Enterprises have moved past the point of asking whether AI agents can do useful work. The more pressing question now is whether teams can stand behind what those agents say and do. Is the agent answering a question based on the definition of revenue that finance actually approved? Can security see a risky tool call before it executes, and can a budget limit stop a runaway agent before it shows up in the quarterly forecast?
If governance and context aren’t built into agents from the start, the potential for conflicting answers and rogue agents compounds with every new agent — not to mention the risk of shadow AI and runaway AI spend. Today at Expedition 2026, Snowflake is announcing capabilities and enhancements that build trust into each layer of the platform by:
- Creating a shared, governed understanding of the business
- Extending Snowflake's control plane to every agent action, connection and dollar spent
- Grounding these capabilities in a data foundation that's open and interoperable by design
A dashboard reports one figure for net revenue, an analyst reports another, and an AI agent returns a third. When every team calculates "net revenue" in a different way and there’s no record of which version is official, even a perfectly governed agent can return an answer that nobody agrees on.
Business Ontology (public preview soon) is the newest addition to Snowflake Horizon Context. It gives organizations one governed place to record what their core business concepts mean, independent of any specific data structure. Business Ontology is also part of a unified approach to context governance: Snowflake Cortex Sense (public preview soon) finds context across your data, Business Ontology makes it authoritative, and Snowflake Semantic Views turn it into governed, consistent metrics.
Thanks to Cortex Sense, you can bootstrap Business Ontology from existing context within Snowflake, giving analysts a starting point for defining concepts and how they relate across teams. Once created, the people who own the concepts can approve their definitions. As business context evolves, Cortex Sense can suggest definitions, but the human owner controls the process of making them official.
You also don’t need to rebuild a glossary from scratch for Business Ontology. Organizations that keep definitions in a third-party catalog or a spreadsheet can migrate them with guided glossary import in a single session with Snowflake CoCo, Snowflake's coding agent. The import pulls out domains, candidate terms, relationships and linked data assets, and marks each one as a draft for its owner to approve through the steward approval workflow.
Each approved definition connects to semantic views, tables and views across teams and domains, so agents in finance, sales and support will all work from the same meaning. When a difference is intentional, such as a metric defined differently by region, Business Ontology keeps each version as an approved definition, labeled with where and when it applies. Meanwhile, Semantic Views run the approved logic when a question comes in.
Besides helping to keep governed context consistent and fresh, Cortex Sense also brings that context into the agent interaction itself, so agents are useful from the moment they're launched. When Cortex Sense infers something that conflicts with an approved definition, agents will use the approved definition and Cortex Sense learns from the experience.
Horizon Context creates a connected foundation for Business Ontology by collecting, enriching and activating context across the data estate. Its out-of-the-box connectors gather schemas, query logs, dashboard definitions and usage signals from sources like Tableau, Power BI, dbt, SQL Server and PostgreSQL. Horizon Context enriches that metadata with end-to-end lineage, tags, AI-generated descriptions and popularity signals.
Read our blog post about Business Ontology for more details.
Govern, observe and optimize every agent action #
Often, teams across every enterprise function are connecting agents to models, MCP servers and enterprise systems faster than IT and security can track them. That leaves leaders with a choice between scaling quickly and accepting potential security, compliance and cost risk or slowing down deployments until the security and cost capabilities catch up.
Leaders don’t need to make that choice with Snowflake, because it connects governed data, agent security capabilities, AI cost controls and agent observability features all in one system.
Bring Snowflake's control plane to every agent with Cortex AI Gateway
Cortex AI Gateway (public preview) is a centralized, governed entry point for agents to access models and the data they act on. It applies a consistent set of policies to agent connections across supported clouds, frameworks and AI providers. Because it governs from where the data lives, every agent inherits your organization’s existing identity and access controls, as well as any new controls as they are applied. Teams no longer have to create and manage a second set of credentials for agents.
Governance is the starting point, and new Snowflake Cortex Sandboxes (private preview) show Snowflake’s commitment to enabling governed autonomous agent actions. Cortex Sandboxes provide control over the code an agent executes: They give agents a secure, private space to run custom code alongside enterprise data in Snowflake, while maintaining and following the permissions of the person running the agent. There’s no need to move data off of Snowflake, and the spaces launch in seconds and shut down automatically upon completion, avoiding additional maintenance and reducing the risk of accidental exposures.
Snowflake is also continuing to enhance governance across AI workflows. With MCP governance through Cortex AI Gateway Tools (public preview), agent tool calls are authenticated with the end user's identity and checked against policy before they run. The system can block unapproved API calls, unauthorized actions and compliance violations before an agent reaches a sensitive system. Cortex AI Guardrails (public preview) adds account-wide runtime protection against prompt injection, jailbreaks and novel attacks to every agent routed through Cortex AI Gateway, including third-party agents.
Visibility comes next, because you can’t manage what you can’t see. Observability (generally available) captures every model call, tool invocation, reasoning step and guardrail decision in governed tables inside the customer's own Snowflake account. Those traces can be queried with standard SQL and joined to the rest of the business's data. Agent Trace Intelligence (private preview) uses that join to find agents wasting money in loops, tools that keep failing and the real cost of each completed task.
Cost is the final part of the picture. Cost Governance capabilities (generally available) attribute AI spend to a team, department, agent or application and enforce hard limits in near real time, helping managers identify and stop runaway agents before they exceed budgets. Dynamic Model Routing (private preview) matches each request to an organization-approved model based on cost, accuracy and latency, keeping routine work off the most expensive models. In internal testing, Dynamic Model Routing delivered up to 3x greater token efficiency than a frontier-model-only path at the same quality.
Bring real-time observability to agents and applications with Observe by Snowflake
Cortex AI Gateway provides a governed entry point for agents, but running AI reliably in production also requires understanding how agents behave and how the applications and infrastructure behind them perform. Observe by Snowflake brings these systems into a unified, AI-powered observability platform, helping engineering teams investigate failures, improve agent quality and optimize cost.
Observe has rebuilt its ingestion pipeline on Snowpipe Streaming (private preview), writing logs, metrics and traces directly into Snowflake tables so engineers responding to an incident see the most current data. Snowflake Data Share into Observe (private preview) turns data that customers already store in their own Snowflake accounts into zero-copy Observe data sets, so there's no need to ingest or store it twice. The O11y Agent in Slack (private preview) lets teams automate the first response to new issues and see the root cause analysis in the team's Slack channel.
For teams building their own agents and LLM applications, Agent Observability (private preview) tracks performance, errors, token usage and cost. It also supports search and debugging across traces, sessions and conversations, and runs online LLM-as-judge evaluations against production traffic. Telemetry is collected through OpenTelemetry and retained in Apache Iceberg™ tables on an architecture built to keep full agent context affordable, so teams can connect agent behavior to business data on Snowflake and assess agent value.
Extend governance across an open lakehouse #
Most enterprises have more than one data platform. They run multiple engines across multiple catalogs, regions and clouds. With each silo, agents have to reason over multiple — and sometimes conflicting — data copies, governance policies and business context layers. This results in answers that are difficult to trust, powered by data that agents maybe shouldn't have access to.
At Snowflake, we believe there’s a better way. Underpinning Business Ontology and Cortex AI Gateway is Snowflake's open-by-design architecture. It is grounded in open source communities across Apache Iceberg for storage, Apache Polaris™ for governance and Apache Ossie (incubating) for semantics. Together, they provide the mechanisms for you to connect all of your data across platforms to build an open and interoperable data foundation with universal governance and context to power your agentic enterprise.
Horizon Catalog, built on Apache Polaris, helps simplify multicatalog and multiengine architectures by offering a universal catalog without lock-in. Define one set of governance policies and enforce them in any Iceberg REST Catalog-compatible engine, regardless of catalog and without migration, with external engine access to external tables (generally available).
Enforcing security is half the equation; you also need to audit every engine interaction. This is why audit history now logs both Snowflake and external engine interaction s with Snowflake-managed and externally-managed Iceberg tables when Horizon Catalog is acting as the universal governance layer. Security and platform teams can investigate access, track adoption and validate integrations from one place, without building separate observability infrastructure.
Silos within your teams are only part of the challenge. What happens when a customer or partner needs access to data? This is where Snowflake AI and Data Sharing comes in, enabling data producers to quickly and easily share their products with consumers. It now supports Open Data Sharing, powered by Apache Polaris (generally available), which unlocks the ability to share data to any Iceberg REST compatible engine.
Snowflake is also bringing two other AI and Data Sharing enhancements to private preview. Cross-Cloud Sharing Without Replication makes provider listings available across regions without up-front replication and storage costs, allowing consumers to access data using remote queries on Iceberg tables. Schema Drift Detection and Copy Grants for Data Sharing helps producers and consumers identify schema or grant changes that could break sharing, while preserving copy grants when views are modified.
Manage AI spend and admin across teams and accounts #
AI spending overruns — or the threat of them — mean executive eyes are keeping a closer watch on credits and dollars. At the same time, managing accounts and performance is an important component of realizing AI ROI.
New FinOps updates give enterprises more cost governance control and flexibility across accounts. Administrators can apply per-user quotas, set weekly per-user limits and apply quotas to Cortex AI Gateway usage (all generally available). Support for org-level budgets (generally available) allows you to set spend targets in dollars or credits across multiple accounts using tags. Tag-based cost anomaly detection (generally available) provides insights into potential spikes and supports up to 50 custom spend cohorts per account, with daily alerts and a sandbox mode. As AI adoption grows, so does the number of Snowflake accounts an enterprise runs. At enterprises running 10+ Snowflake accounts, administrators often have to log into each account or write custom scripts to roll out org-wide changes.
Snowflake is changing that with Snowflake Organizations. It gives GlobalOrgAdmins one Snowsight interface to manage organization users and user groups, configure third-party domain rules and tenant types, and monitor violations. They can also push Trust Center scanner packages to every account at once, streamlining updates (public preview). New cross-account Org Usage Views (generally available) — including organization-wide AI spend by account and service — help make governance measurable and auditable, and admins get an insights-first, unified view of their organization’s health.
Performance tuning is also getting closer to zero maintenance. With Optima Clustering’s Automatic Key Selection feature (generally available), Snowflake makes clustering autonomous. Instead of an admin manually choosing clustering keys, Snowflake now selects keys based on historical query patterns and table access and automatically updates them as patterns change. Optima reclusters the parts of the table that are actually being queried according to workload heat maps, and value-based billing tied to ingest reduces the chance of cost spikes. Paired with Snowflake Adaptive Compute's ability to adjust compute as demand changes, this helps streamline performance maintenance across your enterprise.
Build agents your business can stand behind #
For users to trust an agent, the agent has to work from approved business meaning, act within policy, stay visible to the people accountable for it and spend within agreed limits. Snowflake makes this possible by connecting governed business context, agent security and cost controls in one system, and it extends that governance and context across an open data foundation. With this integrated, interoperable approach, enterprises can deploy more agents with confidence, continuously improve them and get more value from their AI investments. Explore more Snowflake Expedition announcements at snowflake.com/en/blog/expedition-26/, or join the virtual event at snowflake.com/en/expedition/.
Forward-looking statements: This article may contain forward-looking statements, including about future product offerings, and is not a commitment to deliver any product offerings. Actual results and offerings may differ. See our latest 10-Q for more information.
Frequently asked questions #
What is Business Ontology?
Business Ontology is a governance layer in Snowflake Horizon Context, public preview soon. It records what business terms officially mean, who approved them and which data asset computes them. Approved definitions bind to Semantic Views, tables and views and are served to Snowflake CoCo and Snowflake CoWork, so AI agents can answer from the authoritative definition.
What does Cortex AI Gateway do?
Cortex AI Gateway extends Snowflake's control plane to third-party AI agents. It checks MCP tool calls against policy before they run, records model calls and reasoning steps in a server-side observability plane, attributes AI spend and enforces hard budget limits in near real time. It also routes each request to the most efficient model.
How does Snowflake help you control the cost of AI agents?
Snowflake helps control AI agent costs at several levels. Cortex AI Gateway attributes spend to teams, agents and applications and enforces near real-time limits, while Dynamic Model Routing enables you to choose the best model for your request based on cost, accuracy and latency. Organization-level budgets, per-user quotas and tag-based anomaly detection extend cost governance across every account.
Can Snowflake govern Apache Iceberg tables managed in other catalogs?
Yes. Horizon Catalog, built on Apache Polaris, can access externally managed Iceberg tables through Horizon Iceberg REST Catalog and govern them in place without migration. Open Data Sharing, now generally available, delivers governed data to any Iceberg REST Catalog client.
How does Observe by Snowflake support AI agent observability?
Agent Observability, in private preview, collects agent telemetry through OpenTelemetry and tracks performance, errors, token usage and cost. It supports debugging across traces, sessions and conversations and runs LLM-as-judge evaluations on production traffic. Telemetry is retained in Apache Iceberg tables so teams can correlate agent behavior with business data.