How Snowflake AI Research Drew 5,200 Customers

Snowflake is expanding its AI capabilities by integrating them more closely with enterprise data and targeting use cases that differ from consumer-facing applications. The cloud-based data storage and analysis company has been at work to bridge AI with trusted enterprise data, allowing businesses to move from ingestion to insights. 

In fact, this opportunity motivated Dwarak Rajagopal to join Snowflake in 2024, after stints at Meta, Google, and Apple, bringing over a decade of machine learning experience. 

In an exclusive chat with AIM, Rajagopal, vice president of AI engineering and research at Snowflake, said that for AI to be effective in enterprises, it needs the right context. He added that current challenges include hallucination and grounding. “One of the key elements is the right context that you provide for the AI, and Snowflake has that a lot,” he said.

Snowflake now counts over 5,200 companies using its AI products weekly, said Rajagopal.“Bringing AI closer to the data makes it much easier for enterprises and industries to use AI as well as data.”

Enterprise AI Challenges

Rajagopal explained that enterprise AI differs fundamentally from consumer AI. Consumer use cases often focus on retrieval from unstructured web data, while enterprise data is a mix of structured and unstructured formats. 

Structured data — such as invoices and financial records — poses unique interpretation challenges. “When you have unstructured data in SQL tables, you don’t know exactly what it means… because the columns and rows could be arbitrary,” he said.

Rajagopal explained that Snowflake’s access to customer query logs and dashboards enables it to combine structured and unstructured data for richer insights. 

Governance is another critical factor. For example, he cited that AI should be secure enough not to disclose details about a boss’s salary when one’s query is about their own salary. Snowflake’s governance features, built into its platform, ensure compliance and privacy by design.

On the unstructured data front, Snowflake has introduced tools such as Cortex Search, Document AI, and extract SQL functions. Rajagopal said that Cortex Agents APIs can route queries to the appropriate tools for processing. “Depending on different kinds of users, you have different products to ingest and understand unstructured data,” he said.

For instance, Snowflake’s recently launched Cortex AISQL framework enhances traditional SQL by expanding it to multimodal AI, allowing analysts to query unstructured data such as documents, images, and audio. 

Specialised Models 

While other AI labs pursue large, general-purpose models, Snowflake focuses on targeted post-training and reinforcement learning to optimise open-source models for specific enterprise tasks. For example, Cortex Analyst, its text-to-SQL model uses reasoning techniques to improve accuracy with business data. 

Snowflake also builds its own embedding models, such as the Arctic Embed model for search and the Tilt model for document AI, optimised for speed and cost.

Rajagopal shared that the company’s research priorities span three main areas. The first is efficiency and scalability, which includes inference optimisation and model cascading to balance performance and cost. The second is agentic applications, focused on building systems that improve reasoning. The third is proprietary models, such as the extract and embed models used across its products.

Snowflake’s approach combines open-source and commercial models, with many of its intelligence products powered by Anthropic and OpenAI. For specific cost and fine-tuning needs, the company also builds on top of Llama and Mistral models. Customers have the flexibility to choose between open-source and whitelisted frontier models.

Integrating Postgres

Rajgopal said that Snowflake’s acquisition of PostgreSQL company, Crunchy Data, reflects its strategy to support a broader set of workloads. He added that Postgres is particularly relevant for agentic applications and smaller, transactional workloads, while Snowflake SQL remains central for analytics and large-scale processing. “We want to have from ingestion to insights across all data aspects,” he explained.

To reduce hallucinations and increase trust in AI outputs, Snowflake is developing features such as verified queries that are pre-approved by administrators and feedback loops that allow users to validate or flag responses. “Building that kind of human-in-the-loop aspect is going to be critical,” Rajagopal said.

The company is also investing in observability and reinforcement learning for enterprise data. By applying these techniques to structured data usage patterns, Snowflake plans to improve model quality for domain-specific tasks like SQL generation.

AI in Internal Operations

AI is embedded not just in Snowflake’s products but also in its internal workflows. Engineers across the company use AI daily for coding, code reviews, and continuous integration tasks. 

Rajagopal shared that Snowflake’s AI tools, such as SnowConvert AI, migrate Oracle SQL and Snowflake SQL, also assist its professional services teams in migrating customer workloads.

Snowflake’s sales teams use Snowflake Intelligence to prepare for client meetings. Rajagopal explained, “Before I talk to a customer, I would like to understand how they are using Snowflake, what challenges they have seen. All of that information comes very quickly, within a matter of minutes.”

The Future of Enterprise AI

Rajagopal believes AI will act as a force multiplier for all employees. “You would do a lot more things, much faster,” he said. He predicts a future where agents perform many workflows, both in consumer and enterprise contexts, driving higher data and compute demand.

For graduates entering the workforce, his advice is to embrace AI and design products that improve as the underlying models advance. “Build for a future where the underlying technology improves,” he said.

In June, Snowflake announced a new R&D centre in India, recognising the country’s developer talent pool. “India has the second highest developers in the world. This is a super critical market for us,” Rajagopal said, adding that the company is in the process of building its presence.

The post How Snowflake AI Research Drew 5,200 Customers appeared first on Analytics India Magazine.