How to Set Up Sigma Computing AI Query

What Sigma AI Query Does and Why It Matters

Sigma Computing AI Query lets analysts run large language model functions directly inside a spreadsheet-style workbook. Instead of exporting data to a notebook, writing Python, and importing results back into your BI tool, you type a formula in the Sigma formula bar. That formula calls an AI model hosted on your cloud data warehouse and returns the result into a new column.

The practical effect: a customer support team can classify 50,000 tickets by sentiment in a single column formula. A finance team can summarize quarterly earnings call transcripts without leaving the same workbook where they track revenue. All compute stays on your warehouse. No data leaves your infrastructure.

Sigma announced $200M in annual recurring revenue on April 8, 2026, doubling year-over-year. AI Query and AI Agents were cited as the fastest-adopted features driving that growth, with over 1.1 million new active users added in the past fiscal year.

Prerequisites Before You Start

You need three things in place before writing your first AI Query formula.

First, a cloud data warehouse that supports AI functions. Snowflake Cortex, Databricks Foundation Models, and BigQuery ML all qualify. The AI models run on warehouse compute, so your warehouse plan must include access to these services.

Second, the service account Sigma uses to connect to your warehouse must have permission to call AI functions. In Snowflake, this means granting the CORTEX_USER database role to the Sigma service account. In Databricks, the account needs access to the ai_query function. Check your warehouse documentation for the exact permission grants.

Third, a Sigma organization with AI features enabled. An admin must go to Administration, then Account, then AI Features, and toggle on the relevant options. This is a one-time setup per organization.

Step 1: Connect Your Warehouse to Sigma

If you already have a warehouse connection in Sigma, skip to Step 2. Otherwise, log into Sigma and navigate to Administration, then Connections. Click Add Connection and select your warehouse type. Enter your connection credentials: account identifier, warehouse name, database, schema, and the service account credentials.

Test the connection before saving. Sigma will validate that it can reach your warehouse and execute queries. If the test fails, check that your network allows traffic from Sigma's IP ranges and that the service account has the required roles.

Step 2: Enable AI Features in Your Organization

Navigate to Administration, then Account, then AI Features. You will see toggles for several AI capabilities including AI Query, Ask Sigma, and AI Builder. Enable AI Query at minimum. Each toggle controls a distinct feature, so you can enable only what your team needs.

Once enabled, any user with "Can use" access to a warehouse connection can write AI Query formulas in workbooks connected to that warehouse.

Step 3: Write Your First AI Query Formula

Open a workbook connected to your warehouse. Add a new column to any table. In the formula bar, you will use one of Sigma's passthrough functions to call a warehouse AI function.

The passthrough functions determine the return data type. Use CallText when you expect text back (classifications, translations, summaries). Use CallNumber for numeric outputs (scores, counts). Use CallLogical for true/false results. Use CallVariant for structured JSON responses.

Here is a concrete example using Snowflake Cortex. Suppose you have a column called "customer_feedback" with free-text reviews. To classify sentiment, enter this formula:

CallText("snowflake.cortex.sentiment", [customer_feedback])

This sends each row's feedback text to the Cortex sentiment model and returns a label like "positive," "negative," or "neutral" in your new column. The query runs on Snowflake's compute, so processing 100,000 rows might take a few minutes depending on your warehouse size.

For translation, the formula looks like this:

CallText("snowflake.cortex.translate", [french_notes], "fr", "en")

Each passthrough function follows the same pattern: the function name tells Sigma what data type to expect, the first string argument names the warehouse AI function, and the remaining arguments are the inputs that function needs.

Step 4: Classify Unstructured Text at Scale

One of the most practical applications is categorizing unstructured text. Many operations teams have spreadsheets full of support tickets, survey responses, or sales notes that need tagging.

Create a new column and enter a formula like:

CallText("snowflake.cortex.classify_text", [support_ticket], '["billing", "technical", "feature_request", "account"]')

This sends each ticket to the classification model with your category list and returns the best match. You can then use Sigma's built-in pivot tables and charts to visualize ticket distribution across categories, all in the same workbook.

For Databricks users, the equivalent formula uses the ai_query function:

CallText("ai_query", "databricks-meta-llama-3-70b-instruct", Concat("Classify this ticket into one category: billing, technical, feature_request, account. Ticket: ", [support_ticket]))

The syntax varies by warehouse, but the Sigma interface stays the same. You always write in the formula bar, pick the right passthrough function, and point it at your warehouse's AI endpoint.

Step 5: Automate Actions with Sigma Agents

Once you are comfortable with AI Query formulas, Sigma Agents take it further. Agents let you define conditions and thresholds in the spreadsheet interface. When a condition is met, the agent triggers an action: creating a Salesforce opportunity, updating a Jira ticket, sending a Slack alert, or calling a stored procedure.

For example, you could set up an agent that monitors your sentiment classification column. When the 7-day rolling average of negative sentiment exceeds 15%, the agent posts a summary to a Slack channel and creates a Jira ticket for the support lead. You configure all of this in Sigma's UI without writing integration code.

Agent processing compiles to SQL and runs on your warehouse, so it scales with your data and stays within your security perimeter.

Common Errors and How to Fix Them

The most frequent issue is a permissions error when calling the AI function. If you see "insufficient privileges," confirm that the Sigma service account has the correct warehouse role. In Snowflake, run GRANT DATABASE ROLE CORTEX_USER TO ROLE sigma_role in your admin console.

If formulas return null for every row, check that the AI function name matches exactly. Function names are case-sensitive in most warehouses. Also verify that the input column is not empty or null for those rows.

Slow performance usually means your warehouse is undersized for the AI workload. Each row triggers a model inference call. For large tables, consider running AI Query on a filtered subset first, then scaling up after confirming results look correct.

What Existing Guides Miss

Most Sigma documentation covers individual features in isolation. What analysts actually need is the end-to-end flow: connecting a warehouse, enabling permissions, writing formulas, and building something useful. The gap is especially wide for teams using Databricks or BigQuery, where the function syntax differs from Snowflake examples.

If you want to skip the warehouse setup entirely and run AI analysis on uploaded files with no configuration, VSLZ handles that from a CSV upload with a single prompt.

Practical Next Steps

Set up your warehouse connection and enable AI features in Sigma administration. Write passthrough formulas in the formula bar to call warehouse AI models. Start with sentiment analysis or text classification on a small dataset before scaling. Use Sigma Agents to automate actions based on AI Query results. The entire pipeline stays on your warehouse compute with no data movement required.