For a weekend project, Snowflake CEO Sridhar Ramaswamy built himself another company. This one was fictional, with sales and marketing departments whose information overlapped. He wanted to work out how an AI agent should connect a campaign to the sales it generated. Finding the records was only the beginning. The software also needed to understand what that relationship meant to the people running the business.
At Snowflake, he already has software monitoring developments he might otherwise miss. Every 30 minutes, an automation scans Slack conversations for customers praising a feature or reporting a serious bug, then consolidates the findings in a single channel. Ramaswamy created the recurring job simply by describing what he wanted, without writing a single line of code.
That kind of delegation is central to Snowflake’s broader ambitions. Having built its reputation helping enterprises analyze their data, the company now wants its software to understand that information well enough to carry out tasks and become more deeply involved in everyday business decisions. Rivals including Databricks and Microsoft are pursuing similar opportunities, but Ramaswamy believes Snowflake’s established position in enterprise data gives it a stronger starting point.
“Snowflake established itself as, and remains, the best analytics data platform on the planet,” he tells me in an exclusive interview. “Our philosophy has always been to make the platform easy to use while ensuring that all data is connected and trusted. When I say connected data, I mean it should be infinitely shareable. An organization might have an enormous database, but an employee should be able to take one dataset and combine it with another, provided they have the appropriate permissions to do so. Making that possible requires getting governance right while maintaining the reliability that enterprises depend on.”
At its Expedition event, Snowflake introduced CoCo and CoWork Bots, always-on AI agents designed to plan and complete work in the background. CoCo helps technical teams write code and maintain data systems, while CoWork lets business employees ask questions about company information in plain language. The new Bots, currently in private preview, extend those capabilities to ongoing assignments, such as preparing briefings for customer meetings or monitoring data pipelines. They inherit the user’s access permissions and check in when a critical decision or final approval requires human input.
Snowflake Claims It Has Leapfrogged Competition
Ramaswamy laid out his case for Snowflake’s competitive position by first acknowledging a past misstep: the company had been slow to recognize and respond to the growing importance of machine learning.
“We did not innovate quickly enough in that area, while companies like Databricks did. Kudos to them. But we have since closed that gap, and when it comes to AI, I believe we have clearly leapfrogged the competition,” he says. “You can conduct your own in-depth research comparing how customers perceive CoCo and Genie Code, and I think you will find that CoCo stands right alongside it. CoCo is built on a modern agent harness, and the sophistication of the problems it can solve continues to genuinely surprise our customers.”
That confidence was evident in Ramaswamy’s recent public exchange with Databricks CEO Ali Ghodsi on X. On Oct. 28, responding to an essay promoting Databricks by Ben Horowitz, whose firm Andreessen Horowitz led the company’s first institutional funding round in 2013, Ramaswamy pressed the company for greater financial transparency. Databricks had reported an annualized revenue run rate exceeding $7 billion, with growth above 80%. Unlike publicly traded Snowflake, which reports quarterly financial results, privately held Databricks selectively discloses its performance metrics.
“does 80% yoy become 65% when you remove passthru inference?” Ramaswamy wrote. Reselling another provider’s AI model inference can inflate reported revenue while leaving the reseller with a relatively small share of the proceeds.
Rather than addressing those financial questions directly, Ghodsi turned the exchange into a sales pitch. “Since you’re not doing this AI stuff,” he wrote, suggesting that Snowflake consider purchasing its AI tokens through Databricks. He highlighted Unity Gateway’s spending controls, claiming that most of its customers had reduced their token spending by approximately 30% in recent months while increasing usage. Ghodsi went further, proposing that Snowflake integrate Databricks’ Omnigent with CoCo, effectively positioning Snowflake as a prospective customer in the very AI market where Ramaswamy claims a competitive advantage.
“generous. i was asking about the revenue, not contributing to it,” Ramaswamy replied. However, Ghodsi might be pitching controls Snowflake already sells. The platform’s own Cortex AI Gateway provides spending controls, with dynamic model routing currently in private preview. The platform that governs how much AI an enterprise consumes, and at what price, is the prize both companies are now fighting over.
The competition also extends to the data processing workloads traditionally associated with Databricks. Snowflake is increasingly targeting applications built around Apache Spark, with customer migrations providing evidence for its efficiency claims. Klarent, for example, claims migrating its heavy Spark workloads to Snowflake reduced total cost of ownership by 60% while cutting its longest data pipeline from more than 10 hours to less than three.
Ramaswamy also sees opportunities to serve customers who prefer AI agents from other providers. CoCo exposes itself to other agents through the Model Context Protocol and the Agent Client Protocol, and Ramaswamy says developers using tools such as Claude Code can invoke it as a separate agent to execute tasks on their behalf.
“It understands your permissions, which actions are permissible and which are not. Those are significant strengths when working with enterprise information,” he explained. “CoWork is somewhat different. With CoWork, we are taking certain risks because we are reaching out to users who would not ordinarily interact with Snowflake. We are doing so because we believe we have a unique ability to give those users access to important business information through conversational language.”
Microsoft’s workplace software and ServiceNow’s workflows offer competing starting points. Snowflake’s bet is that its authority over enterprise data can remain valuable even when another vendor supplies the interface.
Can Business Ontology Deliver an AI Edge?
Ramaswamy describes AI’s ability to make software development accessible to more employees as the “industrialization of intelligence.” As building applications becomes easier, he argues, their value increasingly depends on how well they understand the businesses they serve.
Snowflake wants to provide that understanding through its Business Ontology, which connects enterprise data to approved business definitions and establishes who is responsible for maintaining them. The capability, which the platform plans to introduce in public preview, allows companies to associate terms such as net revenue with the underlying data used to calculate them. Human-approved definitions take precedence over AI-generated interpretations, while an audit trail records changes over time.
The challenge becomes more complicated when different departments have legitimate reasons to interpret the same information differently. Ramaswamy acknowledges that Snowflake encounters this problem internally. His approach is to support department-specific ontologies, allowing companies to build agents around the information and definitions relevant to individual teams without forcing a single interpretation across the organization.
“We want to make the simple things really easy to do for a straightforward organization, while also making the complex scenarios possible for companies that need them. That goes to the heart of how we think these systems should be engineered,” he says.
The strategy brings Snowflake into more direct competition with Palantir, which has spent years using its ontology to connect enterprise data with business operations. Microsoft’s Fabric IQ Ontology and Databricks’ Genie Ontology pursue similar objectives. Even a comprehensive registry of approved definitions requires employees to recognize when business processes change and update those definitions accordingly.
How Snowflake Governs What AI Agents Can Do
The Model Context Protocol, or MCP, enables agents to connect with external tools, making permission controls increasingly important as those agents begin taking actions on behalf of employees. Snowflake’s internal restrictions illustrate how Ramaswamy distinguishes between allowing an agent to access information and giving it the authority to act on that information.
“We do not allow our Gmail MCP server to send emails externally. Why? Because we consider that a very high-risk action. We allow Gmail to create a draft, but a human must still review it before sending it outside the organization. Having those kinds of enterprise controls in place is what will make AI truly powerful,” Ramaswamy explained. “I personally run a lot of automated jobs, but we have safeguards such as agent policies and restricted session scopes that limit what an agent can do within a particular role. I do not want an agent accessing information from a confidential, CEO-only dataset and accidentally emailing it to someone.”
Likewise, the platform is also introducing Agent Abuse Detection, currently in private preview, to examine patterns of activity for potential misuse, including attempts by employees to circumvent approval processes. However, the company has disclosed limited information about the privacy safeguards governing those investigations or how accountability would be determined if an agent took an authorized action that ultimately harmed the business.
The Financial Reality Behind the AI Boom
Snowflake’s second-quarter fiscal 2027 product revenue reached $1.49 billion, up 37% year over year, marking its third consecutive quarter of accelerating growth. Ramaswamy says AI products accounted for approximately half of the acceleration in revenue growth. CoCo’s 9,100 accounts and CoWork’s 5,800 accounts indicate growing adoption, although account counts do not necessarily represent unique customers or reveal the extent of paid usage.
Asked how he evaluates lasting customer value, Ramaswamy emphasized measurable productivity gains, arguing that CoCo should allow data teams to complete their existing workloads five to 10 times faster. The evidence he offers for that figure is internal.
“We are promising an enormous return in terms of time savings and productivity gains, and we support that promise with data from our own teams, including my data team and our sales team, particularly when it comes to core productivity,” he says. “In fact, my own data team has become an ambassador for what a modern data organization can look like when equipped with powerful tools such as CoCo. We have been using these technologies across our own operations, giving us firsthand experience of the productivity improvements we are describing to customers.”
Delivering those gains, however, carries costs for Snowflake. Its non-GAAP product gross margin declined to 74.7% from 76.4% a year earlier, which the company attributed to a higher mix of fast-growing AI workloads that carry lower margins today. The expansion of always-on agents could intensify that pressure as customers consume more inference resources. Ramaswamy also pointed out Snowflake’s $9 billion in remaining performance obligations (RPO), representing contracted future revenue, as evidence of customer confidence. Although that figure increased 30% year over year, it declined from $9.77 billion at the end of January.
His advice to founders is to look beyond the software they build and identify what makes their businesses valuable to customers, even as the technology itself becomes increasingly easier to replicate.
“If the value you deliver to customers comes primarily from running a particular piece of software, you constantly have to consider the possibility that someone else could recreate that software and offer it at a lower price. That means you need to think beyond the software itself. What is the durable value that is uniquely yours? I think this is a mind-bending time for founders, but in some ways, it also represents a return to fundamentals,” Ramaswamy says. “Assume that anything written in software can be reproduced very quickly. Under those circumstances, do you still have enduring value to offer your customers?”
THE CONTROL LAYER VERDICT
Snowflake’s established role in managing critical enterprise data gives it a strong foundation for expanding into AI-driven workflows. But using that information to make decisions and execute tasks introduces responsibilities that go beyond traditional data management. If Ramaswamy can make AI autonomy dependable without imposing a comparable burden of human oversight, Snowflake could become more deeply embedded in how enterprises operate. Otherwise, businesses risk replacing manual work with the equally demanding task of supervising the agents meant to eliminate it.
