Snowflake CIO says AI agents need a new security layer

Snowflake CIO Mike Blandina said companies need to control data, security, costs and AI agents working across multiple systems. He spoke at Snowflake World Tour in Stockholm before more than 2000 IT leaders.

According to Blandina, model choice is less important than the wider infrastructure around AI. Snowflake’s approach has four parts: data and context, model selection, applications, and a control layer. Different business units may hold different versions of the same information, creating a context problem rather than a model problem.

Snowflake’s Horizon Catalog is intended to collect metadata and business logic from data inside and outside its platform. The company also supports zero copy integrations, a gateway for MCP servers and API connections, and a control layer for agent identities, authentication, permissions, costs and logging.

Blandina said companies should not be locked to one model provider. Rules can determine which model is used and how much a use case may cost. He described Cowork as a personal enterprise agent and Coco as a tool for developers building solutions. In an internal test, one colleague worked only through Cowork for two weeks, creating Slack messages, email and other tasks; the first two days were difficult, he said.

For companies adopting agentic AI, Blandina recommended starting with repetitive use cases that represent 80 percent of the work, rather than the most advanced projects.