Most companies no longer struggle to build an AI agent. They struggle to trust one enough to put it into production. That is the problem Snowflake set out to solve at its Expedition 2026 event, announcing a set of capabilities that put business context, governance, and autonomous execution on one platform: Business Ontology, Cortex AI Gateway, and new always-on Bots for its CoCo and CoWork agents.
The announcements, made on November 3, 2026, frame a shift in enterprise AI. The bottleneck is no longer the model. It is whether an agent knows what the business means, whether it follows the rules, and whether anyone can see what it did and what it cost.
Why Agents Stall in Pilot
Snowflake’s case starts with a number. In a survey of 400 senior data and AI executives conducted by MIT Technology Review Insights in July and August 2026, 55% cited a lack of business context as the primary cause of low-quality AI agent insights, with fragmented or inaccessible data second at 45%. Every respondent worked at a company earning at least $500 million a year.
The failure mode is easy to picture. Ask an agent for a weekly net revenue report and a dashboard, an analyst, and the agent can each return a different number — because nobody agreed on what “net revenue” means. An agent that does not know the company’s official definition gives a confident wrong answer that someone has to catch.
Meanwhile, agents are connecting to models, MCP servers, and enterprise systems faster than IT and security teams can keep up, creating new risks and unexpected bills. The result is that many companies keep their agents in pilot phase, unable to scale them across the business.
Business Ontology: One Definition Everyone Agrees On
The first piece of Snowflake’s answer is Business Ontology, the newest capability in Horizon Context — the home for governed business context inside Snowflake Horizon Catalog. It gives every agent the same approved definitions, so “net revenue” means the same thing to an agent as it does to the CFO.
A Business Ontology governs the official meaning of a term, a Semantic View calculates it, and Cortex Sense infers context that is not already defined. Together they form one authoritative business context for every agent, app, and person in the enterprise. Teams approve official definitions for core concepts, assign each one an accountable owner, and link it to the underlying data, inside and outside of Snowflake. Approved definitions always take precedence, and every change is logged — so when an auditor asks what net revenue meant in a given quarter, the answer is a simple lookup.
Business Ontology is coming to public preview soon. It is not generally available yet, which matters for teams planning around it: the capability is real, but it is still an emerging foundation to validate rather than a production standard to assume.
Cortex AI Gateway: One Set of Rules for Every Agent
An agent that knows the right definition still needs the right controls on how it accesses data, calls models, and uses tools. That is the job of Cortex AI Gateway, the single front door for agents in an organization. As part of Horizon Catalog, it applies one set of policies to agent connections across clouds, frameworks, and AI providers.
Agents act on enterprise data through MCP servers, and the Gateway applies the company’s access policies, authentication, and permissions to those calls. Agents inherit the controls already established for human users, so teams do not have to build a second set of operations for machines. The core gateway capabilities reached general availability on November 3, 2026.
The controls layered on top of it include:
- Cortex AI Gateway Tools (public preview): every tool call an agent makes is checked against fine-grained authorization policies before it runs.
- Cortex AI Guardrails (public preview): runtime protection against prompt injection, jailbreaks, and novel attacks.
- Cost governance (generally available): AI spend is attributed to the team, department, agent, or application that generated it, with hard limits enforced in near real time. Dynamic Model Routing, which sends routine work to lighter, lower-cost models, is in private preview.
- Observability and abuse detection: every step is recorded in the customer’s own Snowflake account, where teams can analyze it with Agent Trace Intelligence. Agent Abuse Detection analyzes complete agent trajectories — not individual prompts — to catch agents operating beyond their permissions. Both are in private preview.
“Which teams and which agents were driving the spend” was the question Andy Owens, VP of Data and AI at Kargo, said his team could not answer before the Gateway. “Cortex AI Gateway gave us that breakdown, plus hard limits so a runaway process doesn’t quietly burn through a budget,” he said.
CoCo and CoWork Bots: Agents That Work While You Sleep
With context and controls in place, Snowflake argues enterprises can do what they have been holding back on: let agents work on their own. CoCo, the coding agent for data teams, and CoWork, the personal work agent for knowledge workers, now take on recurring work for both audiences. CoWork is used by more than 5,800 accounts and CoCo by more than 9,100, according to Snowflake.
CoCo Automations, generally available since November 2, let data teams describe a workflow once and have CoCo run it autonomously at any frequency — a routine pipeline health check, or a weekly report someone used to build by hand. Every run keeps an audit trail and task history, so teams can see exactly what CoCo did and suspend or rerun a job whenever they need to.
The newest step is CoCo and CoWork Bots, now in private preview. Unlike agents that wait for a question, Bots act like an autonomous chief of staff: given a high-level goal such as preparing briefings for every customer meeting or keeping data pipelines healthy, they plan and complete the work in the background, operating strictly within the user’s governed boundaries and surfacing only when a critical decision or final approval is needed. They inherit the user’s exact access rights, role context, and permissions.
The agents also work where people already do, through Snowflake for Slack and CoCo for Excel and Sheets, both in public preview. Skills and Plugin Sharing and Profiles in CoCo mean what one team builds, the whole company can reuse — and the new AI Data Package offers more than 60 curated industry datasets, complete with semantic views and a ready-to-use CoWork agent.
What This Means for the Agentic Enterprise
Snowflake’s pitch is that context, controls, and execution are built together, so each makes the others stronger. Business Ontology makes sure agents give the right answer. Cortex AI Gateway makes sure they get it the right way. CoCo and CoWork turn those answers into finished work. Because all of it runs on one governed platform and one copy of the data, enterprises no longer have to choose between moving fast and staying in control.
“The enterprises that win won’t be the ones with the most agents. They’ll be the ones whose agents know exactly what the business means by accessing the right data, abide by governance and guardrails specified, and can account for actions taken by agents,” said Christian Kleinerman, EVP of Product at Snowflake.
The broader industry is moving the same direction. Confidential computing hardware such as Google’s TPU 8tx, built to protect data while it is being processed, shows how seriously the industry now takes protecting data in use — and the same week brought a reminder of the stakes, with the OpenSSH SSHole flaw showing how quickly an ungoverned gap can be exploited. Snowflake’s bet is that the winning platform will be the one that treats governance as the foundation of the agent, not an afterthought bolted on later.
For data leaders, the practical takeaway is to separate what is ready now from what is still arriving. Cortex AI Gateway’s core capabilities and CoCo Automations are generally available today; Gateway Tools and Guardrails are in public preview; Business Ontology is coming to public preview soon; and the Bots, sandboxes, and abuse detection remain in private preview. The architecture is coherent — but the pieces arrive on different schedules, and any rollout plan should be built around the availability that actually exists.
