What is Snowflake Cortex

Direct Answer

Snowflake Cortex is the managed AI and machine-learning layer built directly into the Snowflake Data Cloud, letting teams run large language models, vector search, and ML functions with SQL or Python — without moving data out of the warehouse. For RevOps, it means you can summarize call transcripts, score lead intent, and forecast pipeline against your governed CRM and product data in the same place that data already lives.

The pitch is simple: bring the model to the data instead of shipping data to the model.

1. What Cortex Actually Is

Snowflake Cortex is a suite of AI features inside Snowflake that splits into two broad buckets. Cortex LLM Functions are serverless calls to hosted large language models — including Anthropic Claude, Meta Llama, Mistral, and Snowflake's own Arctic model — invoked through functions like SNOWFLAKE.CORTEX.COMPLETE(), SUMMARIZE(), SENTIMENT(), TRANSLATE(), and EXTRACT_ANSWER().

Cortex ML Functions are pre-built, SQL-callable machine-learning routines for forecasting, anomaly detection, and classification that any analyst can run without writing model code.

On top of those primitives sit higher-order products. Cortex Search is a hybrid (vector plus keyword) retrieval engine for building RAG applications. Cortex Analyst is a text-to-SQL service that lets business users ask questions in natural language and get governed answers against a semantic model you define.

Cortex Agents orchestrate retrieval and SQL generation across structured and unstructured data.

The architectural point that matters: everything runs inside the Snowflake security perimeter. Your data is not copied to a third-party API, and role-based access controls (RBAC), masking policies, and governance you already configured still apply. Snowflake competes here with Databricks Mosaic AI, Google BigQuery ML, and Amazon Redshift ML — the common thesis across all four is that the warehouse, not a separate ML stack, becomes the place AI happens.

2. Why It Matters for RevOps

RevOps lives on the seam between Salesforce, HubSpot, product telemetry, billing systems, and finance models. Historically, doing AI against that data meant an ETL pipeline to a notebook, a data-science backlog, and a governance review every time someone wanted to score accounts. Cortex collapses that loop.

Concrete plays: pipe Gong or Chorus call transcripts into a Snowflake table and run SUMMARIZE() plus SENTIMENT() to flag deals where buyer sentiment dropped between calls. Use CLASSIFY_TEXT() to tag inbound leads by ICP fit against your own taxonomy. Run FORECAST() on closed-won history to produce a bottoms-up pipeline projection that sits next to rep-submitted commit numbers.

Build a Cortex Analyst semantic layer so a CRO can ask "what's net revenue retention by segment this quarter" and get a number that ties to the same governed definitions finance uses.

The leverage is time-to-insight and trust. Because the answer is computed against governed tables, you avoid the spreadsheet-of-record problem where six teams have six different definitions of qualified pipeline. One semantic model, many questions.

3. The Core Building Blocks

Break Cortex into the pieces a RevOps team will actually wire up:

• LLM Functions — task-specific (SUMMARIZE, SENTIMENT, TRANSLATE, EXTRACT_ANSWER) and general (COMPLETE against a chosen model).
• Cortex Search — managed embeddings plus a retrieval service for grounding answers in your documents, contracts, and playbooks.
• Cortex Analyst — natural-language-to-SQL governed by a YAML semantic model that names your metrics and join logic.
• ML Functions — time-series forecasting, anomaly detection, and classification callable in plain SQL.
• Document AI — extract structured fields from PDFs like order forms and signed contracts.
• Snowflake Copilot — an in-product assistant that writes and explains SQL.

The unifying mechanic is that each of these is a function call billed in credits, not a separate platform to provision. Mistral, Llama 3, and Claude are available as model arguments — you change one string to swap the underlying model.

4. Pricing and the Credit Model

Cortex bills on Snowflake credits, the same consumption unit as compute. LLM functions are priced per token processed, converted into credits per the Snowflake Consumption Table; ML functions and serverless features bill on the compute they consume. There is no separate seat license — cost scales with usage.

The operator implication: a runaway COMPLETE() loop over a million rows can generate a real bill. Treat Cortex like any variable-cost compute — set resource monitors, batch where possible, and cache results in tables rather than re-summarizing the same transcript nightly.

Anthropic and Meta model choices carry different per-token rates, so model selection is a cost lever, not just a quality one.

5. Limitations and Honest Caveats

Cortex is not a silver bullet. Model availability is region-dependent — not every model runs in every Snowflake cloud region, and cross-region inference must be explicitly enabled. Cortex Analyst quality is only as good as your semantic model; vague metric definitions produce confidently wrong SQL.

The LLM functions are stateless — there's no built-in memory, so multi-turn agents need orchestration you build.

You also inherit general LLM risks: hallucination on ungrounded prompts, and the need to keep humans in the loop for anything customer-facing. And while data stays in-perimeter, you should still confirm data-residency and model-provider terms with your security team. Databricks advocates will note their model-serving flexibility; the honest trade-off is that Cortex favors simplicity over open-ended MLOps control.

Central Model

Frameworks at a Glance

• Cortex LLM Functions — serverless SQL/Python calls to hosted models
• Cortex ML Functions — forecasting, anomaly detection, classification
• Cortex Search — hybrid vector + keyword retrieval for RAG
• Cortex Analyst — governed natural-language-to-SQL via semantic model
• Document AI — structured extraction from contracts and forms
• Snowflake Consumption / Credit Model — token-based variable pricing
• RBAC + Masking Policies — inherited governance perimeter

Operating Loop

FAQ

Do I need to be a data scientist to use Cortex? No. The ML and LLM functions are callable in standard SQL, so any analyst comfortable writing a query can run forecasting or summarization. Higher-order tools like Cortex Analyst are designed for business users asking plain-language questions.

Does my data leave Snowflake when I call an LLM? No. Inference runs inside the Snowflake perimeter, and your RBAC and masking policies still apply. This is the core differentiator versus piping data to an external OpenAI or Anthropic API directly.

Which models can I use? Snowflake hosts Anthropic Claude, Meta Llama 3, Mistral, and Snowflake's own Arctic, among others. You select the model as a function argument, and availability varies by cloud region.

How is it different from Databricks? Both put AI next to data. Databricks Mosaic AI leans toward open MLOps and custom model training; Cortex favors managed, SQL-first simplicity. Your choice usually follows whichever platform already holds your data.

What does it cost? Cortex bills in Snowflake credits — LLM functions per token, ML and serverless features per compute consumed. There's no per-seat license, so set resource monitors to control variable spend.

Bottom Line

If your governed revenue data already lives in Snowflake, Cortex lets you run AI against it without a second platform or an ETL detour. Start narrow: one Cortex Analyst semantic model for pipeline metrics, plus SUMMARIZE and SENTIMENT over your call transcripts. Watch the credit burn, prove one decision it improved, then expand.

Sources

• Snowflake Documentation — Cortex LLM and ML Functions (docs.snowflake.com)
• Snowflake — Cortex Analyst and Cortex Search product pages
• Snowflake Consumption Table (Service Consumption Table, 2024)
• Gartner — Magic Quadrant for Cloud Database Management Systems 2024
• Anthropic — Claude model documentation
• Meta AI — Llama 3 model card
• Databricks — Mosaic AI platform documentation (for comparison)
• Forrester Wave: Data Management for Analytics 2024
• Snowflake Summit 2024 keynote announcements
• A16z — "Emerging Architectures for LLM Applications" (Andreessen Horowitz)