Anyone who has attended an enterprise technology conference over the past year could be forgiven for reaching the same conclusion – AI cannot operate at scale – or deliver predictable business outcomes – without a unified data foundation.
According to Madhan Arumugam, Snowflake’s senior vice president of customer experience and frontier engineering, many enterprises still contend with fragmented data and disconnected systems. AI workloads can compound those problems by creating further security silos and operating without the context needed for AI to act reliably.
Snowflake, he argued, was designed to support the transition to an AI-driven enterprise.
“We think you know Snowflake as a platform with an architectural approach that is necessary in order to make the agentic enterprise vision come to life,” Madhan said.
That strategy rests on four core capabilities: enterprise data and context, model choice, connectivity, and a control plane.
High-quality data is essential to effective AI. Model choice enables organisations to select the most appropriate tool for individual tasks. Connectivity to application systems provides agents with the context they need, while a unified control plane brings these elements together and enables governed business actions.
Model choice and sovereignty
Technology vendors that work extensively with enterprise data, including Snowflake, broadly agree on one point: AI must be grounded in trusted enterprise data and context.
“It is the beginning of all beginnings on your AI journey,” Madhan said.
Customers like Ronnie Chooi of Sime Darby made a similar point – “Context is fuel for AI and makes it dependable. If you manage your context, operating AI is better, and answering without context can be dangerous.”
Madhan’s presentation then turned to model choice and AI sovereignty. Rather than locking customers into a single large language model, Snowflake advocates a flexible “model marketplace” approach, in which frontier and open models can operate side by side.
The premise is that there should be no model lock-in. Not every workload requires a top-tier frontier model; the challenge is to find the right combination of models to achieve business outcomes efficiently and effectively.
This flexibility, combined with intelligent model routing, is also positioned as a major cost-optimisation measure. (Watch video)
“You will typically be seeing anywhere from two to eight times cost savings when you have a good model routing layer,” Madhan said, describing “AI tokenomics” as a consistent topic in customer conversations.
Equally important, Snowflake argues, is control over enterprise data. The company positions its platform as a way to support AI workloads without compromising data sovereignty.
“Our fundamental argument today is that your data is yours. We don’t use it to train our models or context layers, and your data never leaves the trusted boundary,” Madhan said.
“This is what true data sovereignty means in practice: completely controllable security in your model selection, and your AI costs.”
Governing AI actions
The SVP of frontier engineering also referenced Snowflake’s intent to acquire Natoma, an enterprise Model Context Protocol (MCP) platform for AI agents.
Through Natoma, Snowflake aims to establish a natively integrated governance and identity layer for AI agents and MCP tool access. The proposed integration would extend governance into AI-driven workflows, helping enterprises manage both data access and the actions agents take across business processes.
When Snowflake announced intent to acquire Natoma last May, its CEO Sridhar Ramaswamy said, “AI agents are quickly becoming part of how enterprises operate, but intelligence without governance creates risk.
“Agents don’t just need access to data. They need the right context, permissions and policy guardrails to operate safely inside the enterprise.”
The Snowflake CEO added that Natoma’s capabilities could allow coding agents to “come alive inside enterprises” in a way that is secure, auditable and scalable.
