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

0:00The adoption by enterprise of both

0:03generative AI and agentic AI has been

0:06much slower than a lot of people

0:09anticipated. They don't trust the AI.

0:12>> Most companies, most organizations, as

0:14you know, during the last 2 years, they

0:16have been doing pilots around the AI,

0:18right?

0:18>> How long is it going to be before all

0:21large enterprises are using agentic AI

0:25in all relevant processes?

0:27>> I think it will take significant time

0:29because at the end of the day, there are

0:30bottlenecks that in many cases are

0:32related to organization or even legal

0:34resources in some cases.

0:35>> Can you talk about the trust problem,

0:39what you guys have been called a trust

0:41gap?

0:44Why don't you start by introducing

0:46yourself, how you came to Denodo, what

0:49Denodo does?

0:51>> Yeah, absolutely. Well, first of all,

0:53thank you. Thank you for having me,

0:54Craig.

0:55So, yeah, I am Alberto Pan. I am chief

0:57technology officer of Denodo. I'm also a

1:00member of the founding team.

1:03And Denodo is a is a global company

1:05today. We are headquartered in Palo Alto

1:08and we have

1:09offices in more than 25 countries.

1:12But our story actually began in A

1:14Coruña, which is a a small city in the

1:16in the northwest of Spain.

1:18And actually, I am still based there

1:20today. I I am talking from

1:23from A Coruña today and and here,

1:25basically, I'm leading

1:27uh our R&D team. The majority of our R&D

1:30team is is based here.

1:33Um about my journey, my journey actually

1:35started in academia. I was for

1:38many years.

1:40Uh I was for many years doing research

1:42on data management. And actually, the

1:45core technology that powers Denodo today

1:48uh grew out from that original research

1:51that me and and other members of the

1:53founding team of Denodo were doing.

1:55>> And Denodo, uh describe the the problem

1:58that you guys set out to solve and then

2:00the solution you've come up with.

2:03>> Denodo, first of all, is a data

2:04management company um

2:06that uh enables organizations, typically

2:10big organizations, to create a unified

2:13real-time

2:15uh access layer, data access layer

2:18across all their data sources, right?

2:20We typically call this a universal

2:23semantic layer

2:25because it provides the data in the

2:28language of the business, so it makes it

2:31easy for people and, of course, also

2:33today for AI agents not only to get

2:36access to the data that they need, but

2:38also to understand how this data should

2:41be used in different business contexts,

2:43right?

2:45And and and a big difference between

2:47Denodo and traditional

2:49uh data management architectures is that

2:51Denodo does not force you to consolidate

2:54everything up front uh

2:57in a central system, like like in

2:59traditional data warehouse or lakehouse

3:01architectures. With Denodo, you can

3:03query the data where it lives.

3:05Uh and this has a number of benefits. Uh

3:07there's it removes the the bottlenecks

3:10that are typically associated to

3:12centralization.

3:14Also allows accessing the data in real

3:16time where it lives, as I mentioned.

3:18And also allows users and agents to get

3:21access to all the data um because

3:24typically in in many organizations, what

3:26you have in the central systems and data

3:28warehouses or lakehouses is only a small

3:29percentage of the data. So, with Denodo,

3:32you can actually get access to

3:34to all the data.

3:35Uh and we call this uh typically um

3:39a logical data management

3:42because well, you don't need

3:44total physical data replication, right?

3:47And the and the underlying technology

3:49that

3:50uh allows this is is is called in the

3:52market typically uh data virtualization.

3:57>> And does semantic layer that allows

4:01both humans and AI agents

4:05to make queries in natural language or

4:09how does that work?

4:11>> Well, um

4:14natural language is one of the of the

4:15interfaces supported and obviously we

4:18are using Genie I for that. But actually

4:21you can also consume the data with more

4:25traditional interfaces. For instance,

4:27actually what we try to do is to deliver

4:30the data so it can be consumed with any

4:32tool. So for instance, if you are

4:34accessing

4:35um the data products as you create with

4:37the NODO with tools like Power BI or

4:40Tableau, you you will probably access

4:42them using technologies like JDBC or

4:44ODBC, right? Technologies oriented to to

4:46SQL. But if you are a data scientist and

4:49you want to access the NODO data with a

4:52notebook, maybe you will use a

4:53technology like Arrow. Or if you are

4:55creating a web application on top of the

4:59data products exposed by the NODO, maybe

5:01you will use REST APIs, right? Or if you

5:04are an AI agent, you will probably use

5:06MCP. All those are examples of

5:08interfaces that are available and that

5:10you can use to to query these

5:13data products

5:15uh um that you create with the NODO

5:17across your data sources.

5:19>> But but the semantic layer in order for

5:24uh

5:24the interface to know what data uh

5:28you're trying to access in uh disparate

5:31locations, does that unify not only

5:36access but the language that you need to

5:40access? I mean, I've spoken to a lot of

5:42people about data ontologies and maybe

5:45one data store calls a customer a

5:48customer and another data store calls

5:51uh

5:52it a client.

5:53You know, something like that. And you

5:55need to link the two. Uh uh you

5:58mentioned graph databases. How were you

6:01saying

6:01>> That problem of solving inconsistencies

6:04across data sources is very very

6:05important in practice.

6:07And And yet, it's crucial in Denodo. And

6:10yes, Denodo provides ways to

6:13explicitly specify the meaning of each

6:16term and each piece of data in each data

6:19source. And also, even to consolidate

6:23uh different um different formats in a

6:26unified format, right?

6:28>> Right.

6:28>> So, you you can go either way. Either

6:31you can either expose in a consolidated

6:34data product the data in a single

6:36format, or you can also say, "No, I have

6:38these different data products. And the

6:41way, for instance, that this KPI

6:43is computed in this data product is this

6:45one, and the way this is computed in

6:47this other data product is this one,

6:48right?" So, the agents or the users

6:51accessing those data products have the

6:53full context to understand, okay, for

6:55instance, what data product should I use

6:57in this particular business context?

6:59>> You describe it as a semantic layer.

7:03Uh

7:03it is Is that uh

7:07um

7:08software that sits in the cloud? Is it

7:11software that sits

7:13on premise at the organization?

7:17>> Well, actually, both options are

7:19available. I would I would say that

7:21today

7:23most of our customers use Denodo in the

7:25cloud.

7:26But, actually,

7:28uh

7:29Denodo can be deployed at any location,

7:31on premises, private cloud,

7:33uh public public cloud. If you're

7:35running Denodo in the cloud, you can

7:37also choose between a SaaS model where

7:40basically we manage the infrastructure

7:41for you, or you can decide to manage the

7:43infrastructure yourself. So, we provide

7:46a lot of uh deployment options. We

7:48typically work with

7:50very big organizations. So, I think for

7:53this type of of organizations, giving

7:55them

7:56the flexibility in the type of

7:57deployment is very important, right?

7:59Because for instance, sometimes they are

8:00accessing very sensitive data. So, maybe

8:03they want that for this particular use

8:06case, they want the node to sit on

8:07premises data, because that data is not

8:09in the cloud. So, we we try we really

8:11try to give that flexibility to to our

8:13customers.

8:14>> How then has evolved with uh with

8:17Agentic AI?

8:19>> Probably the main change between Agentic

8:21AI and the

8:23and the initial iterations of the Gen AI

8:25applications

8:27is that well, we have just mentioned

8:29that a minute ago, right?

8:31If you think about it,

8:33the

8:34the first thing AI applications were for

8:37data analytics.

8:38So, they were about helping humans

8:41uh to get information to make better

8:43decisions.

8:45Um

8:47And now we are seeing a new phase that

8:50maybe we could call them Agentic AI or

8:53operational AI, where basically agents

8:55are being embedded directly into the

8:59into business workflows.

9:01Um I don't know, to manage I don't know,

9:03customer incidents or insurance claims

9:05or logistics.

9:07Um so, this is just it's no longer only

9:11analyzing data.

9:13Uh

9:13it's real-time decision-making.

9:16And I think that's where traditional

9:18data management architectures start to

9:20break down, because if you think about

9:22it, traditional data management

9:24architectures, data warehouse, data

9:26lakehouse, and so on, they were built

9:28for analytics.

9:29So, while AI remained

9:32confined to analytics, I think this was

9:36mostly fine. Obviously, there are also

9:37limitations there, but it mostly fine.

9:40But when we are starting to extend in

9:42the scope, I think that's that changes

9:44things significantly and and that is

9:46having a big impact also in the I think

9:48in the novel one and in the perception

9:51of the novel by the market, yeah.

9:53>> Adoption by enterprise of [snorts] both

9:56generative AI and agentic AI has been

10:00much slower than a lot of people

10:03anticipated

10:05uh and it's primarily a trust problem

10:09uh because

10:12enterprises are reluctant to integrate

10:15generative AI or

10:18agentic AI into their

10:20uh systems

10:22uh beyond maybe a customer-facing

10:24chatbot, but even that is can be

10:27problematic. They don't trust

10:30uh the AI and this is particularly a

10:33problem with

10:35generative AI, which is

10:37probabilistic and uh you know, older

10:41uh systems certainly uh from the expert

10:45system era of AI

10:47uh were deterministic and a lot of

10:50enterprises feel much more

10:52comfortable knowing

10:55uh what the answer is going to be

10:58or the the answer is going to be

10:59deterministic. Can you talk about the

11:02trust problem uh what you guys I think

11:05call the trust gap and then we'll talk

11:08about the study that you you just came

11:11out with?

11:11>> This study basically

11:13uh was was a survey on more than 800

11:17more than 850

11:19um data leaders in big organizations

11:24um with representation of companies from

11:27from the US, from EMEA, from APAC

11:31and all main sectors including financial

11:34services, the public sector, health

11:36care,

11:37you name it.

11:38Uh

11:39focus specifically on big organizations,

11:41organizations with more than 1,000

11:42employees at least.

11:44And yet, what we found is consistent

11:47across industries and also consistent

11:49with other industry surveys from the

11:52likes of McKinsey or Gartner.

11:54And it's that

11:56um

11:57most failure modes of AI agents today

12:01uh

12:02are related to data. Most companies,

12:04most organizations, as you know, during

12:06the last 2 years, they have been doing

12:08pilots around AI, right?

12:10And now, I I would say in the last few

12:12months, they have been analyzing the

12:13results. And I think most of them have

12:17reached this

12:19surprising surprising conclusion

12:21that is

12:23typically, it's not that the model is

12:24not smart enough, it is that it does not

12:27have access to the right data, or maybe

12:30the data is not up to date,

12:32or the agent is not understanding how

12:34the data should be used in a particular

12:36business context, right? And that that

12:38that is what creates hallucinations,

12:41that is what it creates when the agents

12:43overreach what they should do, and this

12:45is basically what is creating this

12:48distrust problem, right? If you read

12:50down in the in the report,

12:53there are some numbers that I would say

12:54and we try to understand exactly what

12:57are

12:58the the root

12:59uh causes of the problems, we get some

13:01striking numbers.

13:03For instance, uh

13:05almost 70% say that

13:08uh the lack of real-time data for the

13:10agents has been a problem.

13:13And the percentage is even bigger if you

13:15take organizations that actually have AI

13:18in production. So, that suggests that,

13:19you know, as organizations go into

13:21production, they realize that this is a

13:23a real problem.

13:25Also, another problem that is mentioned

13:27in the study um is that

13:30uh in these organizations

13:33AI agents need data coming

13:36from an average of 400 data sources

13:39and more than

13:4085% of organizations

13:43uh require more than 100 data sources.

13:46So, these numbers are

13:47are amazing.

13:49And and almost 70% also report problems,

13:53probably because of this data

13:54distribution, report problems getting a

13:57agents to comply with consistent

14:00governance policies and consistent and

14:02consistent workflow, right?

14:04So, I I think that's the main conclusion

14:06that you can get from the report that

14:09the

14:10uh the reason why there is this trust

14:12gap, the reason why

14:14uh the agents sometimes

14:17um yeah, overreach or or or or providing

14:20inconsistent results uh is are mostly

14:22data problems.

14:23>> Can you break some of those down of

14:28live data, for example? Is is that data

14:32architecture problem?

14:33>> You could say it's an architecture

14:35problem because as as you were

14:37mentioning before

14:38initially, AI was focusing on analytics,

14:41so it was based in these data

14:44warehouses, data lake house systems that

14:46force you to first copy all the data

14:48there.

14:49So, when you have a data replication

14:51process, you always have some latency.

14:53So, the data will never be 100% real

14:56time.

14:57That's fine for analytics

14:59typically when you are trying to find

15:01trends or things like that

15:03uh but it's not good enough for many uh

15:07uh

15:08business workflows, right? So

15:11that is why, let's say, the old

15:13foundation for AI is is

15:16is not good enough to to get uh

15:19real-time data. As I said before

15:2270% of the of the organization said that

15:26even data that is 1 minute is stale,

15:29you know, is is is not valid for some AI

15:32agents.

15:33>> How does the the Denodo semantic layer

15:37uh

15:38correct that or or enhance that?

15:41>> Yeah, because with with the with Denodo,

15:44you don't need uh to centralize the data

15:48up front. You don't need to replicate

15:49the data. Denodo is able to query the

15:52data where it lives, right? So, you can

15:54think about these data products that you

15:56create with Denodo as virtual data

15:58products. So, maybe one data product

16:01is created with our lake house and is

16:03pointing to the lake house, but maybe

16:05this other data product is pointing to

16:07your CRM or an operational database,

16:10right? Where data is updated live. So,

16:14uh

16:15actually, the data products that you

16:16create with Denodo can provide this data

16:19up-to-date in real time. Well, for

16:21instance, if you are using a centralized

16:23architecture, you first need to copy the

16:25data from that operational database to

16:27the central system. And that will always

16:29introduce some latency. So, data will

16:31not be in real time.

16:32>> The right data

16:34uh um is obvious.

16:37Uh the you need the right data for for

16:40the right uh

16:42action or answer. But, what's the issue

16:46there? What creates the trust gap in

16:48traditional architectures?

16:50>> There are several issues here, right? Uh

16:52because having the right data means

16:54different things. First, first, you need

16:56access to all the data.

16:58Um if you are confining your AI

17:01applications to a single system, for

17:03instance, to a data warehouse or a lake

17:04house,

17:05then you will probably be missing part

17:08of the data that is needed. So, that's

17:10one part of the problem. And the second

17:11part of the problem is this

17:13uh semantics that we were talking about,

17:15right? First, that the the semantics

17:18need to be described, and second, the

17:20semantics need to be consistent.

17:22Uh to explain the problem of semantics,

17:25I sometimes use an example, right? Um

17:28uh

17:29imagine that you

17:31have a very brilliant DBA, a new hire,

17:34uh

17:35with perfect uh SQL knowledge,

17:38but is a new hire, so uh

17:41lacks the business context, right? If

17:42you ask them for

17:44I don't know, for instance, in the

17:45healthcare space, the readmission rate

17:48of cardiac patients,

17:50and then you give them access to the raw

17:52tables that contain the data to answer

17:54that,

17:56uh

17:56but you only give them that, they will

17:58fail. They will fail because they they

18:00won't know, for instance,

18:02which specific diagnostic codes uh

18:05define all the different cardiac

18:06diseases, or

18:08uh they will not know what is the exact

18:10formula for calculating a readmission in

18:14this case. For instance, if if a patient

18:15is readmitted for a different diagnosis,

18:17is that a readmission or not, right?

18:20Um they might even have problems knowing

18:23what data is in each in each column

18:26table, right? So, without a semantic

18:28layer to define these business rules,

18:30and also in a consistent way across all

18:32data sources,

18:34even the the smartest expert is forced

18:36to to guess.

18:38Uh and in the world of AI, guesses are

18:40basically hallucinations, right?

18:43Um so, that's uh

18:46what we mean with the right data. First,

18:48having access to all the data. Second,

18:50the data with the right level of

18:52freshness. If you need this piece of

18:54data in real time, it should be real

18:55time. And third, with consistent

18:57semantics.

18:59>> From the user's point of view, is is it

19:01a dashboard, or is this all happening

19:05unseen uh to the user?

19:08>> The node is more like this

19:11a catalog of data products.

19:14Uh it's like a a web commerce, you know,

19:16like Amazon or or a a a a a digital

19:19storefront

19:21where the node doesn't actually move the

19:23data to create these data products. It's

19:26like a catalog of these data products.

19:28These data products maybe come from a

19:30cloud database or on-premises system or

19:32a SaaS application, right?

19:34And then these data products can be

19:36consumed

19:38basically using any interface that you

19:39want, right? We actually also have a

19:41data mart our own, let's say, data

19:43marketplace, which is a web application

19:46that human users can use. But also, as I

19:48mentioned before, these data products

19:50can be accessed with

19:52any tool, any tool that want to to

19:54consume that data using all those

19:56interfaces that I mentioned before.

19:58>> On these uh

20:00foundations of uh

20:02trustworthy AI, uh

20:05you you we talked about live data, the

20:07right data, and then guardrails. Can you

20:10talk about the guardrails uh

20:13that are necessary, why they're

20:15necessary?

20:16>> Uh yes. Yes, absolutely. Well,

20:20I I think, you know, there are some

20:21types of war rails that everybody

20:23understand, everybody talks about. I

20:25don't know, things like ensuring that

20:26the prompts are safe,

20:28um protecting yourself against prompt

20:30injection, uh forbidding the agents the

20:33access to I don't know, to certain

20:35tools. Everybody understand that, right?

20:37But I think what sometimes

20:40goes under the radar is that

20:42as I mentioned before and the studies

20:44shows, agents need to pull data from a

20:49massive variety

20:51of data sources, right?

20:53Um

20:54and

20:56and when you have to to to work with

20:59multiple data sources,

21:01then uh you have several problems,

21:03right?

21:04Uh for instance, the problems of the

21:06consistent semantics that we mentioned

21:08before.

21:09But also

21:10how do you enforce

21:12consistent security policies across all

21:15those data sources? So,

21:17for instance, imagine that you have a

21:19customer support agent,

21:21but

21:22because of GDPR or any other regulation

21:24or whatever,

21:25the agent is not allowed to see um

21:29certain data from customers, right?

21:32And that data from customers

21:34will not be or in many cases may be

21:36available in several data sources, in

21:38several places, right?

21:40So, if you don't have something like

21:42Denodo and the agent needs to access

21:44those different data sources, you will

21:45need to implement those security rules

21:48in several places, right?

21:50With Denodo, you can implement that

21:52ruling only one place and you can be

21:54sure that it will be enforced for all

21:56the agents, for all the data sources and

21:58so on, right?

22:00Uh also I I would like to

22:02briefly

22:03mention or briefly comment why agents

22:05need to access to so many data sources,

22:08right?

22:09Well,

22:10for many reasons, right? First because

22:12organizations are very complex, data is

22:14distributed in many places.

22:17Uh also because real-time data sometimes

22:20prevent consolidation as we mentioned

22:22because you need to query the data where

22:23it lives.

22:25But a a and a third reason and this is

22:28very fundamental about about agents is

22:31that agents are unpredictable by

22:32definition.

22:34So, by definition, agents we use agents

22:37when we don't have

22:39a predefined set of rules where we have

22:41a task, where we have a task where a

22:43predefined set of rules

22:45uh

22:46does not cover all the scenarios, right?

22:48Because if we had a predefined set of

22:49rules that covers all scenarios, we

22:51could use a conventional workflow.

22:54So, the whole point point of an agent is

22:56that it needs to find solutions for

22:59problems that they have never seen

23:02before in exactly the same form.

23:05So, and if you think about it, this

23:07means that we don't know in advance

23:11exactly

23:12what data an agent will need.

23:14Uh in a typical workflow, we know, okay,

23:16this workflow needs this piece of data,

23:17this piece of data, and we can expose

23:19only that piece of data.

23:21But with an agent,

23:23uh you have to give them access to all

23:25the data that potentially might be

23:28relevant for the task at hand.

23:30And that also contributes to increase a

23:32lot the number of of data sources.

23:34>> You were talking about live data. Why is

23:37live data becoming critical for

23:39trustworthy AI?

23:41What what breaks when AI acts on stale

23:44data?

23:45>> Basically, that leads to to bad

23:47decisions or has the potential to to to

23:51make the

23:51to cause the agents to make bad

23:54bad decisions. For instance,

23:56if your agent is talking with a customer

23:58about a service down incident,

24:01uh

24:02it needs to know the state

24:04of the service now, not 5 minutes ago.

24:07Or if it is it is deciding how to

24:09allocate a certain resource,

24:12uh

24:13for instance, a seat in a in a plane for

24:15an upgrade or something like that,

24:17right? It needs information about the

24:19resource availability now. The resource

24:21availability

24:2310 minutes ago or 1 hour ago may not be

24:25good enough.

24:26Not all workflows are like this, right?

24:28There will be agents that maybe will not

24:31need real-time data. For instance,

24:33agents that work only for analytics, as

24:34I mentioned before, sometimes don't need

24:36real-time data or or don't need

24:39but there are many many workflows that

24:41actually need real-time data. In other

24:43case,

24:44uh the agents will

24:45will not be able to make good decisions.

24:48>> The challenge of enforcing these

24:50guardrails and governments

24:53>> [clears throat]

24:53>> across distributed environments

24:57uh

24:59how how do you handle that?

25:01Uh and then how does uh a Gen AI make

25:05that even harder?

25:06>> Since the node is this single entry

25:09point for agents to to get the data or

25:11this data products are the single entry

25:13point to get the data, you can be sure

25:15that those policies will be always

25:17enforced, right?

25:18Um behind the scenes, maybe the data for

25:21that product is coming from several data

25:24sources.

25:25But uh the agent never interacts

25:28directly with the data sources. It

25:29always goes through this virtual data

25:31product offered by the node. And it is

25:34there where the

25:36um

25:37security policies and the war rails are

25:39are enforced are defined first and then

25:41enforced.

25:42>> And these problems that we're talking

25:44about underlying the trust gap that that

25:47you you guys see um

25:51in enterprises regarding uh

25:54integration of AI?

25:56>> You know, the organizations were

25:58through several stages in the last 2

26:00years, right? The first stage was,

26:01"Okay, this is very cool. Let's start to

26:03experiment. Let's start running pilots."

26:07Then the second stage was,

26:09you know, uh "Wow, this is this is this

26:12sometimes is, you know, this makes for

26:14impressive demos and sometimes it

26:16actually

26:18uh

26:19provides fantastic insights or fantastic

26:21suggestions, but also

26:23uh very often

26:25makes these strange decisions or or or

26:28or overreaches. Maybe the agent

26:30overreaches.

26:31Does things that it should not do.

26:33And then wow, and and and initially, as

26:36you mentioned before, this is not an

26:38expert system. This is not symbolic,

26:40neuro-symbolic. So, at first, I cannot

26:42know what went wrong, right? So, I think

26:45that that that was the second stage. The

26:47third stage was, "Okay, we

26:49we are investing a lot of money on this.

26:51Let's analyze what's going on. Let's

26:52analyze what's going on.

26:54So, fair first stage is, okay, let's

26:57invest in auditability. Let's invest in

26:58traceability and auditability because

27:00actually

27:01those agents are, you know,

27:04using external tools and so on and so

27:06on, you are quite able to know exactly,

27:08okay,

27:08I got this data, then I analyze the

27:10data, in function of this I decided this

27:13and so on. So, actually you can get good

27:15traceability if you really invest in it.

27:18And then, analyzing these failure modes,

27:21it's when I think the these

27:23organizations discover these gaps that

27:25are mostly data gaps, right?

27:28I I I think that changes the the

27:29strategy. I I I I am for instance

27:32I

27:33I

27:34I I I was in

27:36last week a a very recent report from

27:38IDC

27:39from the from the from the

27:41and they were I was actually comparing

27:43the report from 2026 and the report from

27:452025. The report from 2026 was released

27:47I think in April and the other one I

27:49think is February 2025, right? And the

27:52question was, what is the main barrier

27:54or no, what is your main priority for

27:56investment in AI?

27:58And in 2025

28:00uh

28:01uh revamping your data architecture for

28:03AI was like in fifth place or something

28:05like that or sixth place. And now it's

28:07the first place, right? And I I think uh

28:11that change is because of this

28:13realization

28:14that uh

28:15that where

28:17the the failure modes really really are.

28:19>> That's interesting. So, we're at a point

28:22uh in in the

28:25adoption by enterprise where there was

28:29the pilot phase then they took it uh

28:34a little further and ran into these

28:36problems

28:38uh understood that it's data and that's

28:41where Denodo comes in to to

28:45unify the

28:47disparate data

28:49sources and then provide these

28:52unified universal policies. How should

28:56executives think about balancing

28:59existing lakehouse investments with

29:02the new

29:04layers that a genetic AI requires?

29:08And

29:09and what organizations

29:11what should organizations do

29:15as they scale AI?

29:18>> Yeah, um

29:20well, I think that for most

29:21organizations the lakehouse is is part

29:23of the solution, right?

29:25It contains a a lot of valuable data,

29:28it's a great foundation for data

29:30analytics and data science.

29:32But I also think it's not enough. Even

29:34in the world before AI

29:37there will always be data that is not in

29:39the lakehouse. You will

29:41sometimes need to access real-time data.

29:44You will need a strong semantics across

29:46several data sources. So even in the

29:48world before of AI we think that

29:51lakehouses were not the the final

29:53answer. Very useful, but not the final

29:56answer.

29:57But with AI, all those problems are

30:00multiplied.

30:01And it and that is why we think that

30:04this new data foundation is needed. Uh

30:08So our recommendation would be leverage

30:11the investments that you have, the data

30:13sources that you have,

30:14but prepare this common infrastructure,

30:18right? Uh data infrastructure that goes

30:20beyond central system that actually

30:22covers all your systems.

30:25Uh

30:26and ensure that your semantics, your

30:27governance, and so on is is defined

30:29there.

30:30And this is also very related to

30:33what organizations should avoid, what

30:35organizations should not do.

30:37Well, I I I see a lot of companies

30:40falling into what I call the ad hoc

30:43trap.

30:44Meaning that, you know, since because

30:47these traditional architectures have

30:49limits for the new use cases,

30:52and there's so much pressure to to

30:53deliver results,

30:55it's tempting to take shortcuts, right?

30:58It's tempting, for instance, to create

31:00an ad hoc data layer just just for one

31:03specific AI AI app. And then another

31:06data specific data layer only for this

31:09other uh AI app, right? Uh

31:13or or you may say, "Okay,

31:15this agent will be a customer support

31:17agent. So, I will restrict it to work

31:19only inside the CRM data, right?"

31:24A- And and that might give you a quick

31:27win, but we really think it's a dead end

31:30because we are rapidly moving towards a

31:33multi-agent world.

31:35And in this world, agents will not work

31:37in isolation. Uh they will need to

31:40collaborate. And if every agent is built

31:43on a different data silo with its own

31:45definitions,

31:46uh

31:47they won't be able to talk to each

31:49other. They won't be able to cooperate.

31:52That's why, you know, those are the two

31:54traps that I see for organizations. The

31:56first thing is, "Okay,

31:58let's try to create AI on a single

32:00system. We will try to centralize all

32:02data here." That will never work for big

32:04organizations. And the second trap is,

32:06"Okay, since the first thing is not

32:08working, let's do ad hoc solutions for

32:10every agent." And that will not work in

32:12the long term, either, because agents

32:14need to cooperate and talk to each

32:15other.

32:16>> And these other solutions, centralizing

32:19the data or building

32:21uh data layers for each AI

32:24or agentic uh application, is are

32:28companies, enterprises,

32:31uh before they find another Are they

32:33doing that themselves? Are they using

32:36consultants? Is that a solution that

32:39they come up with

32:41um because

32:43it's it seems an obvious fix or are

32:46there

32:48uh

32:48people in the market telling them to do

32:51that that that's the solution?

32:53>> A a bit of everything, right? Obviously,

32:55on one hand, obviously,

32:57you know, each vendor tries to

33:00bring

33:01the data to their systems and and to

33:03have their AI engines work on data,

33:06right? So, obviously,

33:08the vendors try to do that, but if you

33:10think about it, where where does that

33:13end? If all of them are partially

33:15successful,

33:17then at the end you what you have is

33:18many silos, right? With different AIs

33:20that cannot talk to talk to each other.

33:24So, part of it is the market, of course,

33:26but part is also a natural evolution of

33:29things, right? So, you start to For

33:31instance, as I said before, the first AI

33:33applications that you do are in

33:34analytics.

33:36So, you say, it's natural to think,

33:37"Okay, I will use my analytic systems

33:40for that or my main analytic system for

33:41that." And then only later you realize,

33:44"Oh, this is not enough because now my

33:46agents go beyond analytics."

33:48And then when you realize that, you say,

33:50"Wow, I have a lot of pressure to

33:52deliver, right? Because I I really need

33:54to deliver this." And then the solution

33:56is, "Okay, I will do something ad hoc.

33:58It will work for this and and let's see

34:01what happens, right?" So, there are also

34:03very very natural

34:05um very natural trends uh

34:08uh but I think that if if you think long

34:12term,

34:13you see that you know, you need to go

34:16beyond that.

34:17>> This is

34:18happening

34:20quickly. This

34:21uh the trust gap is has formed quickly

34:25and then

34:26uh

34:27you know, companies are

34:30are building these systems quickly and

34:33and the capability of the systems is

34:35advancing quickly.

34:37Uh how does Denodo uh keep up with all

34:41of that? Is your product also evolving?

34:44>> Well, I think this is something that

34:46obviously is not happening only to

34:47Denodo, also to

34:50to everyone, right? In the space.

34:53Uh

34:53if I I am as I I am the technology

34:56officer, right? So, I own the road map

34:58of the product and if I see my road map,

35:00I don't know, 5 years ago and now, wow,

35:03now it's completely dominated for AI,

35:05right? In those two flavors that I

35:07mentioned before. What AI can do for

35:09Denodo, what Denodo can do for AI,

35:11right?

35:12Uh um to to be honest,

35:15this is because also we see the trend as

35:18unstoppable. Um we are seeing for

35:20instance the evolution of our users and

35:22our our users

35:23and 5 years ago, probably this would

35:26have been unthinkable, but we are seeing

35:27like, okay, now maybe we are starting,

35:30you know, the the number of non-human

35:31users is growing so fast, right? That

35:34actually that changes a lot of things.

35:35That changes a lot of things about your

35:37product and about how you should expose

35:39the data and how you should represent

35:40the data.

35:41Uh it's of course a learning process

35:43because as you said, things are

35:45are changing so quickly that we are all

35:47learning, right? Of course, we are all

35:49learning.

35:50Um

35:51as I said before, I think these

35:54lessons that we were discussing were the

35:57lessons that many organizations got

36:00after this 1 year or 2 years of pilots.

36:03And now I think we are entering a stage

36:05of more maturity,

36:07but there will still be very significant

36:08changes because the landscape keeps

36:10keeps moving.

36:12Yeah, we In our website, if you go to to

36:15denodo.com,

36:16uh

36:17you can download the report from there.

36:19Yeah, and also learn more about about

36:21Denodo.

36:21>> From

36:22Denodo's point of view, has this uh the

36:26the findings in this report, has it uh

36:30changed the way uh

36:32you're you you talked about the the road

36:35map map that you oversee?

36:37Uh

36:38has it uh informed your future

36:43uh

36:43you know, when you define the the trust

36:46problem?

36:48>> Yes. Um

36:50uh for instance, some of the topics, for

36:52instance, limited data access probably

36:55was even that we were anticipating,

36:58right? Um so, for instance, that has had

37:02a an immediate impact right on

37:04optimizing even more those type of

37:06accesses, right?

37:08Uh so, that's only one one example.

37:11Um

37:13I think we were already pretty strong on

37:15those particular areas, right?

37:17But for instance, that would be an

37:18example where we we were actually were

37:21surprised about how the strong how

37:23strong the results were. Uh and that has

37:26obviously also has an impact in on what

37:28we do.

37:28>> Are you guys Is this a period of uh

37:32growth for you guys

37:35now that everyone's adopting AI?

37:38>> Ye- yes, I Well, Denodo, you know, is a

37:41consolidated player in this place. We

37:43have been

37:44in the data space for for many years.

37:46You know, for instance, we have been

37:49uh leaders in the

37:51data integration magic quadrant from

37:52Gartner for six or seven years in a row

37:54or something like that, right? But

37:56certainly AI is is being an accelerator

37:59for us. It's being a clear accelerator

38:00for us. I think it's because

38:02those changes that we have been

38:04discussing, right? And that actually we

38:06we think that fit very well with our

38:08original vision.

38:10Uh so, yeah, uh especially in the last

38:12year, AI is being a

38:15uh

38:16uh a great accelerator for us, yeah.

38:18>> What about uh

38:19sort of general adoption?

38:22uh you've identified uh this the the

38:25data

38:26problems that create the trust gap

38:30that have been slowing adoption.

38:33Uh

38:34if if that's solved with Denodo

38:39uh do you you see AI adoption? People

38:43ask me all this all the time. In in how

38:46long is it going to be before all uh

38:50large enterprises

38:53are using a Gen AI in in in all relevant

38:57processes and

39:00uh

39:00you know is it 5 years, 10 years, or or

39:05uh or less?

39:06>> If we say all organizations in all

39:08relevant processes, I I think it will

39:09take significant time because at the end

39:11of the day there are bottlenecks. I

39:13think many cases are related to

39:14organization or even legal reasons in

39:16some cases, right?

39:18But if we lower the bar the bar a little

39:20bit and and we say, "Okay, let's say

39:23the 25% most advanced companies

39:27using it in critical workflows, critical

39:32workflow let's say 20% of this their

39:34critical production workflows

39:37I would say by the end of 2027

39:41and

39:41and maybe I am being

39:44well, you know, all the you know

39:47uh

39:47I was going to say that maybe I I I am

39:50being too conservative because I I I see

39:52the the thing the things changing

39:55changing quite um fast, but it's also

39:57true that at the end of the day in big

39:59organizations everything takes more than

40:01expected uh

40:02because there are, you know, all types

40:04of bottlenecks, not only technology

40:05bottlenecks. From the technology point

40:08of view, to be honest, I think

40:10the technology is ready, the models are

40:13good enough

40:15for most of the

40:17use cases that we are considering. I

40:19think with solutions like this you can

40:20also solve the data problem. So I think

40:23that from the technology point of view

40:25the technology is mostly there.

40:29Also the data ability has improved a

40:30lot.

40:31I think the technology is mostly there.

40:33So that's why I think at least the most

40:36advanced companies

40:38that are most advanced more mature with

40:40technology it will go quite fast.

40:43Actually

40:44another thing that you see in this

40:46survey and also in other surveys like

40:48the IDC survey that I mentioned before

40:52is that actually the companies with a

40:54high level of data maturity for instance

40:56that already have a well established

41:00data products foundation and so on are

41:02actually actually in those ones the

41:04level of adoption of AI in production is

41:06already significantly higher, right?

41:08So I think technology probably is not

41:12the bottleneck anymore.

41:14>> Yeah. Yeah, that's fascinating.

41:18Okay, is there anything that I haven't

41:22touched on that you'd like listeners to

41:25hear?

41:26>> No, I think you know it was quite

41:29comprehensive,

41:32hopefully it has been also useful for

41:33the listeners. Okay.