At Snowflake Expedition 2026, the company announced what may be the most important structural bet in enterprise agentic AI this year. It is not a faster model, a cheaper token or a bigger context window. It is a governance layer: Business Ontology and Cortex AI Gateway, working together so that agents understand what the business means, act within its controls, and can account for what they did.

The urgency behind the announcement is well documented. In a recent MIT Technology Review survey sponsored by Snowflake, 55% of executives cited a lack of business context as the primary cause of low-quality agent insights, and 45% cited fragmented or inaccessible data. The result is what Snowflake itself describes as the pilot trap: enterprises can build an AI agent in weeks, then spend months wondering whether they can trust it enough to put it into production.

The trust problem in three parts

The survey distils the enterprise trust gap into three interconnected problems. First, agents produce confident answers based on ambiguous definitions. Ask an agent for weekly net revenue and you may get a different figure from the dashboard, the analyst and the bot, because neither the agent nor the human has a single, signed-off definition to work from.

Second, agents are connecting to models, MCP servers and enterprise systems faster than IT and security teams can keep up. Every new connection is a new attack surface, a new data policy and a new cost centre that may not show up on anyone's existing budget until it is too late.

Third, and most consequentially, the gap between build and production is widening. Enterprises that solved the technical question of whether an agent could complete a task are now stuck on the operational question of whether they can let it do so unsupervised. Snowflake's answer is to treat context, controls and execution as a single system rather than three separate procurement decisions.

Business Ontology: one meaning, one owner, one audit trail

Business Ontology, now in public preview inside Snowflake Horizon Catalog, governs what the business means by its core concepts. Teams approve definitions for terms like net revenue, assign each one an accountable owner, and link it to the underlying data tables inside and outside Snowflake. Approved definitions always take precedence over anything inferred by Snowflake's Cortex Sense.

The implications go beyond accuracy. When an auditor asks what net revenue meant in Q2, the answer is a lookup, not a meeting. When an agent produces a report, its reasoning traces back to a definition that a named person approved and timestamped. That shift from inference to provenance is what turns an agent from a conversational tool into a system the finance function can sign off on.

Snowflake also supports migration from third-party catalogues or spreadsheets in a single guided CoCo session, with every change logged. For regulated industries already running on Snowflake, this is a path to production that does not require a separate governance product or a second copy of the data.

Cortex AI Gateway: one front door, one set of rules

Cortex AI Gateway acts as the single access point for every agent, model and tool call in an organisation. As part of Horizon Catalog, it applies one policy set across clouds, frameworks and AI providers. Agents connect to enterprise data through MCP servers, and the gateway enforces the same access policies, authentication and permissions that human users already have.

Snowflake breaks the gateway's functionality into five practical layers. Cortex AI Gateway Tools check every tool call against fine-grained authorisation policies before it runs. Cortex AI Guardrails protect against prompt injection and jailbreaks. Cost governance attributes spend to the owning team and enforces hard limits. Dynamic Model Routing sends routine requests to lighter, lower-cost models. Observability records every step in the customer's own Snowflake account, where teams can analyse it with Agent Trace Intelligence and Agent Abuse Detection, which examines complete agent trajectories rather than individual prompts to catch agents operating beyond their permissions.

Kargo, a logistics technology company, provided a customer quote that illustrates the operational value: "Cortex AI Gateway gave us that breakdown, plus hard limits so a runaway process does not quietly burn through a budget. We are running more agents than ever and we know what each of them costs."

CoCo and CoWork Bots: from query to co-worker

With trust infrastructure in place, Snowflake is now deploying agents that work autonomously rather than waiting for prompts. CoCo Automations let data teams describe a workflow once and run it autonomously on any schedule, keeping an audit trail and task history for every run. CoCo and CoWork Bots act as always-on AI co-workers that operate in the background inside the tools teams already use, surfacing only when a decision or sign-off is required.

According to Shelter Mutual Insurance Company, which has eight agents in production and more underway: "It would take a lot of time to create semantic layers in the past, but things have changed drastically in the last three months since we started using Snowflake CoCo. We now have access to real-time insights on demand directly in our operational processes."

What the announcements mean for enterprise buyers

For procurement and architecture teams, the Snowflake announcements carry four practical implications.

First, the governance question is moving upstream. Organisations that were planning to buy agent tools and bolt on governance later are discovering that retrofitting trust is harder and more expensive than building it in from the start. Snowflake is betting that buyers will prefer platforms where context, controls and execution are already integrated.

Second, the total cost of agent ownership is becoming visible. Token spend has historically been the easiest number to track, but it is only one component. The cost of cleaning up an incorrect report, remediating a policy breach or retraining a model on wrong data can exceed the infrastructure bill by an order of magnitude. Cost governance at the gateway level is a start toward accounting for the full lifecycle cost.

Third, multi-cloud and multi-model deployments are now the default, and they need a unified control plane. The Cortex AI Gateway is designed for heterogeneity: it does not care whether the model is OpenAI, Anthropic, Google or open source, or whether the data sits in Snowflake, S3 or Azure. Buyers evaluating agentic platforms should test whether their preferred vendor can deliver the same cross-cloud policy consistency.

Fourth, the audit and compliance bar is rising. Regulators in the EU, the US and the UK are all drafting AI-specific accountability requirements, and the first enforcement actions have already begun. An agent that cannot explain how it reached a decision, who approved its inputs and what it did with its outputs is an agent that will eventually be shut down by compliance.

What the announcements mean for suppliers

For companies selling agentic AI platforms, Snowflake's moves recalibrate what "enterprise-grade" means. Features are no longer the differentiator. Governance is. Suppliers that can demonstrate mapping between business definitions, agent actions and audit outputs will win procurement cycles against competitors still pitching capability without accountability.

The second signal is that data platform incumbents are becoming agentic control planes. Snowflake already holds the data, governance rights and compute infrastructure for more than 14,500 enterprises. By extending Horizon Catalog into Business Ontology and the Cortex AI Gateway, it is turning that data position into an agentic position. Specialist agent vendors should assume they will be evaluated alongside, and potentially against, the data platforms their buyers already pay for.

Third, the always-on bot is coming. CoCo and CoWork Bots are not assistants that answer when asked. They are autonomous workers that plan, execute and escalate. Suppliers that cannot yet offer this level of autonomy need a credible roadmap, because buyers are already asking for it.

The Agentic Expo angle

Agentic Expo, at Olympia London on 23 to 24 March 2027, is built on the premise that the questions enterprises are asking have shifted. A year ago, the prevailing question was whether AI agents could perform useful work. Today, the question is whether enterprises can trust them enough to let them do it at scale.

Snowflake's Expedition announcements confirm that the answer is not more agents, faster agents, or cheaper agents. It is agents that know what the business means by "net revenue," act within the controls the business has already defined, and can account for every decision when asked. The vendors and suppliers that solve this trust equation are the ones that will own the next phase of enterprise AI adoption. That is precisely the audience Agentic Expo brings together: the buyers deciding which platforms to trust, and the suppliers building platforms worth trusting.

Get Tickets Exhibit at Agentic Expo

Sources: Snowflake, "Snowflake Powers the Agentic Enterprise", 9 November 2026; BizToday, "Snowflake Powers the Agentic Enterprise", 9 November 2026; MIT Technology Review Insights in partnership with Snowflake, "Raising the Quality of Insights in Agentic Analytics", 2026.