Sigma’s AI is designed to stay grounded in your governed warehouse data and metric definitions, which lowers the risk of “made up” answers compared to general-purpose chat tools. When you ask a question in Ask Sigma, it generates a workbook that shows the tables, columns, filters, and calculations behind the result. That means you can inspect the logic, adjust it, and confirm it before you act. If the data or a required metric definition isn’t available, Sigma should surface that instead of guessing.

No. Ask Sigma doesn’t just return a one-line response. You get an editable workbook that shows what it did. You can review the tables it used, how it filtered the data, and what formulas or calculations produced the output. If something looks off, you can change it, or pass it to an analyst to validate. The goal is simple: the person reading the answer should be able to verify it.

It depends on the model you choose and how your admins configure Sigma. Sigma is model-agnostic, so you decide which provider to use. If you use a warehouse-native option like Snowflake Cortex or Databricks-hosted models, AI processing can stay within that platform’s environment. If you choose an external provider like OpenAI or Azure OpenAI, Sigma sends the minimum required inputs to that provider under your organization’s enterprise configuration. Sigma doesn’t route to an external model by default without customer setup.

Sigma doesn’t force a single LLM. Administrators configure the provider in Sigma’s AI settings, and Sigma can work with warehouse-native options as well as external providers. If your organization already has an approved model list or an existing enterprise AI contract, Sigma can be set up to align with it rather than introducing a new provider.

Yes. Sigma’s AI features follow the same access rules that already apply to the user asking the question. In practice, that means the AI can only answer using data the user is authorized to see, including row-level controls and masking patterns enforced in the warehouse. There isn’t a separate “AI permission model” that bypasses your governance. This applies to embedded use cases too, where AI is scoped to the embedded user’s identity and permissions.

MCP (Model Context Protocol) is a standard for connecting AI agents to external tools and data sources in a consistent way. Sigma is building MCP support in two directions. First, Sigma can connect to third-party MCP servers so Ask Sigma can use approved context beyond the warehouse when needed, such as documents or application context in embedded scenarios. Second, Sigma can expose its own capabilities as MCP tools, so external AI clients can call Sigma to query warehouse data, generate analysis, and support governed workflows. MCP support is currently in private beta; your Sigma account team can confirm availability.