In enterprises, agents are most likely to reveal their shortcomings on what seems like a simple issue: the term "active customers" used by sales, "effective contracts" by finance, and "conversions" by operations may each have their own definition. Models can read data, but they don't know which definition is the correct one; once these models are not only capable of answering questions but also able to invoke tools and modify systems, the confusion in definitions can escalate from disputes over reports to real operational risks.
Snowflake released Business Ontology and Cortex AI Gateway on Expedition of November 3, 2026, focusing precisely on this issue. The former incorporates business terminology approved by enterprises into a unified semantic layer, while the latter manages how agents connect to models, invoke tools, and perform actions. They do not promise to make models suddenly smarter, but rather aim to have models work within the same set of definitions, permissions, and audit rules.
A sponsored study cited by Snowflake indicates that 55% of the executives surveyed listed a lack of business context as the primary reason for the low quality of agent insights, while another 45% mentioned data fragmentation or inaccessibility. These figures are self-reported by companies and cannot be directly equated with deployment failure rates, but they highlight a frequently overlooked fact: just because a model has access to a large amount of data does not mean it understands what that data represents within the company.
What business ontologies solve is not "more knowledge," but rather "there is only one authoritative version of the answer."
The role of Business Ontology is to organize core concepts such as revenue, customers, products, and regions, as well as the relationships between them, into approved definitions. When agents query data, they do not need to guess the meaning of fields on the spot, nor do they have to piece together answers from explanatory documents maintained by different teams. For cross-system tasks such as finance, sales, and supply chain, this unified approach is more direct than adding another model parameter.
True value emerges when definitions conflict. In the past, analysts would discover discrepancies between two reports during meetings; in the era of intelligent agents, errors can be replicated within seconds to emails, work orders, and automated approval processes. By placing definitions at the forefront, we transform "first clarify how to calculate" into a system constraint. This approach cannot eliminate dirty data, nor can it replace business owners in verifying indicators, but it can reduce the arbitrary interpretation of models in areas of semantic ambiguity.
Cortex AI Gateway is responsible for the other half of the task: integrating different agents into the model, while MCP manages the pathways to servers and enterprise systems centrally. Enterprises can monitor the usage of the model, tool calls, and associated costs, and they can also link these actions to a unified policy. The focus here is not on forcing all agents into the same product, but rather on ensuring that agents from different teams pass through the same controlled gateway.
From response to execution, governance must follow action.
Errors in read-only Q&A systems usually remain on the screen; however, for agents that can place orders, update customer records, or trigger payments, errors can penetrate into business processes. Therefore, companies cannot simply record which model was used; they also need to know which version of the business definition it was based on, which data was accessed, what tools were invoked, and whether the final actions were approved.
Snowflake emphasizes that every decision and action is auditable and is subject to the same set of policies. This direction is reasonable, but the announcement itself does not prove that all external models, tools, and cloud environments have been seamlessly integrated, nor does it provide a unified accuracy rate or cost reduction figure. Enterprises still need to set up manual confirmation, minimum permissions, and rollback paths for high-risk actions. Gateways can be centrally controlled, but they will not automatically decide for business departments which actions are worthy of approval.
This release is more of a signal: the competition among enterprise agents is shifting from “who can provide more human-like responses” to “who can consistently do one thing correctly within complex organizations.” The capabilities of the model determine whether it can complete the task, the business context determines whether it understands the task, and the governance mechanisms in place determine the extent of losses that will occur if mistakes are made. Without any one of these three elements, the more diligent the agent, the faster the risk will spread.
For teams preparing to launch intelligent agents, it might be more cost-effective to first clarify the definition of metrics and tool permissions than to purchase more models upfront. A shared business ontology provides a source for answers, and a unified execution gateway sets boundaries for actions; what can truly be scaled is not the number of intelligent agents, but the enterprise's ability to explain each response and action.
