Snowflake 推出 Business Ontology，提升 AI 智能體可靠性

Snowflake is hoping to improve the reliability of AI agents with a new tool called Business Ontology, as part of its Horizon Catalog, meant for data discovery, management, and governance.

Horizon Catalog received a boost in June with the launch of Horizon Context, a set of semantic and metadata management capabilities to help enterprises collect, enrich, and organize metadata across their data estates and give AI agents a better understanding of enterprise data, its relationships, and how it is used.

Business Ontology builds on that foundation by allowing enterprises to define and approve business concepts, assign owners to those definitions, and link them directly to the underlying tables and views, Snowflake said in a statement.

That, in turn, should give enterprise AI agents access to authoritative business definitions, such as what constitutes net revenue, rather than relying on competing interpretations of the same data across different systems, it added.

Business Ontology also offers enterprises a way to define and approve business meanings and relationships that have not been formally documented.

Typically, enterprises subscribed to Snowflake have had to rely on Cortex Sense, Snowflake’s AI-powered business context engine, to infer these meanings from available metadata and usage patterns, which may not always reflect the intended business definitions.

With Business Ontology, enterprise users would now be able to review, formalize, and approve these definitions, ensuring AI agents use business-validated meanings rather than relying solely on AI-generated inferences, the company said.

Business Ontology could simplify agent development

For developers, Business Ontology can reduce the effort and time required to stitch context for agents or agentic applications as they no longer have to bake business logic into every agent they build, said Ashish Chaturvedi, executive research leader at HFS Research.

Primarily because agents can draw on the centrally maintained definitions in Business Ontology, allowing developers to reuse the same business context across multiple agents and applications, Chaturvedi added.

This, in turn, could also reduce inconsistencies between agents interpreting the same business data differently, while making it easier to develop and maintain agentic applications as business requirements evolve, echoed Michael Ni, principal analyst at Constellation Research.

For CIOs, that same ability to establish a common set of business definitions could make it easier to scale AI agents across departments without having to worry about different agents interpreting the same business data differently, Chaturvedi said.

However, he warned that keeping business definitions current could still be challenging for enterprises.

“Each core term needs a named owner, a review whenever pricing, products, or reporting rules change, and a link to the data behind it, so a schema change prompts someone to check the definitions that depend on it,” Chaturvedi said.

That process might be further complicated by the fact that different departments may define the same business concept differently.

“Sales, finance, and operations, for instance, may have different interpretations of what constitutes a customer or revenue, based on their respective business requirements. Enterprises will need to decide whether to reconcile these differences into a common definition or retain them as separate, clearly identified terms, a process that requires agreement across business units,” Chaturvedi said.

Business Ontology, according to Snowflake, is expected to be available in public preview soon.

New controls to monitor and govern agent activity

Beyond establishing a common understanding of enterprise data, Snowflake is also adding capabilities to Cortex AI Gateway, its agent control layer, to help enterprises oversee how AI agents execute tasks, access data, and consume resources.

These include Cortex Sandboxes and Agent Trace Intelligence, which provide agents with isolated environments to execute custom code alongside enterprise data and allow developers and data teams to examine agent execution traces against business data, respectively.

For developers, Cortex Sandboxes are likely to increase productivity, Chaturvedi said: “It takes away a chore most developers would happily skip, which is setting up a safe place for agent-written code to run. Today teams either build and secure their own containers, copy data somewhere else, or simply don’t let agents run code.”

“Sandboxes start next to the data, follow the permissions of whoever launched the agent, and shut down once the job is done, so developers can build agents that write and run analysis code without managing servers or juggling credentials,” Chaturvedi added.

That, in turn, for CIOs, should reduce risks, especially in terms of any damage an agent could do, as it will run in a walled-off secure environment, according to Ni.

This means that while an isolated environment can limit the impact of faulty or malicious code on surrounding systems, it does not necessarily prevent an agent from misusing data or access rights that the user is already authorized to exercise.

For Chaturvedi, however, the bigger concern for CIOs is ensuring that the code executed by agents produces accurate and reliable results. While sandboxing can limit the damage caused by faulty or malicious code, it does not validate the logic used or the accuracy of the output, he said.

Both Cortex Sandboxes and Agent Trace Intelligence are currently in private preview.

Alongside Cortex AI Gateway’s capabilities, Snowflake is also introducing Agent Abuse Detection to help enterprises address security concerns around agents.

For its part, Agent Abuse Detection can help enterprises identify agents operating beyond their permissions, such as attempting to bypass approval processes or accessing sensitive data, the company said. It is currently in private preview.