AI Agents and the Fight for Customer Data

0:00There is a new reason to have all your data in one place, which is AI agents need context. If you don't do that, then it's sort of like using ChatGPT from before ChatGPT was connected to the internet. Postgres, contrary to popular belief, is very old technology. It is not a good database simply because it was written a long time ago. It has a lot of technical debt. Asatia has said that there's going to be the collapse of SaaS. Do you think the SaaSpocalypse is a thing and we're going to see a massive shift? The bigger threat is that AI-native companies will just zoom and catch up to the established incumbents and maybe be better.

0:35Like, we'll actually have an HR, and that HR team will onboard AIs as they come. They'll be part of teams, they'll join the Slack. And in that world, these aren't software. That's actually more seats, more consumption of software. And so do you think that for enterprise agents, we're moving more to these, you treat them like humans? Or do you think that that's twofold? For years, companies built data infrastructure to answer questions about the business. Now, they're building it for AI. As agents become more capable, the challenge is no longer collecting data. It's making sure the right systems can access the right context at the right time.

1:12That shift is forcing companies to rethink everything from data platforms and APIs to enterprise software and systems of record. Martin Casado speaks with Fivetran co-founder and CEO George Frazier about AI, data infrastructure, and why the next wave of enterprise software may look very different from the last. So our guest today is George Frazier, who is the CEO of Fivetran. Fivetran announced the merger with DBT. So maybe to start, just give a quick overview of what Fivetran does. So Fivetran, we've been around for a while. We've been around since 2013, had customers since 2015. 2013? Yeah, exactly.

1:54I've been doing this long enough that a slide about the past state in my own slides is the same slide as the future state from when I started. But what Fivetrain does is we help our customers get all of their data from all their systems like Salesforce, NetSuite, all their SaaS tools, their own databases into one place. Getting all your data in one place is not a new thing. Businesses have had the need to do this since filing cabinets. The primary reason historically that people used 5Dran to get all their data in one place was to do business intelligence, was to build reports about things like what's your revenue, what's going on with your sales team, what are we forecasting for this quarter, all those great things.

2:35And now there is a new reason to have all your data in one place, which is if you want to use AI agents in business, AI agents need context. And it turns out that the same data foundations that work well for business intelligence and reporting with some additions and some modifications actually can work really well for AI agents as well. I mean, talk about a sector of the industry which is under a lot of change because of AI. And so maybe, can you give like a high-level overview of how it is evolving? What are some of the considerations about the shifts in data? And in particular, like, we're seeing a lot of changes how vendors view their own data, how the big labs use data.

3:14So just talk a bit about what the industry is. The thing about data in the context of business is it is always born somewhere else. It's always born in systems of record like Salesforce, like Workday, like SAP. And even if it's your own applications data, if you're a software company and you run your own database, the data is born in that database. And since, as I said, time immemorial, businesses have had the need for internal use to centralize a copy of all their data in another location. It doesn't work to just go and do all of your reporting and ask all your questions in each system individually.

3:55Some kinds of questions require you to look across the entire system. And so that is not new. However, these AI agents are new. And there has been in the last year, a reaction, which really started with the stock market. As we saw the SaaS apocalypse happen, and as we saw the stock prices of all these systems of record that I'm talking about plummet, people viewed them as under threat from AIs. We have seen some of these companies start to think that a great strategy for dealing with AI might be to lock it out and to say, your data is our data now, and you can't take it elsewhere. And if you want to use AI on it, you have to use the AI tools that we provide.

4:37Notably, just a couple weeks ago, SAP announced a new API policy that literally said all AI agent access was banned except in a way specifically approved by SAP. Now, if you're an SAP user, don't panic. This is just a policy. You have contracts with SAP. Those are authoritative as to what you are and are allowed to do. So don't overreact to these policy memos. but it just shows how extreme the reaction of some of these companies has been. I just want to tease this apart because I think there's a lot of confusion on what exactly is going on, right? So this is locking down access to the data that an agent would use instead of an app, right?

5:17It's not access to data because you're going to train your own model. That's right. Very few people are in the business of training their own models. Most people, when they want to access their own data in their own systems of record, even if those systems of record are managed by vendors, they are using it for context. They're using it in order to ask and answer questions about what's going on in their business. So the concern is my SaaS app has less value as an interface because now the agents can access the data directly and basically perform the same functions the SaaS app was before. Is that the concern?

5:52I think there are many concerns. I think that one of them... Can you just strawman the set of concerns? Because this is one of the biggest reactions I've seen in the industry in a very long time, and I'm kind of trying to come to grips with what the actual worry is. I think people are worried that their systems that they've spent many years building will simply be less valuable in a world where their users are no longer humans, but they're agents. I think they're worried that... But why isn't this just another seat? I mean, it seems like, I mean, arguably this is positive because there's going to be more consumers of...

6:24So agents don't need as many individual identities. these. When you have AI agents accessing systems, you really just need roles. You don't need the same granularity of users. You have many product managers, each will have their own identity in a system. But if you have a product manager bot, you really might just have one role that it uses, and it might have a single identity and yet do the work of hundreds or thousands of people. So there's not an easy answer like that. Furthermore, these companies have a history of having open APIs. Open APIs are a good thing. If these companies did not have open APIs, they would have been consigned to the dustbin of legacy SaaS decades ago.

7:10I mean, this is a thing that happened in the 90s, right? The evolution of open APIs. And their customers have been using them and depending on them for decades. And those same APIs are the primary target of AI agents. So it is very hard to differentiate whether the users are accessing the APIs in the same way that they always have been, or whether they are accessing them in agentic ways that may substitute for human workflows. I'm going to keep poking on this because I still think it's not a real concern. So let me make another straw of that argument. So let's say you had opened up all your APIs in the 90s, which is the case.

7:47Like, why couldn't I just write a procedural app, which is my own version of the SaaS, and therefore also disintermediate your SaaS? Like, why are agents somehow different than me just writing my own software or my own dashboard? Maybe they're not. This may all be kind of much ado about nothing. I think it's foolish for them to close down their APIs. So you're putting me in a weird position. No, I do. I'm trying to defend a position that I think is stupid. Yeah, yeah, yeah. I agree. So I think a lot of these threats are not new. Like, well, maybe they'll use programmatic access and thereby use less seats.

8:21Maybe they'll move some functionality to their own interfaces. I mean, that is a real thing that has been happening for years. I just want to let you know, I am old enough to remember these discussions in the 90s. The rhetoric was exactly the same. The reaction was exactly, we could never open up APIs. We can never have them do this because they're going to disintermediate us. And it just turns out that if you're buying into a business process, like the operational flow of something, that is set up by the company that you're buying it from. Salesforce knows how to run the Salesforce. And so whether it's an agent that's consuming it or SaaS, I would argue that there's still the value there.

8:54I completely agree. And I will point out another piece of evidence for that claim, which is if you look at the budgets of real companies that are heavy consumers of software, they spend 5 % to 10 % of headcount on software. Software costs are immaterial in the grand scheme of things. Software compared to everything else a typical business spends money on is so cheap. The idea that they're going to use AI to value engineer the number of seats they have on Slack or something is ridiculous. They're going to use AI to go make their business work better in whatever it is that they do. They're not trying to take that 5 % software spend and turn it into four and a half.

9:38That is not the highest, best use of AI. I mean, famously, all the big AI labs, including, I mean, Andreessen Horowitz, and we're all very heavy users of AI, like still use these SaaS tools. And so four years into this, we don't have a lot of evidence that like... As do we, as do OpenAI and Anthropa. Exactly. Who are both FiveTrain customers and we replicate lots of data from these very SaaS tools on their behalf into their data lakes. So if they're still using them, do we really think the company of the future is not going to be? Right, right. But one thing we can both agree on is this is bad for customers, right?

10:14This kind of locking down the APIs is bad for customers. So maybe talk through like A, in-wise bad, which may just be obvious, and then B, kind of your recommendation for how to manage that, like assuming that this is happening industry-wide. So the reason it's bad for customers, anytime vendors put up walls and try to regulate data access, is that you need to have all your data in one place in order to do meaningful reporting, in order to understand what the heck is going on in your business, and in order for AI agents to work in the context of business. If you don't do that, then it's using ChatGPT from before ChatGPT was connected to the internet.

10:55If you used it back then, you remember, it used to have this knowledge cut off and it would tell you I can only answer questions that happened in my training window. Six months ago. Yeah, six months ago was when I was trained. I don't know anything after that because I'm not connected to the internet. That is what using today's AIs is like if you're using them in a business context and they don't have access to your business data. And so it's very important for every company who wants to do things with data to create their own data platform where they have a copy of all of their own data and it's being kept continuously up to date.

11:28And anytime vendors start putting up barriers, it just makes it harder to get that done. And the customers will still do it. They'll just work around these barriers at great cost and complexity. One of my favorite things that you've done as a company to educate the customers on this is this benchmark. Can you maybe talk through what that is? Well, we have a website. Is that what you're talking about? Open Data Infrastructure? Open Data Infrastructure.com. Yeah, Open Data Infrastructure.com. It's a benchmark. Like it does scoring, right? Yeah, so we score. The list is growing. We're trying to make it as long as possible.

12:02We score as many vendors as we've been able to catalog so far on their data access policies. So we basically score them on whether they try to charge egress charges, whether they try to make you pay for getting your own data out, whether they make it impossible to get a complete copy of your data because some vendors will do that or they'll make it just very difficult. And whether they have terms of use restrictions on accessing your own data. So there's a big grid on there Are you comfortable saying who the worst offenders are? Or should you just point people to the website? Just go to the website and read it.

12:46It's all very fact-based and very well evidenced. I mean, the worst offenders historically have been, SAP has always been really bad. I mean, even when I was running a large business, I mean, that was always... It's interesting because they were getting better, and I've sensed in the last few years that there's sort of two camps within SAP, one of whom who regards it as it's the customer's data. The customer's got to be able to do what they need to do with their own data. And then there's sort of the old guard who views it as SAP's data and you'll do with it what we tell you to. And then, you know, historically, Salesforce has been really good with the exception of Slack, where they are terrible.

13:27Yeah, I know. But they've started to get squirrely about this. So it's a moving target. I am hopeful that this is merely a brief flirtation with closed data by most of these vendors, and they will realize this is not a good idea, that they are at war with their own customers, and it's not even going to work anyway. And we will go back to the trend towards ever more openness. You just answered my next question, but I'm asking anyways, which is like, do you feel like this is just like a repeat of the open data API? It'll resolve quickly. Like, you know, we always kind of go through this soul searching, and then we will resolve back to where we were, which is like open data is the right thing.

14:05I do think so. Do you think that's the path or do you think it's different this time? Yes, I think it will be the same. That's my prediction. It will be the same path as mentors will discover that they cannot provide inside their own platform a solution to every data problem that their customers have because they are simply so diverse and instead they simply create a mechanism for the customers to replicate it to their data platform of choice and do with it what they will. And even if they charge fees for that, It's not the end of the world. On opendatainfrastructure.com, you really get yellow for charging fees.

14:36It's only red when you try to actually block it. At the end of the day, if you want to have a little toll, that's not the end of the world. The problem is when you start actually blocking it. Like there is no option. And saying, oh, if you want to do anything with data, you have to come use my tools inside my walled garden, which never works because all the rest of your data is not in that walled garden, and it's not going to be. and you can never create enough tools to support all the different things customers want to do with data. You know, this is probably related to this notion of, you know, or this belief in data gravity.

15:11And I mean, one thing that I've loved working with you over the years is like you're exceedingly smart and you're exceedingly contrarian. It's just so fun to like kind of, you know, watch your opinions diffuse and more often right than wrong. And one thing you have said is that data gravity is either overrated or not real. So do you think that, like, A, do you stand by the statement, and B, do you think that this is driving people to, like, try to do these walled gardens? I think data gravity is completely fake. I am the only person who thinks this. I think you're the only person that says that.

15:44If you want to see evidence... Well, first, maybe describe what data gravity means. Data gravity... People use the term a lot, but, like, I don't think they understand the implications. Data gravity is the idea that business data is so large that it's very expensive to move around because of egress charges of cloud vendors. And if you want to see, and that therefore it's very important that you choose a physical region in the world where all your data is going to live in a specific region of a specific cloud, and then you build all of your data consuming services in that same location, or you can partition it.

16:23Especially because of egress issues. Well, this is Avers. The term data gravity does get used to mean multiple things. It's very general, okay. This is one particular incarnation of the idea of data gravity. And this is the one that I am saying is, Vic, that egress charges are so important. Yeah. And if you want to see evidence against this, come look at the networking dashboard of five trans various AWS and GCP accounts. You will be astonished, despite replicating huge data sets for thousands of companies. We have 7 ,000 customers of size and thousands more little ones. The amount of data being moved at any given time is tiny.

17:04And the reason is that we're doing change data capture. You can have a huge data set, but if you just replicate the changes, the changes are always much smaller than people think. And I think that a lot of this idea of data gravity came from dumb data pipelines that people wrote where they would copy their entire company's data sets out of their database every day, once a day, at midnight. And so they just had this crazy read amplification. You know, they were just repeatedly copying the same data over and over, and it gave them the impression that they had so much data, but they really don't.

17:39It may be a good argument not to roll your own on these things anyways. Yes, yes. This is what happens when people roll their own data pipelines, is they fall back on these patterns that are easy to get right. Like that pattern I just described, it's extremely easy to get a correct replica just by copying it all over and over, but very expensive to operate in the long run. Before we leave this topic, so let's say there's a CIO listening to this right now and the CIO is like, oh, how do I navigate these kind of uncertain three to six months? Like what kind of leverage do I have with these SaaS vendors as they're closing things off?

18:15Do you have any sort of guidance for them? Yeah, I think, number one, you should insist on having a copy of all of your own company data in a data lake that you control. Don't let go of that for any vendor. You have a lot of leverage. These vendors actually have a lot of obligations to let you do that. The reason why they get away with blocking people is simply because people don't fight. So pick fights. and write it into your, into your, if you have big contracts with vendors and you're redlining your MSAs, write language guaranteeing your own data access into those MSAs. Are you seeing those show up now?

18:59Actually, I hadn't even considered that. We have model language on opendatainfrastructure.com that we recommend you incorporate into your MSAs. And even if you don't get it, just by asking for it, you are sending a signal. So, you know, if it's a$10 ,000 contract, don't do it. But if you're signing, you know, $500 ,000 ,000 contracts, insist on data access in your MSA. And you will find surprisingly often that you get it. I want to talk a little bit about agents now. So I feel like we've kind of gone through multiple phases already in agents. So the first one is they're just purely treated as software.

19:40They're like, it was almost like Search++. Like, take all your data, put it into a data lake, and then you have an LLM that has access to it. And then we go to that, and it's like enterprise search all over again. And then we went to, like, the agents-like, like, OpenClaw. And then that model is like a personal agent, but it kind of, like, was part of you. So you'd give it access to your email, and you would give it access to, like, your API keys, and give it access to your accounts. So it was, like, part of you, an extension of you. Now we're saying— I actually set up OpenClaw. That's great. I use NanoClaw.

20:16Well, I set up OpenClaw and then I churned off of it onto NanoClaw because it's such a monstrous piece of over-complex software. And then I actually churned off NanoClaw onto Nanobot, which is what I've stuck with. But I use it to manage my tennis team. But it has its own identity. Okay, so this is what I was going to say. It has its own email. So I think we've all come to this conclusion, which is like, I don't want it having access to my email. I also have the Mac Mini. I run actually now I'm working with somebody like build a harness, but like they're all the same, right? I run it on a VM. Yeah, perfect.

20:47It's got its own WhatsApp number, its own email address. So I got its own phone number, its own email address. And then, you know, now as we think about A16Z, we're actually thinking about like, you know, why don't we just treat all these agents like this? Like we'll actually have an HR and that HR team will onboard AIs as they come. They will train them. They will show them the access to the documents that you need. They'll be part of teams that'll join the Slack, you know, just like humans do. And in that world, and we touched on this a little bit before, but in that world, these aren't software.

21:17There's actually more seats, more consumption of software. And so do you think that for enterprise agents, we're moving more to these, you treat them like humans, or do you think that that's too far? I think it is a intermediate form. I think the reason this works well is because you can slot it in to the existing workflows. without having to refactor the whole universe. So in my example of my agent that manages my USTA tennis team, it can email with the players, it can go to the USTA website, check the schedule, make lineups, check availability. And it works well having its own identity because it can slot into all of these existing workflows that were designed for humans.

22:02At Fivetran, we have an AI agent that helps respond to support tickets that goes and inspects the logs, inspects the code. It uses all the context that we centralize with Fivetran to find out what is going on with this customer, what might be the solution to this problem. And it drafts responses. and right now it slots into the system a lot like a person would but we are working on making the whole thing just a closed loop pure AI system where it will only have one identity. There will just be the connector troubleshooter Borg hive mind. And there may be but let me give the argument for the multiple.

22:51So why do I use a Mac mini for my agent, right? I mean, there's a couple of reasons. One of them is so it can access iMessage because there's no programmatic way to access iMessage. And so, like, you know, it has a desktop. Another one is, like, everybody's moving headless. Like, Salesforce is doing their headless thing and there's headless browsers. But it turns out, let's just talk about browsers. Like, if you have a headless browser, all of the anti-scraping software kicks in and then, like, it's not as functional. So it's actually much better to just give it a fully functional version of Safari.

23:24So you could argue that the interfaces that have evolved over the last 30 years to deal with unpredictable users that know how to use computers is the UIs we have today. And actually the simplest thing to do, rather than try to rewrite all of that stuff, is just to give these agents that have been trained on human data access to full end-to-end systems. I don't know. The systems that I've worked on, browser use has never been necessary. Browser use has a big cost, which is very slow. and it consumes a lot of tokens. Yes. And I have found that in my tennis example, the USTA website does not have any anti-scriping provisions.

24:07You're not trying to like read from LinkedIn or Zillow. Try that. Well, yes. And so I'm just using, I'm using just like Python browser automation. I'm actually using Selenium. And then there's actually a skill that emits just exactly what you want to know so that the agent is not reading HTML and consuming all those tokens all the time. And then at 5Tran, for example, we are working right now on a Salesforce administration agent to basically do continuous integration of small changes into our Salesforce, which is very labor-intensive right now. and it also does not use browser automation because the Salesforce CLI we found does everything we need to do.

24:56We'll use browser automation if we need to, if we come to something that can't be done with a Salesforce CLI, but to their credit, the Salesforce CLI is quite comprehensive. Pretty much anything you can do in the UI, you can do with a CLI command and the agents seem to already know how to use it. So can I do a$1 bet with you for when we're on this podcast in five years? So I think the majority of use for agents in five years is going to be the same interfaces that humans are using just because it's the long tail of integration has already been solved and like all the protections and all the sharing and everything else.