Fivetran's CPO: Closed Data Stacks Won't Survive the Agent Era  🔗 **Source:** [Fivetran's CPO: Closed Data Stacks Won't Survive the Agent Era](https://share.snipd.com/episode/4761b5fe-f8ae-43c2-ad6f-dfd5b7c0720f) ## Highlights - **Fivetran's Internal AI Experimentation And Coaching Program** - Fivetran encourages company-wide AI use and lets employees pick tools while providing coaching and measurement. - Anjan describes a transformation program that trains experts to help others and enforces evals to measure real productivity gains. Transcript: Anjan Kundavaram I mean, we've been a very AI first company for a while. George, our CEO, was like, hey, everybody should use AI all the time. In every meeting, any questions you don't answer, get it on AI. So it's like a very liberating thing for the company, not having to think a lot about budgets and so on and like really investing in AI, I would say. So we highly encourage folks to use AI. The stack looks very similar, like a agent harness. I mean, certainly talk, I can talk about the external point of view in a minute. The internal point of view, yeah. Think about using multiple coding agents. Think about the harnesses. Think about certainly governance and so on. Matt Burns One thing I talked to teams about is if they have a unified AI stack or if they allow their engineers to use what they find most helpful for them, what do you guys do? Anjan Kundavaram Yeah, we don't want to force our engineers to say, go use this model or that model or use this harness or that harness. And it's not just engineers. Engineers, sales, product managers, everyone really benefits. And the reason for that is the AI is sort of evolving so quickly. If you sort of say, look, this is the harness and this is the model, and like you pointed out, like yesterday, two things got, like image two came out, and that could happen, right? And you can get a new model from OpenAI, and all of a sudden it's a lot better. So you want the teams to be able to discover. So it's more guidelines. It's more like, hey, here's what a good stack looks like. You need to think about evals. Are you actually making progress? So one of the things you could kind of, you get this illusion of progress, and you do want to measure it. Hey, are you really getting productivity? Is it really writing good code? Are you actually driving more velocity? And is it higher quality things? And are you really, so you kind of want to spend energy on that. And then the tools, you kind of want to evolve. But I think certainly we're at the point where you can run multiple agents for many hours on a particular task. And that we think is there. ([Time 0:06:46](https://share.snipd.com/snip/fced3698-6622-473b-a527-0ff0781f9461)) - **Enable Team Experimentation With Guardrails** - Allow teams to experiment with multiple models and harnesses instead of enforcing a single stack. - Anjan says guidelines, evals, and measuring productivity ensure discovery without locking teams out of rapid model improvements. Transcript: Anjan Kundavaram So it's more guidelines. It's more like, hey, here's what a good stack looks like. You need to think about evals. Are you actually making progress? So one of the things you could kind of, you get this illusion of progress, and you do want to measure it. Hey, are you really getting productivity? Is it really writing good code? Are you actually driving more velocity? And is it higher quality things? And are you really, so you kind of want to spend energy on that. And then the tools, you kind of want to evolve. But I think certainly we're at the point where you can run multiple agents for many hours on a particular task. And that we think is there. ([Time 0:08:10](https://share.snipd.com/snip/e8a75568-e700-4726-a5ed-bd53c19badbf)) - **Agents Multiply Query Volume And Shift Cost Tradeoffs** - AI agents will massively increase query volume and change cost tradeoffs because they can run many more queries than humans. - Anjan Kundavaram warns closed data ecosystems force every query through expensive engines, likening it to "using a Lamborghini to mow the lawn." Transcript: Anjan Kundavaram A lot of customers aren't ready for this. Number one, there's going to be a cost explosion in how you run your, certainly we're in the database, there's a data query. So I think before you kind of have like an internal analytics team would run queries and it'd be like very curated. Now anybody can ask a question. So like the number of queries that you run went from could go 10x, 100x, I don't know. And I think it's, if you have a stack that is, let's say, built for the human era, either you're going to, as a deal leader, you're going to say, look, I won't be able to support you, and someone's Going to be really mad at them, or you're going to to pay a lot of money and i think sort of it worked in this like closed ecosystem is not going to work anymore like really open right and i Think you will get a lot of cost innovation uh sort of we kind of talk about this in the context of data lakes hey everyone should have an open data foundation. You should be able to talk to any compute engine. And some of them might be a lot cheaper for some of these tasks. And here's the interesting thing about an agent. An agent's not going to behave like us. It could be okay waiting for five minutes. You and I would be like, I need my answer now. And an agent could go spend more time if the agent thinks you're going to save 10x the cost. So you're going to get a fundamental different shift. So an open data foundation semantics are going to be very important. So we all kind of get that intuition. If you give it the right context, you're going to get a better answer. And it's a cheaper answer, too. Don't forget, these tokens are going to cost a lot, which I'm sure you've heard now, Matt. I have a line here about tokens yeah so so the the right context matters because if you if your information is in a lot of different systems you're not disciplined your context is not in One place you don't have a good lake house architecture one your ai returns are going to be poor two your costs are going to double like or maybe at 10x, because you're going to run a lot More queries, and you're going to be foolish about providing context. So it's kind of like a triple whammy. So I think the data foundation for AI is sort of emerging, and we're sort of five times, we're calling that the open data infrastructure. We really are emphasizing openness, because what we've seen in the industry is, early days, everyone's closing down. Everyone's like, it's my data. And that's like kind of silly. And that's going to backfire on those vendors who close. I think our approach is it should be open. And I think AI is going to drive more patterns. The cost economics for agents are going to be different. The cost economics for models we're seeing are exploding so you could make that much worse how do you talk to customers about the the explosion in tokens like do customers even know about Tokens yet i think the right sort of ai diffusion is mixed i would say like there's some companies that are early that are really it's still largely early days in the enterprise i think So they are seeing early signs of this, where costs going up. And one of the data leaders told me at a very large company, hey, our analytics budgets just like queries have gone up a lot. In fact, our own internal analytics leader was like, hey, hold on, I got to put controls. So we're like, no, no, don't put controls. Let's innovate, use an open data foundation, use an open data infrastructure, and like enable these things. So I think the companies now need, have sort of urgency to go change their data stacks, I would say, way more because of AI. And I think the natural intuition from data leaders would be, hey, I'm just going to put budgets. I'm just going to lock everything down. That's good, but then you're not going to get the productivity gains and you're not going to get the innovation gains. ([Time 0:08:58](https://share.snipd.com/snip/ad59fccb-c169-4c91-9a45-da1c00b061bd)) - **Make Your Data Foundation Open For Cost Optimization** - Build an open data foundation that lets agents choose the cheapest appropriate compute engine instead of routing all work through a single expensive platform. - Anjan recommends openness so agents can pick cheap engines for non-urgent queries and avoid constant high-cost paths. Transcript: Anjan Kundavaram You should be able to talk to any compute engine. And some of them might be a lot cheaper for some of these tasks. And here's the interesting thing about an agent. An agent's not going to behave like us. It could be okay waiting for five minutes. You and I would be like, I need my answer now. And an agent could go spend more time if the agent thinks you're going to save 10x the cost. So you're going to get a fundamental different shift. So an open data foundation semantics are going to be very important. So we all kind of get that intuition. If you give it the right context, you're going to get a better answer. And it's a cheaper answer, too. Don't forget, these tokens are going to cost a lot, which I'm sure you've heard now, Matt. I have a line here about tokens yeah so so the the right context matters because if you if your information is in a lot of different systems you're not disciplined your context is not in One place you don't have a good lake house architecture one your ai returns are going to be poor two your costs are going to double like or maybe at 10x, because you're going to run a lot More queries, and you're going to be foolish about providing context. So it's kind of like a triple whammy. So I think the data foundation for AI is sort of emerging, and we're sort of five times, we're calling that the open data infrastructure. We really are emphasizing openness, because what we've seen in the industry is, early days, everyone's closing down. Everyone's like, it's my data. And that's like kind of silly. And that's going to backfire on those vendors who close. I think our approach is it should be open. And I think AI is going to drive more patterns. The cost economics for agents are going to be different. ([Time 0:09:59](https://share.snipd.com/snip/34ea1914-58f4-407b-a359-bd21d6f7a830)) - **Don't Shut Down AI Use; Invest In Open Infrastructure** - Avoid knee-jerk lockdowns when analytics costs rise; invest in open infrastructure and semantics instead of strict access controls. - Kundavaram cautions that locking down will block the productivity and innovation gains from AI. Transcript: Anjan Kundavaram And one of the data leaders told me at a very large company, hey, our analytics budgets just like queries have gone up a lot. In fact, our own internal analytics leader was like, hey, hold on, I got to put controls. So we're like, no, no, don't put controls. Let's innovate, use an open data foundation, use an open data infrastructure, and like enable these things. So I think the companies now need, have sort of urgency to go change their data stacks, I would say, way more because of AI. And I think the natural intuition from data leaders would be, hey, I'm just going to put budgets. ([Time 0:12:00](https://share.snipd.com/snip/4210ecd5-e524-4cb2-9afc-01093cd22d51)) - **Centralize Structured And Unstructured Data For Agents** - Agents need both structured and unstructured data centrally available so they can fetch context efficiently. - Kundavaram says centralizing data and mapping semantics reduces token usage and avoids agents stitching dispersed datasets on the fly. Transcript: Anjan Kundavaram So there's a few dimensions, right? One, the data fundamentally is the same. How you access it is very different. Now with sort of AI, you want, like, it can look at unstructured data. So you kind of want to get structured and unstructured in one place. And FiveFriend, we've been doing well with that for a couple of years now. But once you've got your data in one place, you want it to be accessible for what we call any compute engine. So Google, BigQuery, Databricks. There's a lot of innovation happening in the open source ecosystem, DuckDBC to DB and so on. It's like for some queries that are very expensive, you want the AI to go and say, yeah, go use that. For some queries that are cheap, I might ask a question. It's not that important. I just want to know. I don't want to spend a lot of money on it. And you want the AI to go figure out a very cheap engine. Now, if you have a closed ecosystem, that won't be possible. It's kind of like using a Lamborghini to mow the lawn all the time. You kind of have to go use the most expensive option. So you want innovation. You will get locked out of innovation if you have a closed ecosystem. You absolutely need an open ecosystem. And you can go, okay, for that query, I'm going to use a more cheaper option or fairly low-cost option. And also, here's the other thing. If you keep it open, it's very likely the next analytical engine has not been invented yet. Of course, because the rate at which the AI is going, who knows? You might get a couple of engineers with a lot of insight drive this. We saw this with OpenClaw. Just one engineer did something amazing and ended up being a very, very famous project. So I think closed is a big no-no. You want it to be open. You want it to be scalable. And then the other thing I would emphasize is semantics. You want to teach your context. You want to, hey, here's what my tables are. Here's what it means. This is kind of like my sales table. You kind of need to sort of map all of that and sort of point it to the AI. So when the AI goes, ah, I get it. I know what that means. And the queries are a lot more intelligent and cheaper. ([Time 0:12:55](https://share.snipd.com/snip/7afa242a-30e8-4954-bceb-a3450b9e54de)) - **Semantics Make Agent Queries Accurate And Low Cost** - Strong semantics (clear table meanings and mappings) make agent queries smarter and cheaper by reducing redundant data fetches. - Kundavaram emphasizes teaching AI what tables mean so queries can be precise and cost-effective. Transcript: Anjan Kundavaram You want to teach your context. You want to, hey, here's what my tables are. Here's what it means. This is kind of like my sales table. You kind of need to sort of map all of that and sort of point it to the AI. So when the AI goes, ah, I get it. I know what that means. And the queries are a lot more intelligent and cheaper. ([Time 0:14:53](https://share.snipd.com/snip/9eddaa48-03f6-4447-b5fc-7fa671a39cd9)) - **Treat Agents With Enterprise Security Controls** - Apply the same enterprise security and access controls to agents as you do to queries and transactional data. - Kundavaram urges limiting agent access to only required data and building controls for production usage beyond demos. Transcript: Anjan Kundavaram Yeah, this is good. And fundamentally, they should. Right. They should demand like the same security guarantees that they've had running queries, the same security guarantees that they've had kind of creating transactional data. All of those should permeate to agents. There shouldn't be any other like agents shouldn't be able to be like, I'll go get all of math, data, no. Like only get the data that you really need to. And this is an area we need to do more work as an industry. And certainly in the data space, I think where it's like the cool demos, the cool demos need to go into production, access control, what kind of permissions you sort of need to. And here's the other thing. A lot of the modeling companies have focused from a more B2C point of view, but then they haven't really, as an industry, we haven't solved the how does it really work in an enterprise. Because you can't say like, oh, go, let's let someone go rogue and kind of go do it. You kind of have to put controls. Sure. And not like put your code away on GitHub somewhere and let the whole world see it, for example. ([Time 0:18:20](https://share.snipd.com/snip/12fe7835-36b3-4236-8b71-04e285ab4156)) - **Dashboards Evolve Into Disposable On Demand Views** - Dashboards will persist but become disposable and on-demand as agents create many more ephemeral views. - Kundavaram predicts gold-standard company dashboards will remain while users generate frequent ad-hoc dashboards themselves. Transcript: Anjan Kundavaram Think so. But I think you'll see a lot of disposable software. You'll see on-demand dashboards. I create dashboards all the time now. I create a lot more. Before, I would have to go haggle with my analytics leader. Hey, you need this dashboard. Like, tell me the business reason. Tell me how many people will use it. Like, show me the ROI. You'll probably get mad that I'm saying this. Matt Burns I'm with you. I make a dashboard a day at this point. Anjan Kundavaram Yeah, but now you get insights. And then I think you'll still have the category of dashboards, which are, look, these are gold standards for the company. Don't change them. Everyone needs to look at them. And then now you get insight. It's visually appealing insight. And you can go create them. So I do think they survive. But it'll sort of evolve into something different. ([Time 0:19:50](https://share.snipd.com/snip/3e16c83b-681e-44f2-9575-7d9fe1a9c487))