Snowflake Business Ontology launched on 3 November 2026 at the data cloud company’s Expedition conference, introducing semantic controls and sandboxing features to help enterprises govern autonomous AI agents. The new system aims to prevent software agents from misinterpreting internal metrics by linking corporate definitions directly to governed data tables.
Key points
- Snowflake Business Ontology allows organisations to create, approve and assign ownership to official business definitions for autonomous software.
- New Cortex AI Gateway controls add isolated code execution environments and abuse detection to block agents from exceeding their permissions.
- The platform’s coding and workplace tools will use the central ontology to prevent conflicting answers across departments.
- Cortex Training entered public preview, allowing companies to fine-tune open-weight models on their own infrastructure.
Grounding agents in corporate definitions
As businesses assign broader operational tasks to software agents, discrepancies in business terminology have created unexpected risks. When separate systems define fundamental numbers such as net revenue or active customers differently, automated tools often draw incompatible conclusions from the same underlying databases.
Snowflake Business Ontology tackles this friction within Horizon Catalog, the company’s data governance product. The system lets corporate teams establish authoritative business concepts, nominate individual owners for each definition, and connect those concepts directly to database tables and analytical views. Rather than forcing autonomous systems to guess what data fields represent, the catalog gives them explicit corporate context.
Previously, customers relied on automated inference engines to extract business context from metadata and system usage. While automated scanning surfaces patterns, it often misses intended corporate definitions. Business Ontology introduces a human review step where business units review, document and verify those meanings before deploying agents into production workflows.
Standardising these terms helps external models operate with fewer custom prompt adjustments. Proprietary assistants including Snowflake CoCo, which writes code for data engineers, and Snowflake CoWork, designed for administrative knowledge workers, will draw directly from the new framework to ground their actions.
How does Business Ontology simplify software development?
Building automated systems has traditionally required engineering teams to hardcode bespoke business rules into each individual software workflow. Centralising business terms within the data platform removes that repetitive labour from engineering teams.
Developers no longer need to reconstruct business logic across different agentic tools, said Ashish Chaturvedi, executive research leader at HFS Research. Instead, software agents can reference uniform records stored inside the catalog, allowing teams to repurpose identical definitions across multiple internal systems.
That central repository makes maintaining automated workflows simpler when corporate rules shift. Michael Ni, principal analyst at Constellation Research, pointed out that using verified definitions reduces inconsistencies between separate tools while allowing engineering teams to adjust systems smoothly as commercial demands change.
However, analysts caution that keeping definitions updated over time will demand regular organisational effort. When product catalogues, pricing tiers or accounting standards change, data administrators must review the corresponding documentation. If database structures change, corporate owners must re-examine the definitions attached to those fields to ensure automated reasoning remains accurate.
Reconciling differences between departments presents another hurdle. A sales team may measure customer relationships by contract dates, whereas finance departments count active customers solely through paid invoices. Companies must either negotiate a single shared term or register both interpretations as distinct, documented concepts.
Walled gardens and security guardrails
Establishing semantic accuracy addresses only half of the enterprise challenge. As organisations deploy enterprise AI agents to run code independently, security teams need technical controls to restrict what those automated workers can touch.
Snowflake is expanding Cortex AI Gateway, a central administration layer introduced in July 2026, with features designed to supervise agent behaviour. The system adds Cortex Sandboxes, providing temporary execution spaces where automated tools can run custom code alongside enterprise records without exposing broader production servers.
Setting up secure runtimes for machine-generated scripts has traditionally been a tedious task for IT groups. Developers frequently had to provision private virtual containers, export records out of protected networks, or ban agents from running analysis scripts entirely. Cortex Sandboxes spin up beside enterprise data, match the permissions of the user who triggered the query, and delete themselves once calculation wraps up.
While sandboxes block flawed or hostile code from damaging surrounding enterprise networks, security researchers note that isolation cannot prevent an agent from misusing legitimate user credentials. If an employee already possesses clearance to view confidential payroll files, an isolated agent acting on their behalf can still query that sensitive data.
To mitigate credential misuse, Snowflake introduced Agent Abuse Detection into private preview. The monitoring software flags systems that attempt to bypass corporate approval steps or access data pools outside normal operational routines. The move comes as infrastructure providers build specialised defences against automated misbehaviour, matching efforts like CrowdStrike AI security acquisitions aimed at containing rogue agents across commercial networks.
Observability and the battle for enterprise control
To give auditors clearer visibility, Snowflake is testing Agent Trace Intelligence in private preview. The utility records the exact reasoning path and technical operations an agent executes, allowing data managers to inspect how automated reasoning produced a specific output.
The expansion reflects an intensifying industry contest to serve as the default governance bridge for automated enterprise software. Big business software vendors are pursuing similar control points from application layers. SAP has developed governed systems through its Joule platform, while Salesforce promotes its own control plane architecture to supervise workflows. Other platforms are pursuing infrastructure solutions, such as the Uber MCP Gateway that unifies thousands of developer tools across distributed corporate servers.
Snowflake approaches the governance question from the storage repository. Chief executive Sridhar Ramaswamy framed the transition as a shift in how enterprises value technology, arguing that competitive advantages now belong to companies that can govern proprietary knowledge rather than those relying purely on generic coding talent: “AI is industrializing software creation. Actually, it’s industrializing intelligence.”
Alongside governance tools, Snowflake placed Cortex Training into public preview. The service allows enterprise subscribers to fine-tune open-weight language models using proprietary internal records without leasing large arrays of graphics processing units. Snowflake manages the computing infrastructure behind the scenes.
Christian Kleinerman, executive vice president of product at Snowflake, described the commercial value of custom tuning: “Specialized, smaller models that come with more efficient responses, faster responses, and frequently cost benefits in terms of tokens.”
Targeted models allow regulated enterprises to meet data sovereignty requirements, since keeping training pipelines on controlled infrastructure simplifies technical audits. Separately, Snowflake widened its reach by deploying its platform as a first-party service on Oracle Cloud Infrastructure, giving joint customers direct access to analytical tools across distributed corporate databases.
Snowflake expects to move Business Ontology from private preview into public preview in the coming weeks, while Cortex Sandboxes and Agent Abuse Detection remain in private testing.
