Why Snowflake is no longer just a data warehouse | Equity Podcast

Rebecca, great to be here. Thank you for having me.

Well, I'm an academic by training. Worked in search labs in places like the labs for a while before going to Google. Little did I know that I was joining one of the most amazing businesses of all time, which was the search ads business at Google. I spent over 15 years there eventually running the ads and commerce team at Google, helping them get to 100 plus billion dollars in revenue, which was not something honestly that any of us expected to do. And then of course, Neva and Snowflake after that. I'm blessed to have had so many different opportunities.

Yeah, for most of us AI has meant something like ChatGPT. You could ask good questions. They sometimes be right, sometimes be amazing, sometimes be creative, all kind of hit and miss stuff that doesn't quite work in the context of an enterprise. But if you have questions about revenue, it better be super super accurate. And we could tell even last year that coding agents were going to be a big deal in terms of the speed at which people that only modestly knew coding, say me, could actually get stuff done.

What we're seeing this year though is these techniques coming together. Agentic AI being used more and more to create complicated applications. For example, our support team actually uses a coding agent with aspects of chat, aspects of tools to pretty much drive how they solve problems. Similarly, enterprises use products like Snowflake Intelligence to ask pretty sophisticated questions about how their business is doing. And built on a solid data foundation, these agentic AI products are becoming incredibly useful. Just this morning, I wanted to know how our quarter was doing. But then I wanted to know how our different components, our theaters as they were, the US, EMEA, APJ were contributing to that overall status. Previously, it used to be an analysis that someone had to run for me. This in natural language, I could not only ask the question, but produce something that looks like a dashboard that could be shared internally. That's the sophistication, which is take on governed data sources, things that are secure and auditable and trusted, which Snowflake has always pioneered. Put this layer of agentic AI on top of it. It just makes that data so much more usable. And obviously next up is what kind of actions can you directly take from these agentic applications that can now take actions in the real world. That's where we see that inexorable push towards what is AI doing in the enterprise.

Cortex Code was born of the belief that agentic was going to make it much easier to use on products like Snowflake. A lot of using Snowflake is setting up configuration, connecting to other systems, making sure that data is in good shape, writing code. It's a pretty complex job. It's often tedious, easy to make colossal mistakes, and so making sure that you do it in a safe way is always a concern. Cortex Code was born of the belief that agentic AI had the power to automate much of that. Putting the humans in charge of subjective questions, questions of taste and quality rather than the mechanics of getting work done. It's had a huge impact. Whether it's customers like Whoop that were early adopters or partners like Evolve that used it to leverage how quickly they could get work done with Snowflake. Cortex Code has had a huge impact in just speed of adoption of various features that are in Snowflake.

Now here's where something funny happened at Snowflake. We made Cortex Code available to everyone at Snowflake, not just the engineers, everyone. And what we saw was non-technical users, my sales executives started writing applications, started using Agentic AI to ask questions. That's when it clicked in my head and in a bunch of other people's heads in Snowflake. A coding agent is not just about coding. It's an abstraction agent. That's really what Snowwork is about. Use the power of agentic AI to make data to make actions much simpler so that a CFO can look at travel and expense data within Snowflake. Ask questions in natural language. Take follow-up actions also in natural language all within the safe, governed, auditable environment that Snowflake is. So, it's actually a closer progression than you would initially have thought.

Actually for business users. Yeah. Yeah. That's the magic. It's my HR team that's writing and saying, 'Oh gosh, processing resumes now is so much easier. Being able to figure out who is doing a great job hiring is so much easier.' Obviously, the finance team is using it to get better insight into the data. That spectrum has basically become everyone within an enterprise.

This is a great question. We start on a solid governed foundation. Whenever we create data agents, for example, we help people go through an eval process. That's just a fancy way of saying you're going to ask a bunch of questions, make sure that the answers are right, and every time you make a change, you want to make sure that you didn't break things that were working before. It's what you would think of as standard software engineering. And so whenever capabilities in something like a Snowwork are released, we have them go through the same kind of evaluation process where you ask a series of questions. You make sure that the answers that the coding agent is generating are indeed correct. This is a well understood science. It just has to be applied diligently every time you add new features to something like a Snowwork.

This is rapidly evolving and I think it depends on the progression of capabilities. More and more folks now are comfortable with the idea of a data agent for example built in Snowflake Intelligence asking and answering business questions. We have customers like Whoop that I talked about earlier or TS Imagine that have taken the data that they have in Snowflake and have created some pretty amazing experiences for areas like support and things like autonomous agents having agents take action are slow progressions. It's not, you know, you just decide that you're going to turn everything off. I'll give you our own example. Our SRE team, their site reliability engineers that help us keep Snowflake up. They started using Cortex Code to help them debug problems faster to get to the bottom of why is there a customer issue that say TechCrunch is experiencing with Snowflake. And what they found was that their time to problem detection was down by orders of magnitude. And now they're driving an autonomous version where they're looking at a large volume of incoming requests and saying which are problems that they can reliably answer. And when you can do that, you can also take actions a little bit more autonomously. And you also want to make sure that these are not like big actions, irreversible actions, things like that. So there is very much that element of figuring out how good is a product and how believable are the actions and then you put them in front of a human make sure that it's working really well and when you're very comfortable with how right it is pretty much all the time that's when you say that category of problems you're dealing with can be dealt with autonomously by an agent. It is very much a science. It's not yesterday no autonomous and today everything is autonomous.

It depends on the product. It depends on the area. With Snowflake Intelligence for example, once we have one what we call a use case, which is a pitch for a customer to implement something, most of them make it out to production. But the time that they take to make it out to production can vary anywhere from 4 weeks to 6 months depending on the complexity of the problem. Increasingly, we try to get our customers to try small things and launch them before trying to do bigger things.

The number one reason is going to be I don't know how to ask questions and get answers reliably from a product like I experienced this firsthand even within Snowflake. We have access to a sales agent that's been rolled out to 5,000 of our go to market team used every single day. Now, Snowflake is 10,000 people. 5,000 people is a lot. You can agree that that is scaled usage. But on the other hand, we're rolling out a new agent, for example, with all of our financial data. I was telling someone yesterday that I still don't have a good feel for how to query it, how to talk to it effectively because I don't quite understand the data model effectively enough. There's a lot of complexity in finance data. GAAP revenue is different from raw consumption data. And so you have to know a lot of nuance about how that works before you can feel confident about it. That is an example of something that has to be iterated upon more before it's released to all of the execs at Snowflake itself for example. And so it really depends on how do you create products that your customers your users will relate to and reliably get value from it.