Snowflake VP of AI on Why Enterprises Hide Behind Governance to Avoid Real AI Transformation | Baris Gultekin | E296

Snowflake is the AI Data Cloud behind some of the world's largest enterprises — $4.68 billion in annual revenue, 29% year-over-year growth, and over 760 Forbes Global 2000 companies as customers. Baris Gultekin, VP of AI at Snowflake, leads the product efforts that sit at the center of how those enterprises actually operationalize AI. Before Snowflake, he co-founded Google Assistant and scaled it from 10 million to 500 million monthly users.

What you'll learn:

• Why our data isn't clean enough is a delay tactic — and the scoped approach to move past it
• What the semantic layer is and how it lets AI answer business questions accurately, not just fluently
• Why running AI next to data (instead of sending data to models) makes governance dramatically easier
• How Snowflake deployed AI internally: a CEO-level non-optional mandate combined with bottom-up access to their own Cortex coding agent
• Why context — not just data — is what agents need to operate reliably at enterprise scale

Key takeaways:

• Start with one scoped use case, build the semantic model around it, layer governance — don't wait for perfect data
• Context is a shared reality for agents: unified data + business semantics + codified workflows
• AI adoption compounds when leadership sets a hard mandate and simultaneously gives everyone a tool to experiment with

Credits:
Host: Carlos Gonzalez de Villaumbrosia
Guest: Baris Gultekin

Social Links:

• Find out more about Product School here
• Follow our Podcast on TikTok here
• Follow Product School on LinkedIn here

Transcript

Baris Gultekin | Snowflake 00:00:00

AI is democratizing access to that data. Traditionally, what has happened is you see something going wrong. You have to go ask a data scientist or an analyst to help you, and there's a long queue. You have to wait a couple of days before you get an answer. What AI allows is now you can essentially talk to all of your enterprise data in natural language.

Baris Gultekin | Snowflake 00:00:19

You can get insights within seconds. It empowers all of the business users to be very data savvy and drive their business with up-to-date information. What moved the needle for the whole company was top-down from our CEO — having a very clear mandate that this is not an optional thing. We need to have the whole company operating differently.

Baris Gultekin | Snowflake 00:00:40

We built our own coding agent called Cortex Code, that essentially allows anyone to use that data, ask questions of that data, and build agents and experiences with it. It wasn't a productivity gain only. It was essentially a series of products that we built to increase the value we're offering to our customers, and thereby drive more revenue.

Introduction

Carlos González de Villaumbrosia | Product School 00:01:02

Hey, this is Carlos, CEO at Product School and your host on The Product Podcast. Today's guest is Baris Gultekin, Vice President of AI at Snowflake. Snowflake is the AI Data Cloud — the platform where the world's largest enterprises bring all of their data together in one governed environment to run AI at scale.

Carlos González de Villaumbrosia | Product School 00:01:19

The company crossed $4.68 billion in annual revenue in fiscal year 2026, growing 29% year-over-year, with over 760 of the Forbes Global 2000 as customers. Baris joined Snowflake through the acquisition of Neeva, the AI-powered search startup he co-founded alongside Snowflake's current CEO.

Carlos González de Villaumbrosia | Product School 00:01:37

Before that, he spent over a decade at Google, where he co-founded Google Assistant, scaling it from ten million to five hundred million monthly users. In our conversation, we cover why there is no AI strategy without a data strategy, and what it actually takes to break down data silos before deploying agents.

Carlos González de Villaumbrosia | Product School 00:01:53

The semantic layer — how enterprises define business meaning on top of raw data so AI can answer questions accurately, not just fluently. How Snowflake runs AI next to data rather than sending data out to models, and why that architecture makes governance easier. The practical path to starting small: scoping a first AI use case, building a semantic model, and layering governance before scaling.

Carlos González de Villaumbrosia | Product School 00:02:14

And how Snowflake deployed AI internally — combining a top-down mandate from the CEO with bottom-up access to their own Cortex coding agent. Let's get into it. Welcome to The Product Podcast, Baris.

Baris Gultekin | Snowflake 00:02:26

Thank you. Thanks for having me.

What Is Snowflake?

Carlos González de Villaumbrosia | Product School 00:02:28

You are the VP of AI at Snowflake. Maybe we can start there. That's a really cool title. What does it mean?

Baris Gultekin | Snowflake 00:02:35

I'm a product manager. I lead the product efforts here at Snowflake, focused on AI. VP of AI is essentially leading the AI initiatives at Snowflake.

Carlos González de Villaumbrosia | Product School 00:02:49

Snowflake is obviously a big company — kind of the backbone for a lot of companies. Love for you to explain what it is in simple terms, and then we can go into what is new now that you're doing a huge push with AI.

Baris Gultekin | Snowflake 00:03:03

Certainly. So Snowflake — we call it the AI Data Cloud. Essentially, it is a data platform where companies can bring their data from all different stacks that they have, bring it in one place, govern it, and analyze it at large scale. And as part of that, we bring a lot of AI capabilities to run next to it so that you could use AI to analyze your data, talk to your data, and get more out of your data. The types of use cases you could imagine are anything from powering business intelligence tools, to writing machine learning models, to doing any type of AI extraction and data processing.

Carlos González de Villaumbrosia | Product School 00:03:51

In very simple terms, it seems like it's the data aggregator — there's data living across so many different systems. You connect them all as a single source of truth?

Baris Gultekin | Snowflake 00:04:01

That's right. You connect them all as a single source of truth. You govern it, you make sure anything that needs to be processed and cleaned up is done. It's a large-scale data processing platform.

Why "Dirty Data" Is an Excuse to Delay AI

Carlos González de Villaumbrosia | Product School 00:04:12

I want to have a conversation with you about data — it's top of mind for a lot of large enterprises undergoing AI transformation. The number one excuse or reason they give me when they are not able to get over the hump is that their data's not clean enough. And because the data's not clean enough, they need to wait to clean it before they do something. How do you get over that hump?

Baris Gultekin | Snowflake 00:04:37

We like saying there is no AI strategy without a data strategy — and that's exactly what you're highlighting. What that means is companies are thinking about all of their data assets, and AI feeds on data. So you need to break those data silos, have that data in one place, cross-join the data, and govern it so that only the people who should have access to certain parts of that data have access to it — all within a secure environment. That is what platforms like Snowflake offer. And only after you have that can you bring AI to run next to data.

Baris Gultekin | Snowflake 00:05:10

At Snowflake, we bring the platform so that you can bring various data sources together, create your data assets, and then run AI next to data — rather than the other way around. That makes using AI a lot simpler because you can then respect all of the underlying governance you've put on the AI systems. Any type of access control, any type of monitoring or evaluation becomes a lot easier because you're running AI next to data.

Carlos González de Villaumbrosia | Product School 00:05:51

That's a big effort for a lot of companies. If you have data across different platforms and they're all using their own structures — structured data, unstructured data, different formats — how do you go about integrating all of that and making sure it's in a format that is easy to interpret?

Structured vs. Unstructured Data at Enterprise Scale

Baris Gultekin | Snowflake 00:06:11

Snowflake's bread and butter so far has been structured data — essentially companies processing a lot of documents and data from other places, cleaning it up, preparing it to run their business, to do machine learning or business intelligence. From an AI perspective, what that means is to make the most of structured data, AI needs to write SQL queries. And traditionally it's been really difficult for AI to do this because a typical large enterprise has thousands and thousands of tables — the data is pretty broad and it's difficult to figure out what means what.

Baris Gultekin | Snowflake 00:06:57

What we do at Snowflake is help our customers create the semantic layer on top of that data — it allows them to define "here is how we calculate revenue, this is what this column means, and here is how you join this with this other information" — so that when you bring the business context on your data, AI can do a much better job querying that data and giving you accurate information.

Baris Gultekin | Snowflake 00:07:29

Then there are other solutions for unstructured data, and real-world scenarios need to do both. A lot of conversation about AI tends to be about unstructured data — rightfully so, because unstructured data is massive and we've just unlocked the value of that with AI. But at Snowflake, in any real-world scenario, you want to be able to do both: look at documents and glean information from them, as well as pull structured data to answer questions like "what's my revenue?" or "how am I doing for this product in this region?"

Democratizing Data Access With AI

Carlos González de Villaumbrosia | Product School 00:08:07

Another complaint for large companies trying to solve their AI transformation: access to information. Assuming we were able to integrate all the data across different formats and make it more accessible — it still seems like data is governed by the engineers or the data teams. You need to log into the tool, ask for a dashboard. How can you democratize that type of access so it's easier for all kinds of people to get to those insights faster?

Baris Gultekin | Snowflake 00:08:40

This is exactly what AI is doing — AI is democratizing access to that data. Traditionally, someone is running their business, looking at all of their KPIs, and they see something going wrong. They want to understand what's happening. They have to go ask a data scientist or analyst to help — but usually many people are asking those folks, and there's a long queue. You have to wait a couple of days before you get an answer.

Baris Gultekin | Snowflake 00:09:22

What AI allows is now you can essentially talk to all of your enterprise data in natural language. You can get insights within seconds. You can dive in and direct your analysis. It empowers all of the business users to be very data savvy and drive their business with up-to-date information, versus having to depend on many others to get insights. It's been very powerful. We've had many conversations with customers who are really excited to move to this world — where rather than seeing a dashboard and having three follow-up questions, you can just ask them and understand the why behind what a chart is showing.

The Context Layer: Beyond Raw Data

Carlos González de Villaumbrosia | Product School 00:10:00

If you think about interfaces — one option is to go to the tool, in this case Snowflake, and ask questions there. But we're also seeing it's possible to ask those questions within an LLM like Claude, or even within Slack. How are you thinking about that?

Baris Gultekin | Snowflake 00:10:14

We have deep relationships with all of the labs — Anthropic, OpenAI, and Google — where we bring their large language models to run inside the Snowflake security boundary. So there is that governance part of it. But when you ask a question to Claude outside of Snowflake, if that LLM does not have an understanding of your data assets, you won't get highly accurate information. Especially with structured data, there's only one correct answer. When you ask "what is my revenue in the US?" there's only one answer. To do that well, the LLM needs to understand how you calculate revenue, where all of your information is stored. That requires business context. Snowflake is in the business of creating that context for the LLM — and doing it in a governed and secure way so that you only give that information to the right people and it's highly accurate.

Carlos González de Villaumbrosia | Product School 00:11:17

On the context front — it's hard to make sure you're integrating not just with your data across tools, but also other types of sources, so people can use a shared repository where they can collaborate without worrying about whether the data is correct or up to date. How do you go about creating the right context beyond just the data?

Baris Gultekin | Snowflake 00:11:53

Context means a couple of different things for us. One — context in my mind is a shared reality for agents to operate on. If you want to have a shared reality, you first need to have broken down all the data silos so that the agent has governed access to all of your data. Second, you need to have the business semantics next to that data. It's not enough to say "here are all of my tables" — you need to understand which ones are important, what your metrics are, what the business cares about, how that is calculated. Doing that well requires a semantic layer. We allow our customers to build out that semantic layer and maintain it.

Baris Gultekin | Snowflake 00:12:53

A third pillar is codifying workflows, procedures, and how the company is run — that tends to be about operations, about how work gets done. That's another part of the context layer. So if you have that large context with all of your data, all of the meaning to that data, as well as all of the ways people are operating — the documents, the things that are important — that constitutes context.

Baris Gultekin | Snowflake 00:13:37

The challenges around context are: first, you need governance — you need to make sure that large amount of data is fully governed so that access controls are all in place. The second part is capturing it and constantly maintaining it, because this is a large amount of data and it changes all the time. We build agentic solutions to constantly monitor what is changing and what is important, so that agents have up-to-date information. Then the next layer is — if you ask an agent a question, how do you pick the right part of that context and give you an answer? The retrieval part is really important, and we focus a lot on that — for both structured and unstructured data, what subset of that context is most important to provide accurate information to the LLM to get accurate results back?

What Governance Actually Means

Carlos González de Villaumbrosia | Product School 00:14:10

Let's talk about governance. It's a very used word in the enterprise world, but it's kind of unclear what it really means sometimes. I hope it doesn't mean a 20-page dissertation on company policies. What is that balance between creating the right guardrails while allowing people to start getting value from their own systems?

Baris Gultekin | Snowflake 00:14:35

Governance for us is a series of rich capabilities for companies to create policies around how data can be used. For any type of data, you have role-based access controls — for this role, this is the data you should have access to, but this data you cannot access. Then there is a series of semantic guardrails — you should answer these types of questions, but not those types of questions. Then you have observability — what is the agent doing? Can I see what it's doing? Can I see all of the traces? Can I have tools to evaluate how well it's doing? And finally, there is cost governance — can I set and enforce limits for this organization or this person? It's a full spectrum: from access controls and policies, to observability and monitoring.

Carlos González de Villaumbrosia | Product School 00:15:47

The common pattern I've seen is that a lot of companies use the term governance — "we don't have the right governance," or "these new tools are going to break our governance," or "we don't have the right data" or "we don't have the right talent" — as a way not to get started. What are some of the mechanisms for companies that may never have perfect governance, context, or data to at least find quick wins and get started?

How to Start Small and Find Quick Wins

Baris Gultekin | Snowflake 00:16:29

Governance is incredibly important. A lot of our customers have been going through the transformation to the cloud over the last decade, and as part of that transformation, they're bringing more and more of their data assets online and creating structures of access controls. What we recommend is to start small — start with the core set of use cases that will move the needle in terms of ROI. That is critical. For those use cases, we help them build a semantic model: for this use case, what are the business semantics? Where is your data? Focus on that, and then build the governance layer for that scoped project — first the data layer, and then the agent layer on top of it.

Carlos González de Villaumbrosia | Product School 00:17:33

Snowflake started as a software business. What we're seeing now with a lot of AI transformations is that there's a service layer that helps those software businesses become more adopted. Some companies take the approach of providing the service themselves, others say "we need to bring an implementation company to help us connect the dots." How do you think about making sure that the teams you're supporting are able to get value from your software?

Baris Gultekin | Snowflake 00:18:06

Services is increasingly important in the AI space. We have our own forward-deployed engineering team who can help our customers with large-scale AI transformations. The reason it's important is things are changing very fast, and these transformations can be very meaningful for organizations. Being able to bring in someone who's very deeply connected to the product organization — our FDE team is part of our product organization — and can come and customize the last mile for the company's business accelerates how fast the transformation happens. We can bring the know-how and customize it exactly to how the company runs their business. This runs the spectrum for many companies: both getting their data ready, as well as building a high-quality agentic platform and solution on top of it.

Carlos González de Villaumbrosia | Product School 00:19:06

So if a large company wants to get started with AI transformation, you start with a small use case — function-specific use cases — and provide not only the software but the team that will create the right environment with the context, the governance, and the data. Then as you start scaling, there's a change management situation. Who is really running AI? Who's running the agents? How do you see companies scale AI — is there a centralized function with a Chief AI Officer, or are different people doing their own AI thing across functions?

How Snowflake Scaled AI Internally

Baris Gultekin | Snowflake 00:19:49

Let me give you an example from Snowflake. We believe we need to be at the forefront of that transformation ourselves, so that we have the know-how to bring into how we do work with our customers. At Snowflake, it worked both top-down and bottom-up. We first enabled our teams to use the series of tools we were building, as well as other tools — anything from coding agents to other AI tools.

Baris Gultekin | Snowflake 00:20:22

What moved the needle for the whole company was, first, top-down from our CEO — a very clear mandate that this is not optional. We need to have the whole company operating differently. Second, because we eat our own dog food, we have our data already on Snowflake in a governed environment. We built our own coding agent called Cortex Code that essentially allows anyone to use that data, ask questions of that data, and build agents and experiences with it. Once we launched this and rolled it out to the whole company, that is when we had tremendous AI adoption.

Baris Gultekin | Snowflake 00:21:09

On top of this, we tried to codify it top-down. First, bottom-up, everyone had access to this coding agent and they already started doing interesting things. We had account reps building anomaly detectors for their customers — "when something goes wrong, send me an email" — something they could build very quickly. Everyone started automating their workflows. We then decided to also help from a top-down perspective. We built out, for each function, a series of skills that can be deployed.

Baris Gultekin | Snowflake 00:21:45

These skills are very function-specific. For product management: writing PRDs, reading a code repo and getting information about it, building UI prototypes, and doing all of their routine day-to-day work. We did this for all functions — finance, marketing, HR — and rolled this out along with a series of profiles for these roles. Essentially, both a top-down mandate with a series of capabilities rolled out — all the data assets and connectors — and bottom-up, giving the whole company access to very powerful tools to go build. AI has transformed how we've done work, especially over the last three to four months.

Snowflake's Position in the New AI Competitive Landscape

Carlos González de Villaumbrosia | Product School 00:22:39

I hear you talking about creating skills, building your own coding agents. I traditionally put Snowflake in the category of data cloud, but now there are other LLMs that allow you to create skills and have a coding agent. How do you position yourself in this new competitive landscape?

Baris Gultekin | Snowflake 00:23:02

The way we position ourselves is: we are an AI Data Cloud. Any type of agent you're building that is connected to your company's data — any type of data agent — we want to be the platform to help you build that. We have our own coding agent focused specifically on the data persona: analytics, data engineering, building machine learning models, building data engineering pipelines, and querying data to glean insights. We'd like to be the best place to do that in a governed and high-quality way. We're not in the business of building general-purpose software.

Carlos González de Villaumbrosia | Product School 00:23:53

Right — it's not a general-purpose agent. It's a data-specific agent.

Baris Gultekin | Snowflake 00:23:56

Exactly.

Using AI as a Revenue Driver, Not Just a Productivity Tool

Carlos González de Villaumbrosia | Product School 00:24:00

The next big thing — as more companies adopt AI and become more productive with it, the narrative now is around using AI to generate revenue growth. Tell me more about that. Maybe we can use your own example: using AI not only as a productivity booster, but actually a revenue booster.

Baris Gultekin | Snowflake 00:24:25

It's been a big enabler for us. Snowflake is a consumption-based model — customers pay whenever they are processing data. With AI, there's more data to process, and there are more people who can get more insight from that data. With those two things, we built a series of capabilities where our customers can use AI to drive a lot more value from what they're doing with Snowflake. It wasn't a productivity gain only — it was essentially a series of products we built to increase the value we're offering to our customers and thereby drive more revenue.

Baris Gultekin | Snowflake 00:25:12

The examples are: a lot of our customers use us to analyze data, and we built AI directly into that analytics engine. If you're running a report and you want to say "I want to classify all of my data" or "I want to extract certain information from it" — now AI can help you do that with a lot of ease. That power we gave to customers became very popular, and it drove revenue for us. And then a lot of our customers are now trying to build data agents — providing those capabilities essentially allows them to get more from their data assets, which is highly valuable for our customers. Overall, making the platform more valuable because we're making it easy to use — and easy to use with even more data — has been very profitable for us.

Carlos González de Villaumbrosia | Product School 00:26:08

Baris, thank you for this masterclass on data for AI transformation. This is the classic and important topic that people briefly brush through but never really go in depth on — and I'm glad we had the time to do that.

Baris Gultekin | Snowflake 00:26:23

Of course. Thanks for having me.