Snowflake Launches Project SnowWork For Agentic AI - Let's Data Science
Snowflake wants to be the air traffic control for enterprise AI — and it just filed a flight plan.
The company announced Project SnowWork on March 18, 2026, positioning it as what CEO Sridhar Ramaswamy called a "central control plane" for agentic cohesion across the enterprise. The pitch: enterprises have invested heavily in data infrastructure and AI models, but the actual output — board decks, churn analyses, supply chain reports — still requires human coordination, tickets, and manual handoffs. SnowWork is designed to close that gap by letting non-technical business users request outcomes in plain language and have the system plan and execute multi-step workflows on governed Snowflake data, end-to-end.
It is a research preview, available to a limited set of customers. But the architecture bet is unambiguous: Snowflake wants to sit between the AI model layer and enterprise action, becoming the orchestration substrate that decides what gets done, with what data, under what constraints. That is a different competitive position than selling compute or hosting models — it is closer to being the traffic cop for agentic enterprise workflows. And it is, not incidentally, exactly the kind of central orchestration point that the enterprise security community has spent the past year identifying as a systemic vulnerability.
SnowWork sits in a three-layer stack that Snowflake has been assembling over the past year. Snowflake Intelligence, now generally available, handles natural language Q&A against enterprise data. Cortex Code, which went generally available in February 2026, is a data-native coding agent for developers. SnowWork is the execution layer for business users — finance, sales, marketing, operations — who need finished outputs, not query results. The workflow is continuous: Intelligence surfaces the insight, SnowWork acts on it.
The platform uses pre-built, persona-specific AI profiles configured for common business workflows, terminology, and KPIs per function. A finance user asking for a board-ready forecast deck gets a different agent configuration than a sales ops user running territory rebalancing. Those profiles plan and execute autonomously — querying data, applying analysis, synthesizing insights, generating structured deliverables, and preparing next steps within a single interaction. The governance layer is Snowflake's own RBAC, masking policies, and audit logging applied automatically, so the AI operates within the same trust perimeter as the underlying data. Whether that RBAC model — originally designed for human query patterns — is the right foundation for autonomous multi-step agent execution is an open architectural question that Snowflake is essentially betting will not matter in practice.
"We are entering the era of the agentic enterprise," Ramaswamy said in the announcement. "This shift is about much more than technology, it is about unlocking new levels of productivity and efficiency by embedding intelligence directly into the operating fabric of the enterprise." In an interview with CNBC, he added that SnowWork lets enterprise employees take action across systems like Salesforce and ServiceNow — all orchestrated from within Snowflake's data layer.
The differentiation Snowflake claims is data groundedness. General-purpose AI agents work from documents, emails, and online content — operating from context that is inherently stale and loosely governed. SnowWork is anchored to Snowflake governed enterprise data, shared business definitions, and cross-cloud interoperability. Every action the agent takes is auditable within the same policy framework that governs who can see what data. That is a real architectural advantage for enterprises already invested in the Snowflake ecosystem — the governance model extends directly into agentic execution without a translation layer.
The "control plane" framing is where the competitive picture gets interesting — and where the irony cuts both ways. Snowflake is positioning itself to own the orchestration decision: what gets done, with what constraints, when human intervention is required, how execution coordinates across systems. That is a layer that Microsoft, Salesforce, and ServiceNow are all trying to own as well, and they are not standing still. Microsoft has been building agentic orchestration into its Copilot stack. Salesforce launched AgentForce. ServiceNow has its own AI Control Plane for workflow automation. The question is whether a data platform is better positioned to be that control plane than an application platform or an AI framework vendor — or whether the "control plane" claim is really just another way of saying "the IAM gap we identified in the CSA survey."
"Enterprises have invested heavily in data platforms and AI, yet the last mile of translating governed data into everyday business outcomes remains largely manual," said Sanjeev Mohan, principal analyst at SanjMo, in Snowflake's announcement. "Project SnowWork represents a meaningful shift from AI as an analytical tool to AI as an execution layer embedded directly into enterprise workflows."
The caveat is that SnowWork is a research preview. Snowflake is explicitly limiting access while it gathers feedback and iterates on the architecture. The pre-built persona profiles are pre-configured, not learned from customer data — time-to-value depends on how well those defaults match actual workflows. The "orchestrate across Salesforce and ServiceNow" part suggests integrations beyond Snowflake's own data layer, but the announcement does not detail what those integrations look like or how the security model extends across third-party systems. That is the gap between the demo and production, and it is not a small one.
What Snowflake is really arguing is that the data platform is the right place to build the agentic execution layer — because the data is already there, already governed, already connected to the systems that matter. Build the orchestration plane on top of the data plane, not alongside it. Whether that thesis holds depends on whether enterprises want their AI agents executing across SaaS tools from inside their data warehouse, or whether they would rather have application-layer agents that happen to query the data platform when needed.
The research preview will answer some of those questions empirically. What is already clear is that the "last mile" framing — translating governed data into business outcomes without the manual coordination tax — is the actual product claim. Everything else Snowflake announced is infrastructure to make that claim work. The execution layer is the bet. Data groundedness is the moat. Whether it holds is the story.
Snowflake (NYSE: SNOW), headquartered in Bozeman, Montana, is a data platform company. Project SnowWork is currently in research preview. This article was written March 25, 2026.