theCUBE + NYSE Wired: Mixture of Experts Series | George Fraser, Fivetran & dbt Labs

More from theCUBE + NYSE Wired: Mixture of Experts Series

Gemma Allen

Host, theCUBE + NYSE Wired

theCUBE

George Fraser

CEO

Fivetran

Info

George Fraser, Fivetran & dbt Labs

Gemma Allen

Host, theCUBE + NYSE Wired theCUBE

HOST

George Fraser

CEO Fivetran

search

Gemma Allen

>> Palo Alto Studio Connection, Silicon Valley and Wall Street.

George Fraser

>> I'm John Furrier, co-hosting with David Vellante, my co-host.

Gemma Allen

>> Welcome back to theCUBE Studio here at the New York Stock Exchange. I'm Gemma Allen, co-host of NYSE Wired, Mixture of Experts. And today we're going to have a conversation about how every company wants AI agents. But here's the problem. Most companies aren't actually ready for them. Fivetran's research found that 15% of organizations say they're fully prepared when it comes to AI in production, even as nearly 60% are investing millions in it. Joining me now to unpack that gap is George Fraser, CEO and co-founder of Fivetran. Welcome, George.

George Fraser

>> Great to be with you.

Gemma Allen

>> So I guess we'll just get straight into it, George. Those are some interesting findings you guys released earlier this year. Help us understand the gap. What does it mean to be agentic AI ready in production?

George Fraser

>> Yeah, so I think the thing that we're seeing companies deploy at scale successfully so far is using coding agents for software engineering as assistants to their software engineers. And that's— I don't want to diminish that at all. That's a huge deal. A lot of people, including Fivetran, are getting a lot out of that. But if you want to do general-purpose knowledge work with coding agents, with agents in general, you have to connect those agents to your company's data. And that is something that we are seeing most companies are at a very early stage with. They're still figuring out what that should look like, what the underlying infrastructure needs to be to do that. And the good news is there are some great patterns that can help companies do that. But that's the thing where today I think most companies are at a very early stage.

Gemma Allen

>> So Fivetran, you guys have had an interesting year. You've just fully closed the dbt acquisition. Talk me through— merger, I'm sorry, merger, apologies. They need to be very specific about that. Talk me through this kind of moment that you guys are at from an industry perspective. it's clearly very opportunistic. That's a great synergy, making sure the data is clean and ready and useful, which has been a long-term challenge across enterprise in many industries. But bring me up to speed on where you guys are in September 2026 as the world goes back to school?

George Fraser

>> Well, from our perspective, AI agents are a new audience for the thing that we've been doing all these years, which is helping companies get all their data in one place and organize it. You could say, without oversimplifying too much, that Fivetran gets all your data in one place and dbt organizes it. dbt is the tool that customers use to organize their data. And that has always been important in the past. That's mostly been important for doing things like business intelligence and reporting. But now it turns out that that is the key to connecting AI to your company's data. The same data infrastructure that you've been using for years to prepare your data for decision support, as it is sometimes called, is the ideal infrastructure for connecting AI to your company's data. And the companies that are furthest along on this journey are really the frontier labs, OpenAI and Anthropic. And if you read what they write on their own blogs, this is exactly what they have done. They have taken the data infrastructure that they built to support reporting and they have put their AIs on top of it. And now the AI knows about the internal data as well as the external world.

Gemma Allen

>> So data has always been somewhat of an opportunity and a challenge, right? Especially if you think about the world of enterprise and the verticals that sit within that. Now we're hearing about this whole world of agentic, and we hear all the time data is gold, right? Garbage in, garbage out. Why is an AI agent fundamentally more demanding when it comes to data than, let's say, a chatbot was 3 years ago? what's truly changing underneath the hood?

George Fraser

>> Well, the comparison is really not between agents and chatbots, but between agents and people. So historically, the main carriers of your company's centralized data about everything happening in your company was human beings through reporting interfaces like BI tools and analysts writing SQL queries. And AIs are great at writing SQL queries and they're great at writing ad hoc reports. And so one of the things that we're seeing for the companies who do succeed in getting the infrastructure all wired in, getting their AI agents connected to their company's data, is that it immediately puts a lot of new pressure on their data infrastructure. It's a lot, there's a lot of value that it brings, but it also creates a lot of cost because they run a ton of queries because they are so very good at it. But the good news is that what we've seen is that this is a problem that can be solved with optimization. You don't need to build out a whole new data infrastructure to support AI agents. You can solve this problem by optimizing the infrastructure you have. And we've been doing a lot of work on that at dbt and Fivetran to make it cost feasible to connect your AIs to your data.

Gemma Allen

>> We hear a lot here on theCUBE NYSE Wired about, you know, knowledge capture, right? About tribal knowledge. About context, about the importance of building that into your AI data layer. When you think about the many enterprises you guys operate in and the interoperability of data from system to system, what are you seeing and hearing around, I guess, the role of AI in actually capturing knowledge and ensuring that, you know, this challenge of creating structured data, which has been a long-term challenge, is actually being solved for? In this era of AI? 15%, it's an interesting percentage, right? There's obviously a lot of things still not working as they should. How do you boil down those problems?

George Fraser

>> Well, I think, first of all, recognize that most of the data that you want already exists in a structured form in a system of record somewhere. I think sometimes people do foolish things in this area, like they point AIs at a collection of PDFs of forms filled out by insurance adjusters or something, only to find out that, hey, the contents of those PDFs were generated by a system, and the contents are all sitting in a database somewhere that you could have replicated. So I think you want to make sure to— first and foremost, don't turn this into a more exotic problem than it actually is. The very first thing you should do is get the data from all of the systems of record you already have and get it in one place. And by the way, you're probably already doing that. It's just a matter of incrementally adding new sources, new collections that maybe weren't relevant to the use cases that you were doing before but are relevant to AI. So I think a lot of techniques that are tried and true actually work really well in terms of getting all of your company data organized for AI.

Gemma Allen

>> I wanna ask you about something I know you've spoken about this year, George, and this is this whole concept of the SaaSpocalypse, right? You have certainly had some opinions on what you believe to be true and not so true in that scenario, but you mentioned systems of record. You guys obviously work across many, and there is certainly this view that, you know, are we moving towards one core orchestrator, like one core source of truth 10 years from now. I want to hear from you, like, what are your thoughts on what's happening from the perspective of fragmentation, which has been a challenge, I think, that somewhat fed the SaaSpocalypse narrative, especially as it relates to unified data and, I guess, elevating that proposition in the world of AI. Where do you see the reality there?

George Fraser

>> I think there will be beneficiaries of AI and there will be victims of AI from a company perspective. Companies who see a tailwind, companies that see a headwind. And we don't totally know who those are going to be right now. There's a lot of uncertainty. And I think the markets do price uncertainty. I think the markets maybe overreacted in and around April and maybe they're underreacting a little bit right now. There will be companies that are harmed by AI for sure. I think the other thing that people sometimes miss is even the companies for whom AI is competition or it makes their product less relevant. Change is slow. Adoption of new technology is very slow. And so we're going to see, even after the technology becomes available that maybe means you don't need X thing anymore, or there's a radically better way to do a certain task, you're gonna see shockingly slow progress at adoption of this across the economy because that's just how it always is. Most companies, they don't actually change their behavior unless they're actually under threat. So it's sort of a complicated, nuanced answer. I don't think it's a terribly non-consensus answer, actually, that, there will definitely be winners and losers, but it will also play out at most companies very slowly. And so, for businesses like Fivetran, we just try to figure out a way to be winners and then we try to navigate that over time.

Gemma Allen

>> When we think about what has happened historically and how software has grown up in the world of tech and enterprise, there was always a level of silo that was quite intentional, right? Especially from the perspective of security. And fragmentation, you didn't necessarily have one system of record or one human capable of accessing all variations of systems of record in any one firm for potentially good reason. Now we're hearing a lot about this whole kind of unified era, right? But that security threat, it's still very, very real. And perhaps in the world of agentic AI where it's an agent accessing multiple parts of your architecture at any one time, it's even more real. What are your thoughts around that argument, the argument for agents accessing systems of record and how secure that opportunity is or is not at this moment?

George Fraser

>> I think when it comes to internal corporate data and internal uses of that data, security concerns are mostly a pretense for company politics. At big companies, a lot of people at those big companies like silos. They like it if you want to get an answer to a question about a particular, about something in their domain, you have to come to them. And they get to shape the answer to that question. They don't like all the data to be in one place and accessible to everyone. And so I think a lot of these security concerns are actually about something else. There are legitimate security concerns for sure, but cryptography works. Role-based access control works. We have tools to solve these problems. If you look at companies who are far down the curve of agents accessing internal data, what you typically see, first of all, is they curate that data, usually with dbt, and they curate a subset of data that they are comfortable being broadly accessible to the entire company. What exactly that is is different at every company, But there's a lot of important stuff that you can— that is acceptable to be accessible to everyone. And so this is a simplified view, but the pattern is you have a small number of people, the data team, who have access to everything. They curate it into a subset that can be accessible to the company more broadly. You create a role that has access only to that broader subset. That is the role that the AI agent uses to access the database. these are not new ideas. They do work. So I don't mean to dismiss security concerns. They are real, but we have great tools for dealing with them. We have had these tools for decades. And if you're in a leadership position and you're trying to solve this problem, I would urge you, when people are flagging security concerns, pull on those threads because sometimes hiding behind those security concerns are actually just good old-fashioned corporate politics.

Gemma Allen

>> Well, I certainly see your point, George. So talk me through what's actually happening in the industry at this moment. what are you seeing from the perspective of adoption like that 15% of folks that are getting it right? Are there any verticals you see that seem to be adopting AI and kind of meeting that moment sooner, maybe in a more efficient way? talk me through what you're seeing broadly, the bird's eye view of enterprise AI adoption and agentic adoption right now?

George Fraser

>> Yeah, the companies that are further down that curve are generally technology companies. So you see the Frontier Labs themselves are like this, Fivetran is like this. you've seen companies like Ramp and Cerebras publish details of how they have curated their data for access within the company. And all of these examples are, like I said, using very similar patterns. They take the same data infrastructure that they built for internal reporting. They add additional data sources that didn't make sense until you had agents able to access the data. A lot of unstructured data sources— text was not super useful in the pre-AI era, but now we can actually comprehend it. They use a lot of the same tools. Curation is super important. It turns out dbt, great tool for curating data for AI, just like it was for reporting. And then interestingly, the semantic model seems to be the linchpin of a lot of this. You have to put a semantic model over the data that describes things like, this column of this table, it represents revenue. You can add it if you want to know the relationship to geo Geo is in this table over here, and the way you join to that is through this other column. That kind of information needs to get specified very precisely, but that's kind of the linchpin, at this present moment, it seems, of making AI successfully able to navigate the central database of everything happening in your company. So that's the pattern we see. We see it most commonly in technology companies. A lot of them have written about it, and you can go read about it. And I think this is the pattern that you're going to see everyone else emulate as this spreads throughout the economy.

Gemma Allen

>> It's certainly an interesting moment, right? When we think about some of the tactical challenges, something as simple as a unified taxonomy can actually be a hurdle in this agentic world too. So, George, last question to you. Just finished this merger. Seems like it's a very interesting time in the world of AI and enterprise technology. What's ahead? How are you guys thinking about this next year out as you think about moving forward together with one unified and synergistic offering to the market?

George Fraser

>> Yeah, we are 100% focused on how do we take what we've built and make it as useful as possible for this amazing new use case of connecting AI to data. So we're focused on solving cost problems, making sure that people are going to be able to get all these benefits without having insane data infrastructure costs. And we are focused on the semantics aspect, making sure that we make it as easy as possible for people to curate their data into a simplified form that, as we talked about earlier, is acceptable to give broad read access to the company and represents a simplified view of their business that an agent is going to be able to navigate successfully.

Gemma Allen

>> Well, George, great conversation. Wish you guys all the best as you move ahead on this next chapter. Thanks so much for joining us on NYSE Wired. Thank you very much. I'm Gemma Allen here at theCUBE Studio at the New York Stock Exchange. This is NYSE Wired, Mixture of Experts. Thanks for watching.

George Fraser, Fivetran & dbt Labs

search

Gemma Allen

>> Palo Alto Studio Connection, Silicon Valley and Wall Street.

George Fraser

>> I'm John Furrier, co-hosting with David Vellante, my co-host.

Gemma Allen

>> Welcome back to theCUBE Studio here at the New York Stock Exchange. I'm Gemma Allen, co-host of NYSE Wired, Mixture of Experts. And today we're going to have a conversation about how every company wants AI agents. But here's the problem. Most companies aren't actually ready for them. Fivetran's research found that 15% of organizations say they're fully prepared when it comes to AI in production, even as nearly 60% are investing millions in it. Joining me now to unpack that gap is George Fraser, CEO and co-founder of Fivetran. Welcome, George.

George Fraser

>> Great to be with you.

Gemma Allen

>> So I guess we'll just get straight into it, George. Those are some interesting findings you guys released earlier this year. Help us understand the gap. What does it mean to be agentic AI ready in production?

George Fraser

>> Yeah, so I think the thing that we're seeing companies deploy at scale successfully so far is using coding agents for software engineering as assistants to their software engineers. And that's— I don't want to diminish that at all. That's a huge deal. A lot of people, including Fivetran, are getting a lot out of that. But if you want to do general-purpose knowledge work with coding agents, with agents in general, you have to connect those agents to your company's data. And that is something that we are seeing most companies are at a very early stage with. They're still figuring out what that should look like, what the underlying infrastructure needs to be to do that. And the good news is there are some great patterns that can help companies do that. But that's the thing where today I think most companies are at a very early stage.

Gemma Allen

>> So Fivetran, you guys have had an interesting year. You've just fully closed the dbt acquisition. Talk me through— merger, I'm sorry, merger, apologies. They need to be very specific about that. Talk me through this kind of moment that you guys are at from an industry perspective. it's clearly very opportunistic. That's a great synergy, making sure the data is clean and ready and useful, which has been a long-term challenge across enterprise in many industries. But bring me up to speed on where you guys are in September 2026 as the world goes back to school?

George Fraser

>> Well, from our perspective, AI agents are a new audience for the thing that we've been doing all these years, which is helping companies get all their data in one place and organize it. You could say, without oversimplifying too much, that Fivetran gets all your data in one place and dbt organizes it. dbt is the tool that customers use to organize their data. And that has always been important in the past. That's mostly been important for doing things like business intelligence and reporting. But now it turns out that that is the key to connecting AI to your company's data. The same data infrastructure that you've been using for years to prepare your data for decision support, as it is sometimes called, is the ideal infrastructure for connecting AI to your company's data. And the companies that are furthest along on this journey are really the frontier labs, OpenAI and Anthropic. And if you read what they write on their own blogs, this is exactly what they have done. They have taken the data infrastructure that they built to support reporting and they have put their AIs on top of it. And now the AI knows about the internal data as well as the external world.

Gemma Allen

>> So data has always been somewhat of an opportunity and a challenge, right? Especially if you think about the world of enterprise and the verticals that sit within that. Now we're hearing about this whole world of agentic, and we hear all the time data is gold, right? Garbage in, garbage out. Why is an AI agent fundamentally more demanding when it comes to data than, let's say, a chatbot was 3 years ago? what's truly changing underneath the hood?

George Fraser

>> Well, the comparison is really not between agents and chatbots, but between agents and people. So historically, the main carriers of your company's centralized data about everything happening in your company was human beings through reporting interfaces like BI tools and analysts writing SQL queries. And AIs are great at writing SQL queries and they're great at writing ad hoc reports. And so one of the things that we're seeing for the companies who do succeed in getting the infrastructure all wired in, getting their AI agents connected to their company's data, is that it immediately puts a lot of new pressure on their data infrastructure. It's a lot, there's a lot of value that it brings, but it also creates a lot of cost because they run a ton of queries because they are so very good at it. But the good news is that what we've seen is that this is a problem that can be solved with optimization. You don't need to build out a whole new data infrastructure to support AI agents. You can solve this problem by optimizing the infrastructure you have. And we've been doing a lot of work on that at dbt and Fivetran to make it cost feasible to connect your AIs to your data.

Gemma Allen

>> We hear a lot here on theCUBE NYSE Wired about, you know, knowledge capture, right? About tribal knowledge. About context, about the importance of building that into your AI data layer. When you think about the many enterprises you guys operate in and the interoperability of data from system to system, what are you seeing and hearing around, I guess, the role of AI in actually capturing knowledge and ensuring that, you know, this challenge of creating structured data, which has been a long-term challenge, is actually being solved for? In this era of AI? 15%, it's an interesting percentage, right? There's obviously a lot of things still not working as they should. How do you boil down those problems?

George Fraser

>> Well, I think, first of all, recognize that most of the data that you want already exists in a structured form in a system of record somewhere. I think sometimes people do foolish things in this area, like they point AIs at a collection of PDFs of forms filled out by insurance adjusters or something, only to find out that, hey, the contents of those PDFs were generated by a system, and the contents are all sitting in a database somewhere that you could have replicated. So I think you want to make sure to— first and foremost, don't turn this into a more exotic problem than it actually is. The very first thing you should do is get the data from all of the systems of record you already have and get it in one place. And by the way, you're probably already doing that. It's just a matter of incrementally adding new sources, new collections that maybe weren't relevant to the use cases that you were doing before but are relevant to AI. So I think a lot of techniques that are tried and true actually work really well in terms of getting all of your company data organized for AI.

Gemma Allen

>> I wanna ask you about something I know you've spoken about this year, George, and this is this whole concept of the SaaSpocalypse, right? You have certainly had some opinions on what you believe to be true and not so true in that scenario, but you mentioned systems of record. You guys obviously work across many, and there is certainly this view that, you know, are we moving towards one core orchestrator, like one core source of truth 10 years from now. I want to hear from you, like, what are your thoughts on what's happening from the perspective of fragmentation, which has been a challenge, I think, that somewhat fed the SaaSpocalypse narrative, especially as it relates to unified data and, I guess, elevating that proposition in the world of AI. Where do you see the reality there?

George Fraser

>> I think there will be beneficiaries of AI and there will be victims of AI from a company perspective. Companies who see a tailwind, companies that see a headwind. And we don't totally know who those are going to be right now. There's a lot of uncertainty. And I think the markets do price uncertainty. I think the markets maybe overreacted in and around April and maybe they're underreacting a little bit right now. There will be companies that are harmed by AI for sure. I think the other thing that people sometimes miss is even the companies for whom AI is competition or it makes their product less relevant. Change is slow. Adoption of new technology is very slow. And so we're going to see, even after the technology becomes available that maybe means you don't need X thing anymore, or there's a radically better way to do a certain task, you're gonna see shockingly slow progress at adoption of this across the economy because that's just how it always is. Most companies, they don't actually change their behavior unless they're actually under threat. So it's sort of a complicated, nuanced answer. I don't think it's a terribly non-consensus answer, actually, that, there will definitely be winners and losers, but it will also play out at most companies very slowly. And so, for businesses like Fivetran, we just try to figure out a way to be winners and then we try to navigate that over time.

Gemma Allen

>> When we think about what has happened historically and how software has grown up in the world of tech and enterprise, there was always a level of silo that was quite intentional, right? Especially from the perspective of security. And fragmentation, you didn't necessarily have one system of record or one human capable of accessing all variations of systems of record in any one firm for potentially good reason. Now we're hearing a lot about this whole kind of unified era, right? But that security threat, it's still very, very real. And perhaps in the world of agentic AI where it's an agent accessing multiple parts of your architecture at any one time, it's even more real. What are your thoughts around that argument, the argument for agents accessing systems of record and how secure that opportunity is or is not at this moment?

George Fraser

>> I think when it comes to internal corporate data and internal uses of that data, security concerns are mostly a pretense for company politics. At big companies, a lot of people at those big companies like silos. They like it if you want to get an answer to a question about a particular, about something in their domain, you have to come to them. And they get to shape the answer to that question. They don't like all the data to be in one place and accessible to everyone. And so I think a lot of these security concerns are actually about something else. There are legitimate security concerns for sure, but cryptography works. Role-based access control works. We have tools to solve these problems. If you look at companies who are far down the curve of agents accessing internal data, what you typically see, first of all, is they curate that data, usually with dbt, and they curate a subset of data that they are comfortable being broadly accessible to the entire company. What exactly that is is different at every company, But there's a lot of important stuff that you can— that is acceptable to be accessible to everyone. And so this is a simplified view, but the pattern is you have a small number of people, the data team, who have access to everything. They curate it into a subset that can be accessible to the company more broadly. You create a role that has access only to that broader subset. That is the role that the AI agent uses to access the database. these are not new ideas. They do work. So I don't mean to dismiss security concerns. They are real, but we have great tools for dealing with them. We have had these tools for decades. And if you're in a leadership position and you're trying to solve this problem, I would urge you, when people are flagging security concerns, pull on those threads because sometimes hiding behind those security concerns are actually just good old-fashioned corporate politics.

Gemma Allen

>> Well, I certainly see your point, George. So talk me through what's actually happening in the industry at this moment. what are you seeing from the perspective of adoption like that 15% of folks that are getting it right? Are there any verticals you see that seem to be adopting AI and kind of meeting that moment sooner, maybe in a more efficient way? talk me through what you're seeing broadly, the bird's eye view of enterprise AI adoption and agentic adoption right now?

George Fraser

>> Yeah, the companies that are further down that curve are generally technology companies. So you see the Frontier Labs themselves are like this, Fivetran is like this. you've seen companies like Ramp and Cerebras publish details of how they have curated their data for access within the company. And all of these examples are, like I said, using very similar patterns. They take the same data infrastructure that they built for internal reporting. They add additional data sources that didn't make sense until you had agents able to access the data. A lot of unstructured data sources— text was not super useful in the pre-AI era, but now we can actually comprehend it. They use a lot of the same tools. Curation is super important. It turns out dbt, great tool for curating data for AI, just like it was for reporting. And then interestingly, the semantic model seems to be the linchpin of a lot of this. You have to put a semantic model over the data that describes things like, this column of this table, it represents revenue. You can add it if you want to know the relationship to geo Geo is in this table over here, and the way you join to that is through this other column. That kind of information needs to get specified very precisely, but that's kind of the linchpin, at this present moment, it seems, of making AI successfully able to navigate the central database of everything happening in your company. So that's the pattern we see. We see it most commonly in technology companies. A lot of them have written about it, and you can go read about it. And I think this is the pattern that you're going to see everyone else emulate as this spreads throughout the economy.

Gemma Allen

>> It's certainly an interesting moment, right? When we think about some of the tactical challenges, something as simple as a unified taxonomy can actually be a hurdle in this agentic world too. So, George, last question to you. Just finished this merger. Seems like it's a very interesting time in the world of AI and enterprise technology. What's ahead? How are you guys thinking about this next year out as you think about moving forward together with one unified and synergistic offering to the market?

George Fraser

>> Yeah, we are 100% focused on how do we take what we've built and make it as useful as possible for this amazing new use case of connecting AI to data. So we're focused on solving cost problems, making sure that people are going to be able to get all these benefits without having insane data infrastructure costs. And we are focused on the semantics aspect, making sure that we make it as easy as possible for people to curate their data into a simplified form that, as we talked about earlier, is acceptable to give broad read access to the company and represents a simplified view of their business that an agent is going to be able to navigate successfully.

Gemma Allen

>> Well, George, great conversation. Wish you guys all the best as you move ahead on this next chapter. Thanks so much for joining us on NYSE Wired. Thank you very much. I'm Gemma Allen here at theCUBE Studio at the New York Stock Exchange. This is NYSE Wired, Mixture of Experts. Thanks for watching.