dashboard or is this all happening unseen to the user?

19:08Denodo is more like this catalog of data products. It's like a web commerce, you know, like Amazon or a digital storefront where Denodo doesn't actually move the data to create these data products. It's like a catalog of these data products. These data products maybe come from a cloud database or on-premises system or a SaaS application, right? And then these data products can be consumed basically using any interface that you want, right? We actually also have our own, let's say, data marketplace, which is a web application that human users can use. But also, as I mentioned before, these data products can be accessed with any tool, Any tool that wants to consume that data using all those interfaces that I mentioned before.

19:58On these foundations of trust with the AI, we talked about live data, the right data, and then guardrails. Can you talk about the guardrails that are necessary, why they're necessary? Yes. Yes, absolutely. Well, I think, you know, there are some types of worries that everybody understands, everybody talks about. I don't know, things like ensuring that the prompts are safe, protecting yourself against prompt injection, forbidding the agents the access to, I don't know, to certain tools. Everybody understands that, right? But I think what sometimes goes under the radar is that, as I mentioned before, and the study shows, agents need to pull data from a massive variety of data sources.

20:55And when you have to work with multiple data sources, then you have several problems. For instance, the problems of the consistent semantics that we mentioned before, but also how do you enforce consistent security policies across all those data sources? So, for instance, imagine that you have a customer support agent, but because of GDPR or any other regulation or whatever, the agent is not allowed to see certain data from customers. And that data from customers will not be, in many cases, may be available in several data sources, in several places. So if you don't have something like Denodo and the agent needs to access those different data sources, you will need to implement those security rules in several places.

21:50With Denodo, you can implement that rule in only one place and you can be sure that it will be enforced for all the agents, for all the data sources, and so on. right also i would like to briefly mention or briefly comment why agents need to access to so many data sources right well for many reasons right first because organizations are very complex data is distributed in many places also because real-time data sometimes prevent consolidation as we mentioned because you need to query the data where it lives but and a third reason and And this is very fundamental about agents, is that agents are unpredictable by definition.

22:34So by definition, agents, we use agents when we don't have a predefined set of rules, where we have a task, when we have a task where a predefined set of rules does not cover all the scenarios, right? Because if we had a predefined set of rules that covers all scenarios, we could use a conventional workflow. So the whole point of an agent is that it needs to find solutions for problems that they have never seen before in exactly the same form. And if you think about it, this means that we don't know in advance exactly what data an agent will need. In a typical workflow, we know, okay, this workflow needs this piece of data, this piece of data, and we can expose only that piece of data.

23:21But with an agent, you have to give them access to all the data that potentially may be relevant for the task at hand. And that also contributes to increase a lot the number of data sources. You were talking about live data. Why is live data becoming critical for trustworthy AI? What breaks when AI acts on stale data? Basically, that leads to bad decisions or has the potential to cause the agents to make bad decisions. For instance, if your agent is talking with a customer about a service down incident, it needs to know the state of the service now, not five minutes ago. or if it is deciding how to locate a certain resource, for instance, a seat in a plane for an upgrade or something like that, right?

24:17In this information about the resource availability now, the resource availability 10 minutes ago, one hour ago may not be good enough. Not all workflows are like this, right? There will be agents that maybe will not need real-time data. For instance, agents that work only for analytics, as I mentioned before, sometimes don't need real-time data. But there are many, many workflows that actually need real-time data. In other case, the agents will not be able to make good decisions. The challenge of enforcing these guardrails in governments across distributed environments, how do you handle that?

25:01And how does agentic AI make that even harder? Since Denodo is this single entry point for agents to get the data, or these data products are the single entry point to get the data, you can be sure that those policies will be always enforced, right? Behind the scenes, maybe the data for that product is coming from several data sources, But the agent never interacts directly with the data sources. It always goes through this virtual data product offered by Denodo. And it is there where the security policies and the warrails are defined first and then enforced. And these problems that we're talking about underlying the trust gap that you guys see in enterprises regarding integration of AI?

25:56You know, the organizations were through several stages in the last two years, right? The first stage was, okay, this is very cool. Let's start to experiment. Let's start running pilots. then the second stage was you know wow this is this is the sometimes is you know this makes for impressive demos and sometimes it actually provides fantastic insights or fantastic suggestions but also very often makes these strange decisions or or or overreaches maybe the agent overreaches does things that it should not do. And then, wow. And initially, as you mentioned before, this is not an expert system. This is not symbolic, neurosymbolic.

26:41So at first, I cannot know what went wrong, right? So I think that that was the second stage. The third stage was, okay, we are investing a lot of money on this. Let's analyze what's going on. Let's analyze what's going on. So first stage is, okay, let's invest in auditability. Let's invest in traceability and auditability because actually those agents are, you know, using external tools and so on, you are quite able to know exactly, okay, I got this data, then I analyzed the data in function of this, I decided this and so on. So actually you can get good traceability if you're really investing it.

27:17And then analyzing these failure modes, it's when I think these organizations discover these gaps that are mostly data gaps, right? and I think that changed the strategy. For instance, I was seeing last week a very recent report from IDC, from the annual, and I was actually comparing the report from 2026 and the report from 2025. The report from 2026 was released, I think, in April, and the other one, I think, is February 2025, right? And the question was, what is the main barrier? No, what is your main priority for investment in AI? And in 2025, revamping your data architecture for AI was like in fifth place or something like that or sixth place.

28:06And now it's the first place, right? And I think that change is because of this realization that where the failure modes really are. That's interesting. So we're at a point in the adoption by enterprise where there was the pilot phase, then they took it a little further and ran into these problems, understood that it's data. And that's where Donoto comes in to unify the disparate data sources and then provide these unified universal policies. How should executives think about balancing existing lake house investments with the new layers that agentic AI requires? and what organizations, what should organizations do as they scale AI?

29:18Yeah. Well, I think that for most organizations, the lighthouse is part of the solution, right? It contains a lot of valuable data. It's a great foundation for data analytics and data science. But I also think it's not enough. even in the world before AI, there will always be data that is not in the lighthouse. You will sometimes need to access real-time data. You will need a strong semantics across several data sources. So even in the world before of AI, we think that layhouses were not the final answer. Very useful, but not the final answer. But with AI, all those problems are multiplied. And that is why we think that this new data foundation is needed.

30:07So our recommendation would be leverage the investments that you have, the data sources that you have, but prepare this common infrastructure, right? Data infrastructure that goes beyond central system that actually covers all your systems and ensure that your semantics, your governance and so on is defined there. And this is also very related to what organizations should avoid, what organizations should not do. Well, I see a lot of companies falling into what I call the ad hoc trap, meaning that, you know, since because these traditional architectures have limits for the new use cases, and there's so much pressure to deliver results, it's tempting to take shortcuts, right?

30:58is tending, for instance, to create an ad hoc data layer just for one specific AI app, and then another specific data layer only for this other AI app, right? Or you may say, okay, this agent will be a customer support agent, so I will restrict it to work only inside the CRM data, right? And that might give you a quick win, but we really think it's a dead end because we are rapidly moving towards a multi-agent world. And in this world, agents will not work in isolation. They will need to collaborate. And if every agent is built on a different data silo with its own definitions, they won't be able to talk to each other.

31:49They won't be able to cooperate. That's why, you know, those are the two traps that I see for organizations. The first thing is, okay, let's try to create AI on a single system. We will try to centralize all data here. That will never work for big organizations. And the second trap is, okay, since the first thing is not working, let's do a dog solutions for every agent. And that will not work in the long term either, because agents need to cooperate and talk to each other. And these other solutions, centralizing the data or building data layers for each AI or agentic application, are companies, enterprises, before they find Denodo, are they doing that themselves?

32:34Are they using consultants? Is that a solution that they come up with because it seems an obvious fix? Or are there people in the market telling them to do that, that that's the solution? A bit of everything, right? Obviously, on one hand, obviously, you know, its vendor tries to bring the data to their systems and to have their AI engines work on data, right? So obviously, the vendors try to do that. But if you think about it, where does that end? If all of them are partially successful, then at the end, what you have is many silos, right? With different AIs that cannot talk to each other. So part of it is the market, of course.