Snowflake Cortex & Databricks Agent Bricks

The Data Gravity Argument

Most banks have spent years -- and tens of millions of dollars -- building modern data platforms. Whether your institution runs Snowflake, Databricks, or both, your analytics infrastructure already holds the cleaned, governed, and compliance-ready data that AI applications need.

Moving that data to a separate AI platform introduces risk, cost, and governance overhead. Snowflake Cortex AI and Databricks Agent Bricks (formerly Mosaic AI) take the opposite approach: they bring AI capabilities to where the data already lives.

Snowflake Cortex AI

Snowflake Cortex AI adds foundation modelFoundation ModelA large AI model trained on broad data that can be adapted to many tasks. Examples include OpenAI's GPT, Anthropic's Claude, Google's Gemini and Meta's Llama families. Banks evaluate these for capabilities, safety, and regulatory fit.See glossary capabilities directly into the Snowflake Data Cloud. You invoke AI functions using SQL -- the same language your analysts and data engineers already use. Models from several providers run inside Snowflake, including OpenAI's (through a $200 million partnership announced in February 2026) and Anthropic's.

Core Capabilities

Cortex AI Functions. Call foundation models directly from SQL queries, enabling AI processing inline with data transformations. You can summarize text columns, classify records, extract entities, and generate insights without leaving Snowflake.

Cortex Search. A managed retrieval-augmented generationRetrieval-Augmented Generation (RAG)A pattern that combines document retrieval with LLM generation. The system searches a knowledge base for relevant context, then feeds it to the model to produce grounded, accurate answers.See glossary service that creates searchable indexes over your Snowflake data. Point it at a table of documents, and it handles chunking, embedding, and vector storageVector DatabaseA specialized database optimized for storing and querying high-dimensional vectors (embeddings). Enables fast similarity search across millions of documents for RAG and recommendation systems.See glossary automatically. Your applications query it with natural language and get grounded answers backed by your data.

Fine-tuning. Fine-tuneFine-TuningThe process of further training a pre-trained model on a specific dataset to specialize its behavior for a particular domain or task, such as banking compliance language.See glossary supported models on your proprietary data without the data leaving Snowflake's security boundary. At its June 2026 Summit, Snowflake also announced Cortex Training, managed GPUs for fine-tuning open models inside the platform.

Cortex Analyst. A natural language interface to structured data. Business users describe what they want to know in plain English, and Cortex Analyst generates and executes the SQL -- democratizing data access for non-technical stakeholders.

Cortex Agents and Snowflake CoWork. Cortex Agents (generally available November 2025) let you build agents that work over your Snowflake data. Snowflake Intelligence, the business-user agent built on them, was renamed Snowflake CoWork in June 2026.

Cortex AI Guardrails (generally available May 2026) add safety controls to these agents. Snowflake has also announced a Cortex AI Gateway for governing agents, but as of October 2026 it is announced only, not yet an available control.

Banking-Specific Value

For banks already running Snowflake for analytics and reporting, Cortex AI offers three significant advantages:

Zero data movement. Your customer data, transaction records, and regulatory reports stay in Snowflake. AI processes them in place, eliminating the data copy, transfer, and secondary governance that separate AI platforms require.

Existing access controls. Snowflake's role-based access control extends to Cortex AI functions. If an analyst cannot see PII columns in a table, they cannot pass those columns to an AI function. Your existing data governance policies apply automatically.

SQL interface. Your data engineering team does not need to learn Python, new frameworks, or AI-specific tooling. SQL-callable AI functions integrate into existing ETL pipelines, reporting workflows, and data products.

Databricks Agent Bricks

In March 2026 Databricks made Agent Bricks the umbrella name for all its AI capabilities; the older "Mosaic AI" label survives only on some components, such as Mosaic AI Model Serving. Databricks provides a full MLOps platform with LLM serving and agents alongside traditional machine learning. Where Snowflake emphasizes SQL simplicity, Databricks emphasizes flexibility and control.

Core Capabilities

Model Serving. Deploy foundation models and custom-trained models as scalable API endpoints within your Databricks environment. Supports both pay-per-token and provisioned throughput pricing, with automatic scaling based on demand.

Unity AI Gateway. A governed front door to multiple inferenceInferenceThe process of running a trained model to generate predictions or outputs from new input data. Inference cost, latency, and throughput are key factors in enterprise AI deployment.See glossary endpoints -- including external providers alongside self-hosted models -- with cost monitoring, access controls, and agent tracing. The Unity Catalog registry now covers agents, tools and models, not just data.

Vector Search. Native vector database capabilities built into the Databricks Lakehouse. Create vector indexes over Delta tables and perform similarity search without a separate vector database service -- your embeddings live alongside your structured data.

Knowledge Assistant. Agent Bricks' managed document question-and-answer agent, with citations -- the successor to earlier RAG tooling. Since March 2026 it is available by default to workspaces that use Databricks' compliance security profile with HIPAA controls selected.

MLflow Integration. All AI experiments, model versions, and deployments are tracked through MLflow -- the open-source ML lifecycle platform that Databricks created. This provides the model registry, experiment tracking, and deployment management that model risk management teams require.

Banking-Specific Value

Unified ML and LLM platform. Banks running traditional ML models (credit scoring, fraud detection) on Databricks can add LLM workloads and agents to the same platform. One governance framework, one model registry, one monitoring system for all AI.

Open-source foundation. Databricks is built on open formats -- Delta Lake, MLflow, Apache Spark. Banks concerned about vendor lock-in find comfort in the portability of their data and model artifacts.

Lakehouse architecture. The Databricks Lakehouse unifies structured data (transaction records, financial metrics) with unstructured data (documents, emails, call transcripts) in one platform. AI applications that need both types of data -- which describes most banking use cases -- benefit from this unified architecture.

Government-grade compliance. Databricks holds FedRAMP High authorization on AWS GovCloud.

Comparing the Two Approaches

When Data Platform AI Makes Sense

Data-platform-adjacent AI is strongest when:

1. Your data is already in the platform. If your loan portfolio, customer records, and regulatory data live in Snowflake or Databricks, running AI there avoids the cost and risk of data movement.

2. Your AI use cases are data-centric. Summarizing database records, classifying transactions, extracting entities from document tables -- these map naturally to SQL-callable AI functions.

3. Your team is data-engineering-heavy. If your strongest technical talent is in SQL and data pipelines, Snowflake Cortex meets them where they are. If you have a strong data science team, Databricks Agent Bricks gives them the flexibility they expect.

4. Governance is your top concern. Keeping AI processing within your existing data governance boundary -- with the same access controls, audit logging, and compliance infrastructure -- reduces the incremental risk of AI adoption.