Snowflake Cortex AI: What It Actually Does and When to Use It

Why Cortex AI Matters

Here is the problem most data teams face. You have valuable data sitting in Snowflake. You want to use AI on that data. So you extract it, send it to an external API, wait for results, and load them back. That workflow is slow, expensive, and creates data governance headaches. Every API call is a potential data leak. Every extraction step adds latency and complexity.

Cortex AI eliminates that entire loop. You write SQL. The AI runs inside Snowflake. Results land in your tables. Your existing Snowflake access controls and policies apply automatically. No external APIs to manage. No data egress charges. No new security reviews.

That matters a lot for regulated industries where data cannot leave the platform. It also matters for teams that just want to get AI into production without building a separate ML infrastructure stack.

The Core Cortex AI Functions

Cortex AI functions went generally available in November 2025. Each function is designed for a specific AI task and can be called directly in SQL. Let us walk through them one by one.

AI_COMPLETE: General-Purpose Text Generation

This is the most flexible function. You give it a prompt and it returns generated text. You pick the model. You can use it for summarization, content generation, question answering, or any open-ended text task.

SELECT
  ticket_id,
  AI_COMPLETE(
    'claude-opus-4-6',
    'Summarize this customer support ticket in two sentences: ' || ticket_text
  ) AS summary
FROM support_tickets
WHERE created_date = CURRENT_DATE();

You can also pass system prompts and structured options for more control over the output. The function supports every model available in Cortex, so you can use a cheaper, smaller model for simple tasks and reserve the larger models for complex reasoning.

AI_CLASSIFY: Text Classification

Need to categorize text into predefined buckets? AI_CLASSIFY handles that. You provide the text and an array of categories, and it returns the best match. No training data required.

SELECT
  feedback_id,
  feedback_text,
  AI_CLASSIFY(
    feedback_text,
    ARRAY_CONSTRUCT('billing issue', 'product bug', 'feature request', 'praise', 'other')
  ) AS category
FROM customer_feedback;

This works surprisingly well for most classification tasks. For a tagging pipeline that previously required training a custom model, AI_CLASSIFY can get you 80 to 90 percent accuracy with zero training effort. If you need higher accuracy for a specific domain, consider fine-tuning, but start here.

AI_SENTIMENT: Sentiment Analysis

Returns a sentiment score between -1 (very negative) and 1 (very positive). Straightforward and effective for tracking customer sentiment at scale.

SELECT
  review_id,
  review_text,
  AI_SENTIMENT(review_text) AS sentiment_score,
  CASE
    WHEN AI_SENTIMENT(review_text) > 0.3 THEN 'Positive'
    WHEN AI_SENTIMENT(review_text) < -0.3 THEN 'Negative'
    ELSE 'Neutral'
  END AS sentiment_label
FROM product_reviews
WHERE review_date >= '2026-01-01';

AI_TRANSLATE, AI_EXTRACT, and AI_TRANSCRIBE

AI_TRANSLATE converts text between languages. Pass the source text, the source language, and the target language. It handles dozens of language pairs.

AI_EXTRACT pulls structured information from unstructured text. Give it a text blob and tell it what to extract (names, dates, amounts, product codes), and it returns structured output. Great for parsing invoices, contracts, or any document where you need specific fields.

AI_TRANSCRIBE converts audio to text. If you store audio files in Snowflake stages, you can transcribe them directly without sending data to an external speech-to-text service.

-- Translate customer feedback from Spanish to English
SELECT
  AI_TRANSLATE(feedback_text, 'es', 'en') AS translated_feedback
FROM international_feedback
WHERE language_code = 'es';

-- Extract key fields from invoice text
SELECT
  AI_EXTRACT(
    invoice_text,
    OBJECT_CONSTRUCT('vendor_name', 'string', 'total_amount', 'number', 'due_date', 'date')
  ) AS extracted_fields
FROM raw_invoices;

Supported Models: Choosing the Right One

Cortex does not lock you into a single model. You choose per function call, which means you can optimize for cost and quality on a task-by-task basis. Here is what is available as of early 2026:

• Snowflake Arctic: Snowflake's own open-source LLM. 480 billion parameters total with 17 billion active at inference time (mixture-of-experts architecture). Licensed under Apache 2.0. Good for cost-sensitive workloads where you need decent quality without premium model pricing.
• Claude Opus 4.6 (Anthropic): Strong at reasoning, code generation, and long document analysis. One of the best options for complex analytical tasks.
• GPT-5.2 (OpenAI): General-purpose model with broad capabilities. Good default choice for most text tasks.
• Pixtral Large (Mistral AI): Available since April 2025. Strong multimodal capabilities for tasks that involve both text and images.

The practical advice: use Arctic or a smaller model for high-volume, simple tasks like sentiment analysis and classification. Save the premium models for tasks that need deeper reasoning, like summarizing complex legal documents or generating nuanced analysis. This keeps costs manageable. For a broader look at how AI fits into your data stack, see our AI/ML services overview.

Cortex Search: RAG Without the Plumbing

Retrieval-augmented generation (RAG) is the standard pattern for building AI search over your own data. The typical approach involves setting up a vector database, building an embedding pipeline, creating a retrieval layer, and connecting it all to an LLM. It works, but it requires significant engineering effort.

Cortex Search packages all of that into a managed service. You point it at your data, define a search service, and it handles the embedding, indexing, and retrieval automatically. When a user asks a question, Cortex Search finds the relevant documents and uses them as context for generating an answer.

This is particularly valuable for internal knowledge bases, customer support documentation, product catalogs, and any scenario where users need to search through large volumes of unstructured text. You get RAG functionality without building and maintaining the infrastructure yourself.

Cortex Analyst: Natural Language to SQL

Cortex Analyst lets business users query structured data using natural language. Instead of writing SQL, they type questions like "What were our top 10 products by revenue last quarter?" and Cortex Analyst generates and executes the SQL.

Under the hood, it uses a semantic model that you define. The semantic model describes your tables, columns, relationships, and business terminology. This is crucial because it gives the AI the context it needs to generate accurate SQL. Without it, the AI would have to guess what "revenue" means in your schema.

Cortex AISQL, announced alongside Analyst, extends this capability to more complex analytical queries. It is designed to handle multi-step analytical questions that require joins, aggregations, and window functions.

A word of caution: natural language to SQL is not magic. It works well for straightforward queries against well-modeled data. Complex ad-hoc analysis still benefits from a skilled analyst writing SQL. Think of Cortex Analyst as a tool that handles 70 percent of routine questions so your data team can focus on the hard 30 percent. If your data engineering foundation is solid, the results are much better.

Cortex Code: The AI Coding Agent

Announced on February 3, 2026, Cortex Code is Snowflake's AI coding agent. It lives inside Snowflake Notebooks and helps developers write, debug, and optimize SQL and Python code.

What makes Cortex Code different from general-purpose AI coding assistants is context. It understands your Snowflake environment. It knows your schemas, your tables, your query history, and your access patterns. When you ask it to write a query, it can reference your actual table structures instead of guessing.

Early use cases we see being most valuable: generating boilerplate transformation code, debugging slow queries by analyzing query profiles, and helping junior engineers understand complex existing queries. It will not replace experienced data engineers, but it can significantly speed up routine development work.

When to Use Each Cortex Feature

Here is a practical decision framework:

• Batch text processing (sentiment, classification, translation): Use the task-specific functions (AI_SENTIMENT, AI_CLASSIFY, AI_TRANSLATE). They are optimized for these tasks and cost less than using AI_COMPLETE with custom prompts.
• Custom text generation or complex prompts: Use AI_COMPLETE with the appropriate model. Choose the model based on task complexity and cost tolerance.
• Search over documents and unstructured data: Use Cortex Search. It saves you from building RAG infrastructure from scratch.
• Business user self-service analytics: Use Cortex Analyst. Pair it with a well-defined semantic model for best results.
• Extracting structured data from text: Use AI_EXTRACT. Ideal for parsing invoices, contracts, emails, and similar documents.
• Development acceleration: Use Cortex Code in Snowflake Notebooks for writing and debugging SQL and Python.

Governance and Security

One of the strongest selling points of Cortex AI is that it inherits Snowflake's existing security model. Role-based access control, column-level masking, row access policies, and network policies all apply to Cortex AI functions. If a user does not have access to a column, they cannot run AI functions on that column. Period.

This is a massive advantage over external AI solutions where you have to build a separate access control layer. With Cortex, the governance you already have in place just works. For teams operating under strict cost and governance requirements, this simplifies the compliance conversation significantly.

Practical Tips from Our Projects

We have deployed Cortex AI across several client environments. Here is what we have learned:

• Start with AI_SENTIMENT and AI_CLASSIFY. They are the easiest to deploy and provide immediate value. Running sentiment analysis across your customer feedback takes one SQL query and produces actionable insights on the same day.
• Use smaller models for bulk processing. Running Claude Opus 4.6 on a million rows of simple classification is expensive and unnecessary. Arctic handles these tasks well at a fraction of the cost.
• Invest time in your Cortex Analyst semantic model. The quality of natural language to SQL output depends entirely on how well you describe your data. Spend time on clear column descriptions, business term definitions, and relationship mapping.
• Monitor token usage. Cortex AI billing is per-token. Long input texts (like full documents) consume more tokens. Consider truncating or summarizing inputs before passing them to expensive models.
• Combine with Snowflake Tasks for automation. Schedule Cortex AI functions to run on new data automatically. For example, classify and tag every new support ticket as it arrives using a Snowflake Task that runs every 15 minutes.

Key Takeaways

• Cortex AI brings LLM capabilities directly into Snowflake SQL, eliminating the need for external AI APIs and data movement
• Task-specific functions (AI_SENTIMENT, AI_CLASSIFY, AI_TRANSLATE, AI_EXTRACT, AI_TRANSCRIBE) are optimized and cost-effective for their respective tasks
• AI_COMPLETE gives you flexible text generation with your choice of model, including Arctic, Claude Opus 4.6, GPT-5.2, and Pixtral Large
• Cortex Search provides managed RAG for unstructured data search without building vector database infrastructure
• Cortex Analyst enables natural language to SQL, with quality depending on your semantic model definition
• Cortex Code (announced February 2026) adds AI-assisted coding in Snowflake Notebooks
• All Cortex AI features inherit Snowflake's existing security and governance controls