“No good AI strategy without data strategy”: Snowflake’s Jeff Hollan

At the recent Snowflake Summit, Jeff Hollan, Head of Apps and Developer Platform, laid out a vision for how the cloud data company aims to embed AI deeper into enterprise workflows. Speaking to a room of conference atendees, Hollan introduced Snowflake Intelligence, a ready-to-run business application designed to provide ChatGPT-style interactions, but specifically grounded in an organization’s internal data.

The pitch is clear: make corporate data as easy to query as a Google search, but with enterprise-grade security and control. What Snowflake is actually building promises to push AI as a central component of enterprise decision-making.

Making Enterprise Data Conversational

The core problem Snowflake Intelligence addresses is accessibility. As Hollan noted, tools like ChatGPT have raised expectations for how people interact with information. “I love ChatGPT,” he said. “But when I come to work and ask questions about Snowflake customers, support cases, or usage metrics, it has no idea.”

Since standard LLMs don’t have access to structured enterprise data,  Snowflake Intelligence looks to bridge that gap, combining natural language interfaces with direct, secure access to company databases. The product leverages Anthropic’s Claude models, running inside Snowflake’s infrastructure, to provide answers that are traceable, role-governed, and context-aware.

In internal testing with Snowflake’s sales team, agents built into the platform could return meaningful, personalized insights: complete with charts and recommended actions in about 10–15 seconds.

Beyond Text Documents: AI SQL and Data Complexity

One differentiator, Hollan emphasized, is that Snowflake Intelligence doesn’t just operate on documents. It understands and generates SQL, allowing users to ask complex business questions that require querying structured data.

To support this, Snowflake launched features under the “AI SQL” banner, such as AI Aggregate for summarizing data and AI Join for fuzzy entity matching (e.g., recognizing “Facebook,” “Meta,” and “FB” as the same company). These features aim to make it easier for organizations with messy data, especially in manufacturing or financial sectors: to apply AI without first undergoing months of data wrangling.

To that end, Snowflake is investing in semantic modeling. Instead of requiring perfect schemas, the system can now infer business models from things like Tableau dashboards or query history. This could shrink the data prep phase from several months to just weeks, or eventually, days.

Hallucinations and Human Oversight

While language models often struggle with accuracy (what the industry calls “hallucinations”,) Snowflake claims a significantly lower error rate. Hollan said Snowflake Intelligence retrieves relevant data 90–95% of the time before generating an answer, compared to ChatGPT’s estimated 60%.

Still, hallucinations aren’t entirely avoidable. To manage them, Snowflake provides observability dashboards that score answers based on relevance and groundedness. Importantly, the system shows how it arrived at each response through SQL queries or cited documents, and allows teams to verify or override outputs. For more sensitive use cases, like finance or healthcare, Snowflake’s position is clear: AI should be a co-pilot, not an autopilot.

Interoperability Over Isolation

Rather than building a monolithic solution, Snowflake is betting on extensibility. The company is aligning with open protocols like MCP (Model Context Protocol) and Google’s agent-to-agent standards, enabling Snowflake Intelligence to cooperate with external agents such as Microsoft Copilot or Salesforce’s Einstein.

In practice, that means a marketing campaign might originate in Adobe’s agent, while Snowflake supplies the underlying demographic insights. “We don’t need to do everything,” Hollan said. “But when it comes to pulling data, that’s where we want to be the best.”

Monetizing a Smarter Marketplace

A lesser-known element of the strategy is the Snowflake Marketplace. Here, organizations can license premium data, such as news from AP or Gannett, via Cortex knowledge extensions. Rather than brokering exclusive one-off agreements, Snowflake is promoting a more scalable, ecosystem-first approach.

Snowflake also announced a partnership with Canva for content creation tools, although Hollan admitted some challenges around multimodal consistency still remain.

From Platform to Product

Snowflake Intelligence builds on years of internal development. Hollan noted that features like Cortex Analyst, Cortex Search, and semantic views all laid the foundation. Intelligence, in this context, is more of a wrapper: a UI and orchestration layer that makes AI functionality usable for the average employee.

There’s no extra pricing for Snowflake Intelligence; it runs on standard consumption-based billing. A demo version for journalists, featuring synthetic data, was hinted at and may arrive at future summits.

Staying Grounded in Data

As for the bigger question: whether Snowflake is becoming an AI company, Hollan was clear. “Our mission is still to help organizations unlock the full potential of their data,” he said. “AI just happens to be the most powerful way to do that right now.” In other words, Snowflake is building AI where it adds leverage and sticking to its strength: data.