Snowflake continues to add AI, boost Cortex capabilities
Features that enable users to develop agents, coupled with the integration of agentic capabilities into the vendor's own platform, show the vendor's growth beyond data management.
SAN FRANCISCO -- As recently as two years ago, Snowflake's Cortex environment for AI development didn't exist, and the vendor's plans for generative AI were nascent.
OpenAI's November 2022 launch of ChatGPT, which represented a significant improvement in generative AI (GenAI) technology, sparked a surge of interest among enterprises in developing AI applications that could potentially make workers better informed and more efficient.
Tech giant Microsoft responded almost immediately, investing in OpenAI and unveiling its line of GenAI copilots. Databricks, Snowflake's longtime rival, similarly responded quickly, introducing its own large language model in March 2023 and acquiring MosaicML to form the foundation of an AI development environment for its customers.
Snowflake, which historically provided data management capabilities that enable users to easily query and analyze data without having to move it out of the vendor's environment, was slower to respond. Snowflake introduced Cortex in November 2023, but its features were far from being generally available.
However, when Sridhar Ramaswamy was named CEO in February 2024, replacing Frank Slootman, Snowflake's focus shifted. Providing users with the tools to develop AI applications with Cortex became Snowflake's priority, and since then, most of Snowflake's product development has focused on enabling customers to build and deploy AI applications.
Most significantly, Snowflake turned Cortex, which was largely in preview well into 2024, into a platform for AI application development.
In June 2024, Snowflake provided tools to enable users to quickly develop GenAI chatbots using their own data and unveiled Cortex Agents in preview in February to help customers develop agentic AI applications. Most recently, Snowflake introduced a conversational interface to simplify development and analysis Tuesday during Snowflake Summit, the vendor's annual user conference. It also rolled out an agent to assist data scientists by automating routine tasks.
In an interview during Summit, Jeff Hollan, Snowflake's director of product for Cortex AI agents and applications, discussed the company's focus on enabling AI development. In addition, he spoke about the agentic AI capabilities Snowflake is integrating throughout its own platform to speed and simplify applications, the vendor's competitive standing and its AI development roadmap.
Jeff Hollan
What was application development like before generative AI was available to assist developers and simplify processes?
Jeff Hollan: It wasn't terrible, but it is incredible how much easier it is with GenAI.
The classic flow was having a code editor on one side of the screen and Google Search on the other, and developers were constantly jumping between Google Search, Stack Overflow and the code editor. To write a very simple thing, I knew in code how to do it, but if I didn't remember the exact syntax, I had to go find some code sample to refresh my memory. Then I would write 10 lines of code, run it, find three different bugs and squash those three different bugs. And I would repeat and repeat and repeat.
It worked, but it meant that creating a simple website or dashboard would take two days.
With GenAI and all the assistance it provides, how long does that same application development process take?
Hollan: An hour.
It's because so many of those things go away. I don't have to go to Google to search for things. I just tell my agent or copilot what I want to build, and it will create 70% of it without me having to touch any code, and then any time I have to add my piece, instead of going to Google, I tell it what I need to do, and it takes care of it. The other exciting part is the debugging piece. If you still hit bugs, you tell the agent or copilot, and it diagnoses where the bug is coming from and fixes the line of code, rather than me having to hunt for it.
Snowflake recently introduced Cortex Agents to facilitate agentic AI application development. How does it differ from the Native Apps development framework introduced in 2022?
Hollan: I think of Cortex Agents as a pattern, while Native Apps is a specific offering, like an app in an app store.
Cortex Agents is a way to take powerful generative AI models and layer on reasoning and thinking. Two years ago, if you asked ChatGPT a complicated question, it would get very confused and hallucinate. If you ask an agent a complicated question, it breaks it down into smaller questions, answers those and then gives the big answer. It's remarkable to see the reasoning. The other thing Cortex Agents brings is access to different tools or data sets. It can look up data on its own or perform actions on its own, unsurprisingly with more interest in the data access piece.
By layering on reasoning and thinking, am I right to assume that reasoning comes from proprietary data for an enterprise?
Hollan: I love ChatGPT and use it multiple times a day with random questions, such as what to do when my son has an ear infection. I don't find as much value in ChatGPT doing my job, which isn't a ding on ChatGPT. It's that the questions I care about are things such as what products Snowflake users are using the most, and who the newest customers are that are using certain products. ChatGPT doesn't have access to that data.
With a Cortex Agent, customers can bring their Snowflake data to a model through an agentic pattern, and the model will be able to answer business-specific questions -- it will give the answer because it will look up the right data. Cortex Agents makes it super easy to build an agent that pulls in the right context from a business and joins that with powerful models.
What does Cortex Agents enable users to do that they couldn't before within Snowflake's AI application development environment?
Hollan: We were talking about AI a lot at Summit last year, but at that time it was up to the person building the application to know the right data source to ask questions. Agents just takes care of that for you. The agent can look at an enterprise's Snowflake data and choose the right data source.
Also, as recently as a few months ago, you could ask an application to show something like sales over the last six months, and it would provide a chart. What it could not do was, if you saw a spike in sales at a certain point, explain why sales spiked. With Agents, because there is that reasoning step, it can think through a bunch of possibilities, explore them and do deep research to explain the spike.
What was the impetus for developing Cortex Agents -- was it something customers were asking for as agentic AI became a trend, or was it something Snowflake decided to develop based on its own market observations?
We have had customers for a year and a half who would say they've heard a lot around agents and want to know Snowflake's story there. But that alone is not why we started to do agents. The deeper question is why do customers care about agents, and the signal we hear very strongly is the business stakeholders need to get access to data much faster.
Jeff HollanDirector of product for Cortex AI agents and applications, Snowflake
Hollan: One thing that's top of mind is pulling apart the hype and the meat we need to make sure we're diving into.
We have had customers for a year and a half who would say they've heard a lot around agents and want to know Snowflake's story there. But that alone is not why we started to do agents. The deeper question is why do customers care about agents, and the signal we hear very strongly is the business stakeholders need to get access to data much faster. They're either developers who have an endless list of dashboards that they're that they're being asked to build, or they're analysts fed up with too many dashboards.
Customers asked us to help democratize all of the nuggets, truths and insights hiding within their Snowflake account. That's the signal we heard, and we saw that agents could be an effective tool to help them. Then, we worked with customers to find out their most pressing questions, and the result was the Cortex Agents announcement in February and Snowflake Intelligence this week.
What has to happen to move Cortex Agents out of preview and make the service generally available?
Hollan: What we've been learning through preview, and what we've been continuing to iterate on, is that people want to ask very complex questions. Unsurprisingly, when they ask complex questions, they really want to make sure that the quality is right. Everyone's biggest fear is giving their CEO access to an agent and having their CEO ask a question and get the wrong answer.
We've been learning that we really need to beef up our reasoning, which we've now gotten to. Most of the engineering work is done, and we have to wait for it to bake.
Do you have a timeline for when that might happen?
Hollan: It will be this year, for sure, and I would guess late summer. But no commitment there.
Snowflake just unveiled Data Science Agent as part of its platform to make data scientists more efficient -- beyond providing tools for customers to develop AI agents with Cortex, does Snowflake plan to develop more of its own agents?
Hollan: There's a huge amount of opportunity there. Data Science Agent is a big one, but there are lots of agents in the works, everything from helping users understand their spending on Snowflake to stuff around development, whether building a machine learning model or creating a notebook or a data engineering pipeline. Our vision is that whatever your task is, we'll have specialized agents.
Now that we have the Cortex Agent platform and pieces like Data Science Agent, teams are asking how to get more of these things lit up. By the end of this year, you will definitely see more agents popping up in the Snowflake experience to help users be more productive.
One common critique of Snowflake is that it got off to a slow start and needs to catch up to its competition. With the AI era still so nascent, does it matter if one vendor is a little behind at this point?
Hollan: There are different dimensions to the AI and machine learning pieces.
I think back to the first two or three years of smartphones. If you were an iPhone user or a non-iPhone user, there was a huge difference. Now, an iPhone and an Android are basically doing the same thing. Everything has caught up. There are some degrees of that happening [with AI], although I think there are some ways where Snowflake is quite cutting edge. Where I would say Snowflake is in the lead is in the aspect of [AI inference], which is making sure that when you ask a very specific question, you get the right data for that question. Getting that highly accurate is very difficult.
Am I anxious about whether Snowflake is the greatest to build custom LLMs? We want to be good enough. But do we want to be the best at surfacing the right data for an agentic application? One hundred percent.
Looking ahead, what is Snowflake's roadmap for Cortex and adding further tools to enable AI development?
Hollan: For AI development, interoperability is a big focus. I'm super excited to see Model Context Protocol and Agent2Agent Protocol. I'm shocked at how quickly we've been able to reach some consensus on industry standards, and we absolutely want to do everything we can to participate and be good citizens, so making sure interoperability is easy is a big focus.
And then, I'm excited because in three months, there's some new industry-leading frontier model, and the reasoning gets better and better. The ability for Snowflake Intelligence to do more analysis, to help build dashboards and applications dynamically, to share those insights and answer more complex questions is just going to get more advanced. It's continuing that thread, but the pace of innovation is exponential. It's fun, but it's pretty crazy.
Editor's note: This interview has been edited for clarity and conciseness.
Eric Avidon is a senior news writer for Informa TechTarget and a journalist with more than 25 years of experience. He covers analytics and data management.