Democratizing Enterprise AI: Snowflake’s New AI Capabilities Further Accelerate and Simplify Data-Driven Innovation

At Snowflake Summit 2025, we introduced innovations to make AI easy, efficient and trusted. Fully managed within Snowflake’s secure perimeter, these capabilities enable business users and data scientists to turn structured and unstructured data into actionable insights, without complex tooling or infrastructure.

Whether customers are analyzing tables, processing documents, deploying AI agents or training models, these capabilities are integrated into Snowflake’s secure platform with unified governance, eliminating infrastructure management or complex tooling.

At Summit, we unveiled four major AI advancements:

1. Data agents: We’re offering intelligent capabilities with Snowflake Intelligence (in public preview soon) and Snowflake Cortex Agents (generally available soon). Snowflake Intelligence turns structured and unstructured data into rapid, trustworthy business actions while Cortex Agents orchestrates multistep tasks and integrates into Microsoft Teams or custom apps.

2. Cortex AISQL and analytics: By bringing multimodal data processing using AI into familiar SQL, Cortex AISQL (in public preview) makes complex AI workflows accessible and complements Snowflake's comprehensive unstructured data insights offering. This includes enhanced Document AI with schema-aware table extractions and a new extraction model "Arctic-Extract." This supports extraction from documents in 29 languages, which can be called through our new function ai_extract (public preview soon).

3. Generative AI industry-leading models, evaluation, observability and AI gateway: AI Observability in Snowflake Cortex AI (generally available soon) enables no-code and pro-code monitoring of gen AI apps. Snowflake also provides access to LLMs from OpenAI through Microsoft Azure OpenAI Service, Anthropic, Meta, Mistral and other leading providers, all within Snowflake’s secure perimeter. The AI Governance Gateway provides features that allow customers to easily implement governance polices, including robust access control, granular usage tracking and budget enforcement (private preview coming soon).

4. Machine learning (ML): Build and serve production-ready models with improved scalability and flexibility, including the introduction of the autonomous Data Science Agent to increase productivity across development workflows. 

Together, these launches form a unified AI foundation that simplifies development, scales reliably and preserves trust within Snowflake's governed environment.

1. Data agents: Build agents and analyze multimodal data at scale

Data agents enable self-service insights with Snowflake Intelligence, giving business users a natural language interface to query all their data. Users can ask complex questions and receive governed, explainable answers in seconds without dashboards or SQL.

Cortex Agents and tools such as Cortex Analyst and Cortex Search let developers build trusted, production-ready AI applications that reason across structured and unstructured data, orchestrating workflows with LLMs, SQL and semantic search.

Model Context Protocol (MCP) provides an open standard for connecting AI systems with data sources. We are excited to share that MCP server support will be available (private preview soon) on Snowflake. At launch, developers will be able to serve Cortex Analyst and Cortex Search as tools with the Snowflake MCP server. 

1a. Unlock self-service intelligence for business teams

Snowflake Intelligence gives business users an AI-driven natural language interface to engage with both structured and unstructured data. Users can ask complex questions in plain English and receive explainable, governed answers in seconds — no SQL or dashboards needed. These agents run inside Snowflake’s security perimeter, automatically enforcing role-based access, masking and audit controls. They can reason across enterprise data, identify relationships between diverse data sources and return synthesized answers from tables, PDFs, Jira, Salesforce, Zendesk and more.

1b. Simplify data insights for everyone with data agents

Data agents let nontechnical teams extract insights using natural language. Users can see how insights are generated with automatic charting, lineage traceability and explainability. Developers can deploy new use cases quickly and embed analytics into any app of their choice, accelerating innovation and impact.

1c. Build trusted conversational applications

With Cortex Agents (generally available soon), developers can build gen AI applications that reason over both structured and unstructured data. These agents enable high-quality, explainable results by orchestrating workflows that combine LLMs, SQL and semantic search. Powered by models such as Claude 3.7 Sonnet, OpenAI GPT-4.1 and o4-mini (generally available soon), these agents plan, execute and refine tasks for accurate results. Built-in explainability and API access enable fast deployment and integrations with Microsoft Teams and Copilot, letting users engage with AI directly within their collaboration tools. 

2. Cortex AISQL and analytics: Redefining multimodal data to insights with Snowflake

Unstructured data remains underutilized due to its complexity. Cortex AISQL solves this by enabling teams to analyze documents, images and other formats using familiar SQL syntax, without specialized tools. At Summit, we introduced:

• SQL meets AI: Extract metadata, classify sentiment or search embeddings, all within SQL.

• Value extraction from unstructured data: Use Document AI, now supporting schema-aware table extraction (in public preview), to pull structured tables from complex PDFs with minimal cleanup. 

• Automatic semantic model generation (private preview): Eliminate manual model setup, explore insights with native chart visualization and build branded experiences using Snowpark Container Services.

2a. Turns analysts into AI developers with Cortex AISQL

Cortex AISQL reimagines SQL as the core language for enterprise AI. Its native AI operators let teams build multimodal workflows, combining text, audio, images and structured data, without learning new tools or duplicating data. 

It delivers 30–70% performance improvements over several traditional pipelines (based on internal benchmark results, implemented per normal usage), powered by optimized batch inference and a performance optimization algorithm (private preview), empowering analysts to become AI developers. 

Use cases include record matching, fraud detection and enterprise-scale semantic joins, all written in SQL.

2b. Extract value from unstructured data

We also introduced our next next-gen vision model for Document AI called "arctic-extract" (private preview). It supports a total of 29 languages (including Japanese, Koren, German, French, Spanish and Chinese) and enhanced reasoning capabilities, including classification and normalization.

On the retrieval side, Cortex Search adds:

• Batch fuzzy search for high-throughput tasks such as entity resolution and fraud detection

• Advanced APIs (generally available) for multifield search, scoring and ranking by metadata

• Admin UI in Snowsight (generally available) and Quality Evaluation Studio (public preview soon) for no-code search management, diagnostics and relevance tuning

Teams can also bring custom vector embeddings (public preview) to power Cortex Search, combining Snowflake's secure platform with proprietary model outputs for greater performance and control.

2c. Accelerate data insights with automated semantic models and charts preview

Automatic semantic model generation (private preview) makes creating semantic models for Cortex Analyst easier and faster. By analyzing schema metadata, query history and dashboards, it constructs performant, reusable models, eliminating weeks of manual work. The Charts capability (in public preview) lets users explore insights visually alongside AI results. 

Snowpark Container Services (generally available on AWS and Azure, coming soon to Google Cloud Platform) offers a scalable runtime to host full-stack apps and APIs natively within Snowflake, with centralized logging, governance and security.

3. Generative AI observability, model choice and scalable infrastructure: Deploy AI with confidence

To help organizations scale AI securely and reliably, Snowflake offers: 

• AI Observability: No-code evaluation tools for gen AI accuracy and performance

• Model access: Top LLMs from Meta, OpenAI, Anthropic and Mistral

• Provisioned throughput: Predictable inference performance at production scale

3a. Evaluation and tracing within Cortex AI 

AI Observability (generally available) in Snowsight helps teams measure accuracy and coverage using evaluation data sets. LLM-as-a-judge scoring evaluates groundedness, helpfulness and harmfulness, executed securely inside Snowflake. Features like agent trace logs and model comparisons simplify debugging, prompt refinement and governance.

3b. Access OpenAI, Anthropic and more models — securely on Snowflake

Snowflake’s model ecosystem now includes access to industry-leading LLMs from OpenAI, Anthropic, Meta and Mistral — including the latest generation models such as OpenAI’s GPT-4.1 and o4-mini through Microsoft Azure OpenAI Service in Azure AI Foundry as well as Anthropic’s Claude Opus 4 and Claude Sonnet 4. These models run inside Snowflake’s security boundary, so data remains protected and is never used for training.

Customers can match the best model to each use case, summarization, classification, translation, agentic reasoning and more, without managing infrastructure.

Cortex AI is also expanding to Google Cloud Platform. With Snowpark Container Services on Google Cloud Platform (generally available soon), customers can deploy open source models in GCP regions, avoiding data movement and maintaining governance.

Cortex AI is also expanding to Google Cloud Platform. With Snowpark Container Services on Google Cloud Platform (generally available soon), customers can deploy open source models in GCP regions, avoiding data movement and maintaining governance.

3c. Provisioned throughput for enterprise-ready AI

Provisioned throughput (generally available on AWS and Azure) gives teams dedicated inference capacity for gen AI apps. Accessible via REST API across all Snowflake regions, it enables consistent performance without unpredictability of shared services. It's ideal for moving from POC to production, without the overhead of infrastructure setup. 

3d. AI Governance Gateway: Enterprise control for gen AI

The AI Governance Gateway offers a single pane of glass for customers to access industry-leading LLMs via SQL (or REST APIs) directly within the secure Snowflake perimeter. Through comprehensive role-based access control (RBAC), customers can implement robust governance policies. Granular usage tracking views for each AI feature, combined with budget enforcement controls (in private preview soon), enable customers to monitor and manage generative AI usage across their organizations. Customers can drive responsible AI with Cortex Guard to filter harmful content, while AI Observability allows customers to evaluate, debug and optimize their gen AI apps for accuracy and performance. This improves trust and transparency for production deployments. Cortex AI brings AI to your data, and with AI Governance Gateway, customers can accelerate gen AI application delivery.

4. Models built and operationalized in production with Snowflake ML

Predictive ML is still a critical cornerstone for use cases such as fraud detection, customer segmentation and recommendation engines. However, building and deploying such models often requires stitching together multiple, disparate tools that can be difficult to govern and costly to maintain. 

With Snowflake ML, enterprises now have a modern ML solution that is tightly integrated with governed data across end-to-end workflows. Customers such as Coinbase and Cloudbeds are driving predictive insights. Scene+, a large customer loyalty program in Canada, cut time to production by over 60% and cut costs by over 35% for more than 30 models using Snowflake ML.

At Summit, we continued our rapid pace of innovation with a suite of new announcements focused on scalable, flexible ML, which will allow customers to: 

• Boost data scientist productivity by automating the generation of ML pipelines with Data Science Agent (private preview soon)

• Build production-ready models faster with distributed training APIs in Container Runtime (generally available) and manage training jobs easily with native experiment tracking (private preview soon)

• Easily deploy and orchestrate ML pipelines over Snowflake data, operating from any IDE of choice with ML Jobs (generally available soon)

• Serve features for low-latency, online predictions (private preview soon) on scalable compute from Snowflake Feature Store 

All this is integrated with built-in ML Observability for easy monitoring and alerting with support for custom metrics. 

4a. Bring agentic AI to trusted ML for accelerated productivity.

At Snowflake, we’re committed to making it easy and efficient for all our customers to use the latest industry-leading technologies, including gen AI. We see an untapped opportunity to apply the latest innovations from LLMs to empower data scientists. At Summit, we announced that we are applying agentic AI to turbocharge productivity in predictive ML with a Data Science Agent (private preview soon in AWS) that autonomously iterates, adjusts and generates a fully executable ML pipeline from simple natural language prompts. 

With Anthropic reasoning models running under the hood, Data Science Agent uses multistep planning to break down a problem into distinct steps and chooses the best-performing technique for each phase of the ML workflow, including data preparation, feature engineering and training. 

After quickly generating a verified ML pipeline, teams can provide follow-ups based on their domain knowledge to easily iterate on performance and accuracy for the next-best version. The output is a production-ready, fully functional ML pipeline that can be easily executed from Snowflake Notebooks on Container Runtime. By automating the tedious experimentation and debugging work, data scientists can save hours of manual work and focus on higher-impact initiatives. 

4b. Build and orchestrate scalable ML pipelines on Snowflake data starting from any IDE 

For ML development, we announced a suite of new features that make building models on Snowflake data easier and more powerful from either natively integrated Snowflake Notebooks or any external IDE of choice with convenient pushdown mechanisms. Customers can now easily access Distributed ML APIs in Container Runtime (generally available on AWS and Azure) to accelerate data loading, model training and hyperparameter tuning from any IDE.

As model versions are iterated during training runs, the best-performing model can be quickly identified, shared and reproduced from natively integrated experiment tracking (private preview soon), accessible via APIs or the Snowsight UI. 

To facilitate developing and automating ML pipelines, orchestrated by Snowflake Tasks or external tooling such as Airflow, ML Jobs (generally available soon on AWS and Azure) offers a convenient mechanism to trigger remote ML code execution. Furthermore, the interfaces included in ML Jobs also enable teams that prefer working from an external IDE (VS Code, PyCharm, SageMaker Notebooks) to dispatch functions, files or modules down to Snowflake’s Container Runtime.

4c. Deploy ML models built anywhere at scale 

Regardless of where or how a model is built, it can be logged in Snowflake Model Registry and served for scalable inference on Snowpark Container Services (generally available on AWS and Azure), using CPU or GPU compute for real-time or batch predictions. Customers can also deploy models to a REST API endpoint for low-latency inference applications. This includes support for easy deployment of models from Hugging Face (private preview soon) in one click without downloading any client-side model. By just pointing to the model handle and task for logging and serving in Snowflake, teams can get instant access to the top trained Hugging Face models ranging from image classification to sentence similarity and object detection. 

We also announced that Snowflake Feature Store now also supports the ability to serve features for low-latency, online inference (private preview soon) in addition to batch to support use cases such as personalized recommendations, pricing, anomaly detection and customer service. 

Models storing inference in Snowflake are automatically integrated with Snowflake’s ML Observability (generally available) capabilities to enable reliable predictions over time.

Learn more and view resources

At Summit, we introduced innovations to make enterprise AI easy to use, efficient to deploy and trusted to run. Snowflake Intelligence turns structured and unstructured data into actionable insights, while Cortex Agents orchestrate complex multistep tasks across data. 

Cortex AISQL brings multimodal data processing to familiar SQL workflows, and AI Observability delivers robust monitoring and evaluation tools to scale gen AI applications. For traditional workflows, Snowflake ML is making it easier and more flexible for customers to build and serve models in production.

With these advancements, Snowflake enables organizations to confidently transform their data into intelligent action at scale. 

Get started with Snowflake for AI using the following resources: 

• Productionize ML models: Try building an end-to-end ML workflow from Snowflake Notebooks (available from the free trial) or any IDE of choice

Forward Looking Statements

This article contains forward-looking statements, including about our future product offerings, and are not commitments to deliver any product offerings. Actual results and offerings may differ and are subject to known and unknown risk and uncertainties. See our latest 10-Q for more information.