#327 Baris Gultekin: The Next Phase of AI - Agents That Understand Your Company's Data

0:00Can you walk me through what it's been like? I mean, the last two or three years have just been crazy and it doesn't seem to be letting up. And how Snowflake AI has evolved.

0:12Baris Gultekin:AI is incredibly exciting. So we've been building a lot of product over the last two, two and a half years. anywhere from building AI to run right next to data to just being able to kind of talk to your data in natural language and democratize that access and glean a lot more insight from data. So that's the high-level goal of the set of products that we're building. And that's not great. Most AI is just speech-to-text plus a language model. It's full for reading transcripts, not understanding conversations. Velma from Modulate, an AI built on ensemble listening model architecture, specializes in audio analysis.

0:53It orchestrates hundreds of smaller submodels purpose-built to understand the nuances of voice like tone, timing, stress, and intent. Perfect for fraud defense, deep fake detection, agent attrition prevention, or customer service moderation. Check out the live Velma preview at preview.modulate.ai. That's preview.modulate.ai. To see how the model breaks down audio, providing timestamped, explainable signals. Stop transcribing, start listening with Modulate.ai.

2:04Baris Gultekin:AI was super hot. I love the space. I've been in the space for a long time. And at Snowflake, my job is to build out Snowflake's AI strategy, as well as a series of products that we're building for our customers. AI is incredibly exciting. So we've been building a lot of product over the last two to a half years, anywhere from building AI to run right next to data to just being able to kind of talk to your data in natural language and democratize that access and glean a lot more insight from data. So that's the high level goal of the set of products that we're building at Snowflake. Yeah. And I have to apologize.

2:38I said Boris. I'm sure you get that along a lot. It's pronounced Barish?

2:43Baris Gultekin:It's Barish, yeah. Barish. Okay. And can you walk me through what it's been like? I mean, the last two or three years have just been crazy and it doesn't seem to be letting up. Now Snowflake AI has evolved. Yeah, certainly. I mean, the journey started about two and a half years ago. The feedback we got from our customers is AI is incredibly transformative. But of course, governance is also equally important. And for Snowflake, data governance security is of utmost importance for our customers. So the request was to run AI next to data versus trying to move data to where the AI is, which creates a series of governance and security challenges.

3:26Baris Gultekin:So that's what we set out to do. What that means is we're bringing in all the large language models to run within the Snowflake security boundary, OpenAI, Anthropic, Gemini, Meta, and so forth. And that brings a lot of governance benefits. All of the security and governance that our customers put on their data automatically gets respected by the AI solutions. Then on top of this, we have kind of two flavors of products, if you will. One is Snowflake being an analytics engine. We're bringing AI to run inside that analytics engine. So that means any of our customers who want to analyze large amounts of data can now use AI to do that for classification, extraction of data, summarization of large amounts of data that they have on Snowflake.

4:09Baris Gultekin:Customer calls, feedback messages, a lot of text and images and video that they have on the platform can now be processed with AI very easily at large scale in batch. Then we have a series of products that are what we call data agents. So making it easy for customers to talk to their data using natural language. And when you said you bring OpenAI or Anthropic into Snowflake's cloud or yeah, how do you do that? Because OpenAI is, they're not going to give you the model. So, yeah. So we have, again, deep partnerships with the model providers. And what we do is, so Snowflake runs on top of all three cloud providers, Google, AWS, and Azure.

4:59Baris Gultekin:So we are essentially working with the likes of Anthropic and OpenAI to use their models on those cloud providers inside the Snowflake security boundary with the data governance requirements that our customers expect. So this means there is an element of privacy and data residency guarantees that we give for our customers. So you're not sending data out to the model providers or through an API or something like that? Essentially, only Snowflake has access to the data. And of course, the models are processing the data, but no data gets stored or no data can be used for training purposes and so forth.

5:44Yeah. And what does that philosophy of bringing AI to the data mean? and why is it important for enterprise? Is it purely security or is it also latency?

5:56Baris Gultekin:Yeah, so security is important for regulatory purposes. Data residency requirements need to be respected. The best example is our customers are putting a lot of governance at the full stack. From data perspective, best example is we built our own sales agent, for instance, inside the company. All of our sellers are using an agent. And when one seller asks about their customers and their customer spend, it should be different from the other seller asking it. Setting that level of granular access control so that by definition, by design, it's all respected, makes things both governed as well as easy to build with governance in mind.

6:33Baris Gultekin:So that's the philosophy. In addition to, of course, the security benefits and the fact that you don't have to replicate both your data and your governance in multiple places and manage them in multiple places. Yeah. And can you talk about agents? There's a wave of agentic tools that are coming out, tools and platforms. And Snowflake is also in that space at a high level. Can you talk about how Snowflake approaches agentic AI and how you differentiate yourself in the market? The way we think about it is ultimately AI solutions are as good as the data that they have access to, as good as the context that they have access to.

7:13Baris Gultekin:So we've focused a lot on having very high quality data agents by focusing on the retrieval quality of both structured data. So text to SQL generation, being able to talk to structured data, as well as just documents and unstructured data. So our agents are built on top of a product that we call Snowflake Intelligence, where you can just, in natural language, ask questions of all your data, be it structured or unstructured, and get high-quality answers, results, charts, and so forth. What that differentiates Snowflake from the others are, one, it turns out it's actually very difficult to have very high-quality structured data retrieval.

7:53Baris Gultekin:So we excel at being able to answer relatively difficult questions, business intelligence questions, being able to understand business trends, why things are going the way they are. So essentially bringing insights in a very approachable way with very high quality. And of course, then also this is all about our customers' governed data. So being able to build these agents with high quality very easily with governance at the core of it is something that differentiates us as well. And these products are resonating in the market. We have Snowflake Intelligence is the fastest growing product that Snowflake ever had.

8:29Baris Gultekin:Thousands of customers are using it. There's a lot of demand in the market for building high quality data agents. Yeah. So Snowflake Intelligence is the interface for analyzing the data or talking to the data in natural language and getting answers back. Or is it for building agents and other software to work with the data? So Snowflake Intelligence is the interface for business users to interact with data and ask questions. And then we provide a series of tools for builders to build these agents and then deploy them for their organizations. I see. Yeah. Yeah. And Snowflake has partnerships, as you said, with OpenAI, Anthropic, Meta, and others.

9:17How do you think about building products as opposed to partnering with these?

9:22Baris Gultekin:The way I think about this is, first of all, it's been incredibly exciting to see the momentum and the progress with the Premier Labs. The models are continuously getting better. For our customers, they want model choice, and we'd like to offer that choice to our customers. And the way our customers think about it is, in some cases, each customer has approved certain models to be used for their organizations for a variety of reasons. Sometimes it is a customer is in one cloud, and they want to use the model available in that cloud. Sometimes it's for different reasons. So we provide all of those options for our customers to choose from.

10:01Baris Gultekin:The way we think about product development is we take these models and then we are building the next layer up. We have the retrieval layer, we have the data layer, and then the application layer on top of it, be it analytics, integrating it into our SQL engine or products like Snowflake Intelligence for natural language interfaces. We do build our own focused task-specific models as well. Those are for focused use cases, such as extraction, which is really important for our customers, or text-to-SQL generation, which is, again, really important for our customers. So certain aspects of the product from the modeling perspective are things that we build, but then we rely heavily on general-purpose models that are the frontier models, and we have our partnerships with the model providers for those.

10:45Yeah. I've been talking to people about post-transformer models. How wide a range of models do you offer? I mean, are you pretty much focused on the top model makers or do you offer other architectures?

11:01Baris Gultekin:Yeah, so we are predominantly offering the top models from the model providers. We are looking at other architectures as well. We have customers' partnerships, for instance, with companies like Kumo, who are experimenting with more kind of tabular foundation models. And so there's other areas that we're looking at as well. Yeah. Do you see, you work with customers, obviously, a lot of Snowflake customers. Do you see patterns in the teams that are really getting value from AI quickly? I do. In most cases, actually, it is a culture shift. and usually culture shifts happen when there is a top-down mandate as well as top-down depth of understanding of transformative nature.

11:51Baris Gultekin:So we have a lot more of a kind of attraction when that mandate is coming top-down, when that depth of understanding is coming from top-down. On top of this, coding agents have transformed how products are being built. So we're seeing that the more of our customers adopt and embrace these coding agents, the faster they're able to kind of build and deploy these AI solutions with us. Yeah, yeah, that's interesting. And agents in general are making big waves. Although I keep hearing from enterprises that they're skeptical because of the non-deterministic quality of agents and they're afraid to let them make critical business decisions.

12:40Do you see that changing and how can organizations build AI systems that are trustworthy enough for that?

12:50Baris Gultekin:I mean, I think trust is at the core of everything we do. If we don't help our customers build AI solutions that we trust, it will not go anywhere. So that is a very core focus of the way we operate. And trust has different layers. I talked about trust at the governance layer, making sure that the data does not leak unintentionally to someone who should not be seeing that data, for instance. The trust also, there's a security element. There's a guardrails element to make sure that AI doesn't respond. AI only responds in the tone and to the questions that you want it to respond. Then there's trust at the quality layer.

13:27Baris Gultekin:I want to only have high quality answers. This is especially true if I'm asking factual data questions. If I say, what was my revenue in the last month? There's only one answer to that question that is right. So getting high quality is incredibly important. The way we help our customers get there is first from a design principle perspective, there is governance at the full stack that we give a lot of control and ease of creation of these agents. Then we built our agent observability and evaluations products directly in the agent creation flows. So our customers can have a lot of control on how they want the agent to respond and what good answers are.

14:08Baris Gultekin:And then they can also see how well the agent is answering and then they can tune that relatively easily. Yeah. With agents, is this in effect retrieval, augmentation, what is it, generation issue? I mean, where does the problem of reliability come in? is it's not in finding the right data. And presumably then the models are using that data to formulate an answer and there shouldn't be any hallucinations there. But then when you're taking action, that's where people get spooked. Yeah. I mean, even in the retrieval step, there is a ton of things that can go wrong if it's not set up properly. So what we see, for instance, is especially in structured data, understand the semantics of the business, understanding the semantics of the data system becomes incredibly important.

15:05Baris Gultekin:Semantic models have become a really core part of how our customers are developing agents. So we help them build these semantic models, optimize them, making sure that the semantic model is highly accurate and up to date. So that becomes a critical part of that kind of structured data retrieval, if you will, the text to SQL generation. But then on top of that, evaluations are quite critical, helping our customers build evaluation data sets, evaluate the responses. So that is another core part. And we want agents to also continuously learn and get feedback. So being able to get the feedback from usage, passing it on to the developers who are building it, and making it very easy for such developers to quickly optimize the performance of these agents is part of the development.

15:50Baris Gultekin:And increasingly more and more of the agent memory. So that is important. Agents are automatically learning and then correcting themselves as well so that they keep getting better as they get used. Yeah. And these are agents being built on the Snowplake platform, is that right? That's right. So how do you manage those things? I mean, does it depend on the underlying model or is there architecture at this Gentec layer that you use to manage these? So our customers can choose for each of the agents that they're creating, which model they'd like to use. Or they can just use the default model that's available to them.

16:31Baris Gultekin:And then each agent has its own set of orchestration capabilities. So agents have a series of data sources that they're given access to, series of tools that they're given access to, to either take action or to do things like searches and so forth. And all of those are part of the platform. So that's kind of how our customers are building these agents. One thing that also customers want to make sure happen is they want interoperability, right? When I'm building an agent here, I want to make sure that this agent can also talk to this other solution that I have elsewhere. So we support open source interoperable platforms like protocols like MCP and soon agent-to-agent communication so that our customers can build agents in one place and have these agents talk to other solutions.

17:16And talk to other agents outside the Snowflake universe, yeah. And there's a feeling that LLMs in particular are getting commoditized. So where do you see durable competitive advantage coming from?

17:34Baris Gultekin:I mean, for our customers, ultimately, it's all about data is their core asset, their data, their relationships with their customers. And LLMs are incredibly capable tools that you could bring on top of all of those assets to unleash new capabilities, to unleash a lot more value out of them. And this is exactly what we're seeing. We're seeing significant acceleration in product development by our customers. And also, our customers are able to get a lot more insight from a lot more data to then use for either decision-making or building new products. Yeah. I hear that, I mean, maybe I'm a little behind the news, but that a lot of the experimentation with agents through 2025 have not led to production models or production systems.

18:27Do you see that this agentic AI is really penetrating the enterprise, or is it still relatively simple tasks? Because there's a lot of talk about making business critical decisions, but how much do you see that agents are really being deployed to do that?

18:49Baris Gultekin:Yeah, I'm absolutely seeing that we've crossed the threshold. And now we're in the scaling phase. We have a lot of customers who are already in production and they're getting already a lot of value. And the value starts with productivity gains. I was just on a call with a customer. They were saying, hey, we're saving 2 ,000 hours of time because we're able to automate a series of processes, especially in this case in their call centers and call analytics. And that starts from productivity gains, but then also goes into kind of more and more automation and more capabilities being developed. So overall, I think we're now at a point where LLMs have gotten very capable.

19:31Baris Gultekin:They are now able to execute on longer and longer running tasks. They're able to be a lot more accurate. So many tasks are now possible to automate. But we're also seeing, especially coming from a data perspective, when you just democratize the access to the data, not just automation, just democratizing access to the data, many customers are able to make a lot more data-informed decisions a lot faster. And that results in significant changes in the way our customers are operating. Yeah. And who is making those data-informed decisions? Are you talking at all levels of the organization, even at the C-suite?

20:12Baris Gultekin:Certainly. There are many customers are building products for their whole organization. This is starting with the sales force, being able to create a series of agents for the sales organization, the marketing organization, the finance organization, and the C-suite as well. And then Snowflake also, we eat our own dog food, right? So we have Snow on Snow projects where we're using our own products to develop and deploy AI across the organization. And from our CEO on to everyone in the organization is using such agents across the board. Do you think this is going to not only increase productivity, but accelerate business generally in the way I tell my kids?

20:52I remember typing on carbon paper if you needed more than one copy of a letter. And you look back, it's just amazing that any business got done, given the speed that things get done now. But with agents, do you see just the pace of business increasing?

21:13Baris Gultekin:I live it day to day and absolutely see it everywhere. I think the pace of innovation is increasing. The pace of product development is increasing. The pace of work getting done is increasing because of AI. And we see it across the board, right? The simple examples are very quickly being able to write custom emails for thousands of people that are just very clear. That's a simple example. But then once you start thinking about automating processes that were all kind of required a lot of manual approvals, manual processes, you're now shrinking something that used to take 30 days to hours. It makes a big difference from a productivity perspective.

21:52Yeah. So as agents take on more decision-making responsibilities, I mean, we talked about reliability, but how do organizations, how should organizations think about accountability and risk? Does that lie with whoever is managing the department where the agent is working? Is it with the IT department? Is it with the model providers if something goes wrong?

22:20Baris Gultekin:Yeah, I mean, ultimately it is, it lies with the owners of these agents. Agents are becoming more and more capable. And then there's this long stack of technology and people behind all of this. The way I'm seeing agents deployed is there's usually an owner at the organization who is describing, here are use cases we'd like to develop. These are really important. And then you go down that list. And then there's a rollout process. You develop the agents. There's a POC. There's a small deployment. pilot, and then broad deployment. And all of those are built so that you're kind of both improving the quality of the agent, you're putting the guardrails in place, and you're measuring the productivity gains, the ROI on this agent.

23:02Baris Gultekin:And it's no different than kind of software development, really, but this is incredibly powerful set of tools and capabilities. Yeah. And what does accountability mean when agents are making decisions without human overs? because there's the human in the loop where the agent tees up a decision and the human has to click proceed or but then there are other decisions and increasingly decisions as reliability problem issues go away that agents are making decisions really unseen from humans deep in the computer systems of an organization how do you think about account accountability there Right now, what is happening is, especially in cases where there's an automation, a workflow being set up, the workflows are kind of well established.

23:55Baris Gultekin:There's a series of steps. And anytime there is critical decision making that contains risks, there is usually a human in the loop that can take a look at it and then control it. Over time, as trust on these systems increases, more and more of these kind of human in the loop and guardrails will start getting automated. But ultimately, it will get automated because the AI is doing exactly what it's told. The instruction following capabilities of these models is increasing day to day. Accountability lies in capturing these business processes and then documenting them. And AI's job is to follow to the T exactly what it's asked to do.