Snowflake Inc. today launched agentic artificial intelligence capabilities that allow users to query combinations of structured and unstructured data using enhanced versions of its Cortex Analyst and Cortex Search query tools in combination with the Claude 3.5 Sonnet large language model from Anthropic PBC.

AI agents are autonomous systems that perceive their environment, process information and take actions to achieve specific goals. They use a combination of machine learning, natural language processing and reasoning techniques to make decisions and recommendations with minimal human supervision.

That makes them a step beyond generative AI, said Christian Kleinerman, Snowflake’s executive vice president of product. “The one-question-one-response pattern may have been sufficient for some tasks, but once you can do multistep reasoning-based workflows, you can start to dream of what is possible,” he said.

Trust equation

Snowflake said its foray into agentic AI is predicated on delivering high accuracy levels, trusted results and governed access to data across a variety of sources. Trust has been a major factor in enterprise deployment of AI in general because models are nondeterministic, meaning they don’t always deliver the same outputs in response to the same inputs. Accenture Plc’s 2025 Tech Vision report found 80% of business leaders believe trust needs to be part of their organizations’ technology strategy.

Cortex Agents, a feature now in public preview, orchestrates structured and unstructured data sources ranging from Snowflake tables to PDF files, breaks down complex queries, retrieves relevant data and generates answers.

Agents use Cortex Analyst to operate on structured data while Cortex Search mines through unstructured sources. LLMs analyze and generate answers.

In a demonstration, Snowflake executives showed how the agentic system can answer questions such as “How many insurance claims are currently open and how many are $10,000 or higher?”. Processing that query requires extracting data from unstructured documents and converting them into a structured form for a SQL operation.

Executives said Cortex Agents can automatically determine the best way to retrieve information. For example, a request to list all the appraisal clauses related to snowmobiles across all insurance contracts automatically generates a Cortex Search function that retrieves the information across multiple data sets.

Agents can analyze different possibilities before selecting an option if queries are ambiguous and can split a single request into multiple subtasks and assemble partial responses into a single result. They route across multiple semantic models and search indices and can be configured to generate SQL from natural language.

Beyond text-to-SQL

“Unlike typical text-to-SQL systems that rely on pattern matching, Cortex Analyst uses a semantic model that maps business terms to underlying data,” said Baris Gultekin, Snowflake’s head of AI. “Customers can ask question in natural language and Cortex Analyst translates that into SQL and provides the response.”

Using Claude to aggregate and interpret results has enabled Snowflake to achieve 90% accuracy on text-to-SQL tasks, the company said. In a media briefing, executives praised Anthropic’s commitment to safe and responsible AI deployment.

Anthropic’s model “simplifies the experience and completely aligns with the philosophy we have at Snowflake to help customers achieve what they want to do and not wrestle with infrastructure,” Kleinerman said.

Anthropic Chief Product Officer Mike Krieger said his company’s LLM is designed to respect the privacy of customer data. “We have no access to the data; the inference happens close to the data in a way that is very secure,” he said.

That message has resonated with enterprises, Krieger said. “Having a model that has the right safeguards in place, is hard to jailbreak and has been trained responsibly is actually a net plus that it is enhancing the trust at the deployment side.”

Functional upgrades

Snowflake also announced enhancements to Cortex Analyst, Search and Observability. A new advanced JOIN validation mitigates issues such as JOIN hallucinations and double-counting and enables Analyst to support multitable queries without compromising precision.

A public preview of the new Snowsight web interface simplifies the process of building and refining semantic models. Administrators can select tables and columns and use LLMs running within Snowflake’s secure perimeter to generate a starting semantic model file in YAML, a data serialization format used for configuration files and data exchange between programming languages.

Cortex Search can now index tables with more than 100 million rows. Infrastructure optimizations have reduced serving costs by 30%. Users can now choose to perform multilingual searches using the snowflake-arctic-embed-l-v2.0 or Voyage AI Innovation Inc.’s voyage-multilingual-2 vector embedding model.

Enhancements to AI Observability evaluate the performance of Cortex Agents and applications using techniques such as LLM-as-a-judge, in which LLMs evaluate decisions or outputs independently or as an aid to human decision-making.

AI Observability can also now report metrics such as relevance, “groundedness” and harmfulness to help in refining and improving agent performance.  Users can compare evaluation runs side-by-side and assess the quality and accuracy of responses across different LLM configurations. Traces can log every step of agent executions across input prompts, tool use and final response generation.

Image: Snowflake