Snowflake Cortex is a set of AI functions that run inside Snowflake — LLM functions (COMPLETE, SUMMARIZE, EXTRACT_ANSWER, SENTIMENT), Cortex Search (vector + hybrid), Document AI, and Cortex Analyst (text-to-SQL). No data leaves your warehouse to reach the model.
This page is part of Milemarker's Snowflake cluster for financial services. See the Snowflake for Financial Services pillar for the full landscape, or explore related capabilities: Snowflake Data Sharing and the Snowflake Marketplace.
Why AI Belongs in the Warehouse, Not Beside It
The standard architecture for AI in financial services has been: extract data from your systems, send it to an external API, receive the model's output, and store it somewhere. This pattern is operationally simple but creates a governance gap that regulators and compliance teams are increasingly focused on. The moment data leaves your controlled environment to reach a model endpoint, it is outside your security boundary, outside your audit trail, and subject to the data handling policies of whoever operates that endpoint.
Data Movement Is a Governance Failure
For wealth management firms, the sensitivity of the data makes this gap unacceptable at scale. Client names, account values, social security numbers, and investment profiles are all potentially in scope for any AI workflow that touches client records. Sending that data to an external LLM API — even a reputable one — means that data has left your environment. Your SOC 2 auditors want to know where regulated data goes. Cortex's answer is that it never goes anywhere: inference runs inside Snowflake, on data that never leaves your account.
The AI-readiness conversation in wealth management typically focuses on data quality and normalization. But AI readiness also means architectural readiness — having an inference layer that respects your security boundary. Cortex is that layer for firms already on Snowflake.
What Cortex Can Do for a Wealth Firm
Cortex functions are callable as SQL expressions, which means any team with access to Snowflake can apply AI to the data they already work with. The following represent the highest-value applications for wealth management operations and advisory practices.
Pre-Meeting Client Briefings
Summarize the last 12 months of CRM interaction notes, account changes, and life events into a structured briefing before an advisor's client review. The advisor walks into the meeting with context that would otherwise take 30 minutes to manually compile.
IPS Extraction from PDFs
Apply Document AI to ingest investment policy statements, extract structured data — target allocations, prohibited securities, risk tolerance bands — and populate your compliance monitoring database automatically. IPS updates propagate into compliance rules without manual rekeying.
Email and Note Sentiment
Run SENTIMENT scoring on incoming client communication to surface at-risk relationships before formal complaint or attrition. Advisors see a rolling sentiment score for each client household — a data point that no manual review process can produce at scale.
Semantic Research Search
Index firm-generated research, analyst commentary, and external reports using Cortex Search. Operations and advisory staff search for answers in natural language across thousands of documents — without reading every file or knowing which documents contain the relevant section.
Risk Anomaly Explanations
When a portfolio triggers a risk alert — concentration breach, drawdown threshold, factor exposure limit — use COMPLETE to generate a plain-language explanation of what changed and why, ready for advisor or compliance review without waiting for a quant team to write the narrative.
Natural-Language Reporting
Use Cortex Analyst to let operations, compliance, and advisory staff ask questions about firm data in plain English. Questions like "what is our average household AUM by advisor this quarter?" return accurate, SQL-generated answers without requiring anyone to write a query or open a BI tool.
The entire Cortex API is standard SQL.
Cortex functions call directly from SELECT statements, CTEs, and stored procedures. Teams that already know SQL can apply AI to their data immediately — no Python environment, no API client library, no external service to configure.
SELECT household_id,
SNOWFLAKE.CORTEX.SUMMARIZE(
LISTAGG(note_text, ' | ') WITHIN GROUP (ORDER BY note_date)
) AS client_briefing
FROM crm_notes
WHERE note_date >= DATEADD('month', -6, CURRENT_DATE())
GROUP BY household_id;
Cortex Search and Vector Search
Traditional keyword search fails on financial content because the vocabulary of wealth management is imprecise. A note that says "the client is cautious about market exposure" and a query for "conservative risk tolerance" share no keywords but convey the same concept. Cortex Search solves this by representing text as vectors in high-dimensional space — similar meanings cluster together regardless of exact word choice.
RAG Patterns Over Firm Content
Retrieval-augmented generation (RAG) is the architecture that makes Cortex Search operationally useful. Instead of asking an LLM to answer from its training data alone, RAG retrieves the most relevant documents from your corpus, feeds them into the prompt as context, and asks the model to answer from that context. The model's response is grounded in your firm's actual data rather than general world knowledge.
For wealth management, RAG over Cortex Search enables use cases that closed-corpus LLMs cannot: answering questions about a specific client's situation using that client's actual records, explaining a portfolio decision using the firm's own research at the time, surfacing relevant precedent from the firm's compliance history. See the AI agents in wealth management page for how these patterns connect to orchestrated workflows.
Hybrid Search
Cortex Search supports hybrid search that combines vector similarity with keyword matching. This is the right approach for financial content where precision matters alongside recall. A search for "Regulation Best Interest documentation for the Henderson account" benefits from vector similarity (understanding the regulatory concept) and from keyword matching (returning results that actually mention the Henderson account specifically). Hybrid search provides both without requiring two separate search systems.
Cortex Analyst: Text-to-SQL for Non-Technical Teams
One of the persistent limitations of analytics in wealth management is that the people who most need answers — advisors, compliance officers, operations managers — are not the people who can write SQL. The result is a queue of data requests that flows through analytics or IT teams, adding days or weeks to the cycle time for basic business questions.
Cortex Analyst addresses this directly. It accepts a natural-language question, consults a semantic model that maps business concepts to your warehouse schema, generates a SQL query, executes it, and returns the result. The semantic model is the critical piece: it defines what "AUM" means, what "client" means, what "advisor" means in the context of your specific data model — eliminating the ambiguity that makes naive text-to-SQL unreliable.
Democratized Analytics Across the Firm
When Cortex Analyst is deployed against Milemarker's wealth management data model, advisors and operations staff can ask questions like "which clients have not had a scheduled review in the last 12 months?" or "what is the aggregate equity allocation for clients over 70?" and receive accurate, SQL-generated answers in seconds. The analytics team shifts from being a query production service to being a semantic model maintainer — a much higher-value role.
Explore how this capability fits into a broader wealth management data platform and the Snowflake for RIAs and Snowflake for asset managers use cases.
Cortex vs. External LLM APIs
This is an honest comparison. Cortex and external LLM APIs each have genuine strengths, and the right architecture for most firms is not one or the other — it is a combination, with Cortex handling the majority of high-volume, governed workflows and external APIs handling tasks that require frontier model capability.
| Dimension | Cortex (In-Warehouse) | External LLM API |
|---|---|---|
| Data governance | Data stays inside Snowflake; no egress; audited by Access History | Data sent to external endpoint; subject to provider's data retention policy; not audited in Snowflake |
| Data residency | Inference in the Snowflake region where your data lives | Data routed to provider's inference infrastructure; region may differ from your data region |
| Model capability | Curated open-weight and proprietary models; strong for summarization, extraction, classification, RAG | Access to frontier models (GPT-4o, Claude Opus, Gemini Ultra); stronger for complex reasoning and nuanced generation |
| Cost model | Snowflake credits per million tokens; no separate API contract; no egress cost | Per-token pricing from API provider; plus data egress cost; plus engineering to manage API client |
| Latency | Comparable to external API for batch; may be slower for real-time single-prompt use cases | Low latency for interactive single-prompt; scales well with streaming |
| Interface | Standard SQL — no additional SDK or environment required | REST API or SDK; requires application code to call and parse results |
| Audit trail | Built-in — every Cortex call in Access History with user, time, and data accessed | Depends on provider's logging; not integrated with Snowflake access logging |
The practical guidance: use Cortex for high-volume, data-intensive workflows — summarizing thousands of client notes, scoring sentiment across millions of records, extracting structured data from large document sets. Use external APIs for interactive, high-judgment tasks — generating investment narrative for a specific client, drafting a complex compliance letter, synthesizing a research argument from multiple conflicting sources. Milemarker Navigator is designed precisely for this combination, using Cortex for warehouse-native operations and Claude for reasoning tasks that benefit from frontier model capability.
Where Milemarker Fits
Milemarker Navigator runs Claude-powered AI agents inside Snowflake using Cortex and Snowpark. This is a partner-positioned, not competitive, relationship with Snowflake — Milemarker uses the Cortex layer for in-warehouse functions (summarization, extraction, vector search) and external Claude inference via Snowpark for tasks that require more sophisticated reasoning, while keeping the data model and orchestration inside the Snowflake security boundary.
The practical implication for a wealth firm: deploying Milemarker's platform means that Cortex capabilities are available against Milemarker's pre-normalized wealth data model from day one. Firms do not need to build their own semantic layer, their own vector index, or their own document processing pipeline. The Snowflake for Financial Services infrastructure and the AI layer land together.
Pre-Built Semantic Model
Cortex Analyst is deployable against Milemarker's wealth data model without custom semantic model authoring. Wealth-specific concepts — household, advisor, AUM, sleeve — are pre-defined.
Cortex Search Index
Milemarker includes a configurable Cortex Search index over CRM notes, client documents, and firm research as part of the platform deployment.
Document AI Pipelines
Pre-built Document AI workflows for common wealth management documents — IPS, client agreements, custodian statements — extract structured data directly into the normalized data model.
Navigator Agent Orchestration
Milemarker Navigator orchestrates Cortex functions alongside Claude inference via Snowpark, giving firms access to both in-warehouse AI and frontier model capability from a single platform.
Governed by Design
Every AI function in the Milemarker platform — Cortex or external — operates under Snowflake's RBAC framework. No AI workflow bypasses the access controls that govern the underlying data.
Milemarker Augments, Not Replaces
Milemarker is designed to extend Snowflake's capabilities for wealth management — providing the data model, integration library, and AI orchestration layer that makes Snowflake and Cortex immediately productive.