Snowflake Builds Enterprise Controls for AI Agents

Governance becomes the product layer

Snowflake’s November 6 announcements at Expedition 2026 focus on the control layer enterprises need before AI agents can handle recurring business work. The central industry judgment is that agentic AI adoption increasingly depends less on the number of agents deployed than on whether organizations can define their business context, constrain agent actions and audit what those agents do.

Snowflake presented this as an integrated system rather than a collection of standalone assistants. Business Ontology extends Horizon Catalog with approved, accountable definitions linked to underlying data. Semantic Views handle calculations, while Cortex Sense infers concepts that have not already been defined. The intended result is consistency between definitions used by agents, applications and employees.

The company also said its platform serves more than 14,500 global enterprises. That installed base gives Snowflake a position in the data, models and workflows on which agents depend, but the announcements do not establish new customer orders or quantified adoption of the new capabilities.

A gateway for permissions and costs

Cortex AI Gateway is intended to govern models, tools and actions across multicloud environments, frameworks and AI providers. Its reported mechanisms include inheriting existing human access permissions, authentication and authorization; inspecting tool calls against granular policies; filtering prompt injections and jailbreak attempts; and recording execution steps in customers’ Snowflake accounts.

Snowflake also described spend controls, including hard limits, department- or agent-level compute tracking and Dynamic Model Routing for routing routine work to smaller models. These features address a practical barrier to production deployment: agents can connect to models, Model Context Protocol servers and enterprise systems faster than security and IT teams can supervise their behavior.

The evidence supports a governance and cost-control thesis, but does not disclose adoption levels, customer implementation results or commercial terms for the new products.

Two agents and background workers

Snowflake CoCo is a coding agent for data-engineering and data teams, while CoWork is a personal work agent for knowledge workers. New CoCo Bots and CoWork Bots are described as always-on background instances that work directly in tools used by teams and seek human input only when a decision or sign-off is required.

This design shifts the competitive emphasis from interactive assistance toward recurring execution. It also places Snowflake closer to workflow platforms and agent-control infrastructure, although the announcements do not show that autonomous bots have general availability, named customers or proven production performance.

Trust remains the adoption constraint

Snowflake cited a recent MIT Technology Review report sponsored by the company in which 55% of executives identified a lack of business context as the primary cause of low-quality AI-agent insights. Fragmented or inaccessible data ranked second at 45%.

Those figures shape the company’s product strategy: Snowflake is addressing the gap between agent creation and enterprise deployment by joining approved context, governance and autonomous execution. Because the report was sponsored by Snowflake, the percentages are best understood as the basis for its stated market diagnosis rather than independent proof of an industry-wide adoption barrier.

The next observable milestone is company-disclosed evidence of production use—such as customer implementations, availability details or adoption metrics. The reviewed announcements do not provide a date for those disclosures.

Industry and product implications