Snowflake Unleashes AI Agents to Unlock Enterprise Data

Snowflake today announced the public preview of Cortex Agents, a new class of AI agents that can perform a range of business tasks using structured and unstructured data. The company also announced that, thanks to Anthropic’s latest Claude 3.5 Sonnet model, Cortex Agents are turning in best-in-class agent performance.

The potential for AI is great, and companies are investing billions of dollars to gain a competitive advantage using the latest language models. However, those investments will be for naught if they can’t effectively tap into their enterprise data and expose it with AI, according Snowflake’s SVP of Product Christian Kleinerman.

“We believe that a strong data foundation is important for AI,” Kleinerman said in a Snowflake press conference this week. “I have access to a great model. But if I don’t have my data in order, if I don’t know governance, if I don’t know what data sets I have, it is difficult to get value out of AI.”

Snowflake has worked to help customers break down their data silos and get access to the data using their choice of data architecture while preserving the privacy and security of the data, Kleinerman said. The latest example of that comes in the form of Cortex Agents, which Snowflake previously announced but which are now available as a public preview.

Snowflake’s AI applications framework (Image courtesy Snowflake)

Cortex Agents are a new class of generative AI applications that are designed to function like human workers. They can automate multi-step, data-intensive tasks, such as accessing structured and unstructured data, running complex queries, and serve the data back to the users.

There are currently two Cortex Agents agents, including:

Cortex Analyst, which is based on Anthropic’s latest Claude model, provides text-to-SQL conversion and is designed for working with structured data, such as database table and views;
And Cortex Search, which is based on Snowflake’s Arctic AI models, is designed for answering questions on unstructured data, such as PDF or .txt documents.

When called via a REST API, Cortex Agents securely kick off a four-part workflow that involves:

Snowflake claims Cortex Analyst exceeds Databricks and OpenAI in text-to-SQL accuracy (Image courtesy Snowflake)

understanding the request and planning out the specific subtasks;
executing the tasks via Cortex Analyst, Cortex Search, or SQL generation from natural language;
evaluating the effectiveness of the answer and determining whether it meets accuracy requirements;
and monitoring and iterating on the workflow to refine behavior for continuous improvement, Snowflake said.

Cortex Agents are based on Cortex AI, the service Snowflake launched back in November 2023 that lets users run a range of large language model (LLM)-powered tasks in the cloud. In addition to Cortex Search and Cortex Analyst, Snowflake at that time launched Document AI, which lets users ask questions of unstructured data using natural language (from Snowflake’s presentation, it appears the image-based AI capability in Document AI has been subsumed into Cortex Search).

These agents are ready to take on a variety of tasks currently handled by human knowledge workers in areas like brand and product planning; claims processing; supplier management; customer support, insurance underwriting, and order management. Companies currenlty building agents with Snowflake include Bayer, Nissan, Penske, Siemes Energy, S&P Global, and Zoom, according to Snowflake’s presentation.

Security and privacy are key aspects of Snowflake’s AI strategy. By first registering data within Snowflake’s Horizon Catalog, customers can subsequently control access to both structured and unstructured data sets using Horizon’s role-based access control (RBAC) capabilities.

Snowflake’s internal benchmarks for the latest Cortex Search release (image courtesy Snowflake)

In addition to making the AI agents available in a public preview, Snowflake announced some additions. Cortex Analyst has been bolstered with better support for database JOINs, which will improve accuracy with complex queries, the company said, while a new user interface will simplify the process of building and refining semantic models.

Cortex Analyst also supports more customized business logic, such as specifying fiscal year start dates, thanks to the general availability of a feature called Custom Instructions. Finally, Snowflake says that with Anthropic’s Claude 3.5 Sonnet, it has achieved 90% accuracy for text-to-SQL use cases according to internal benchmarks.

The latest release of Cortex Search, meanwhile, offers scalability and affordability improvements. Customers can now index hundreds of millions of rows with Cortex Search, the company said, while serving costs are up to 30% lower thanks to “infrastructure optimizations.” The unstructured data agent offers boasts better customizability, the company said, thanks to updates to the two multilingual models, snowflake-arctic-embed-l-v2.0 and voyage-multilingual-2. Finally, Cortex Search brings support for date-range filtering on metadata columns.

With the right combination of data, AI algoirthms, and compute, Snowflake Cortex AI agents are poised to have a big impact on customer operations, said Baris Gultekin, head of AI at Snowflake.

“At Snowflake, we believe that AI agents will soon be essential to the enterprise workforce,” he said. “They’ll enhance the productivity for many teams such as customer support analytics, engineering, and they’ll free up employee time to focus on higher value things.”

Snowflake Gives Cloud Customers What They Need and Want at Summit 2024

It’s a Snowday! Here’s the New Stuff Snowflake Is Giving Customers

The post Snowflake Unleashes AI Agents to Unlock Enterprise Data appeared first on BigDATAwire.