Snowflake Says Trust and Security Are Key to Unlocking the AI Agentic Enterprise

Uphadyay said the year has been focused on how to govern AI. He described two sides to the problem: ensuring that somebody else’s AI does not break into a company’s perimeter, and ensuring that a company’s own AI is well behaved, so it does not exfiltrate data and does not inadvertently go somewhere or break out of its perimeter. “When people trust us with their data, we need to make sure our platform is absolutely secure against agentic attacks,” he said.

Uphadyay echoed the Snowflake World Tour London keynote in describing what goes into a successful agentic enterprise. He likened the choice of AI model to the human brain and the agent being used to the body, all working on a foundation of data that needs to be “rock solid.”

Asked how Snowflake balances customer excitement around AI with the need to ensure trust and privacy, Uphadyay said customers are at different points on the AI adoption spectrum. The big difference, he said, is that security is “now the enabler, no longer the blocker.” Quoting former race driver Mario Andretti, Uphadyay said that although pressing the brakes is associated with going slow, doing so gives confidence and trust when turning, and without brakes a driver would be going really slowly anyway. He said security helps build that trust, and customers can leave it to Snowflake to provide the right guardrails and controls, with explanations of how they work.

This is largely done through the company’s Trust Center tool, which runs between the three layers Uphadyay mentioned, connects them and protects the customer while providing visibility into the system and what improvements might be needed. Snowflake has also been running for over 12 years, he added, a reputation that has provided customers with trust. Uphadyay said there is “a lot of innovation” happening in building and establishing new guardrails. He noted that when people work at a company for long periods, they acquire permissions for systems and services they may no longer need, which can be an issue when they run an agent as themselves. Ensuring and providing the right set of permissions for AI agents, which are exploratory in nature, can help prevent issues and keep data secure.

At World Tour London, Snowflake gave a demo of its Coco and CoWork platforms, and Uphadyay described how such services will benefit companies. “It can be a combination of stuff that a human being is going to do, and along with automation, what you might normally do that was repetitive, you can easily imagine an AI to automate than for you,” he said. He added that the work is no longer deterministic and that there are points where automation must be strung together and checked by a human. He compared it to a pyramid: a huge base of extremely mundane work, a layer above that which is less mundane, and at the top extremely mission-critical work that needs a human being for sure. “In this pyramid, you could say you want to automate, you want to move up, the more confidence you build, the higher you go — and that’s how enterprises build this confidence,” he said. Uphadyay closed with a direct assessment of the moment: “Things are here, things are real — it’s here and now.”