Snowflake queries that report to Slack and email

The warehouse holds the numbers; the people who need them are in Slack and email. The Snowflake integration closes that gap: a workflow runs a SQL statement against your warehouse, an AI agent reads the result and writes the two sentences that matter, and the message lands where the decision gets made, on a schedule, on a webhook, or when someone asks the bot. The surface is one general operation, which covers SELECTs for reporting and INSERT or MERGE for writing results back, and a human approval is the right gate for AI-written SQL and for anything that writes to the warehouse. Describe it in plain English and MadMax writes the statement, binds your connection and runs the query during the build so you see real rows before it is scheduled.

What teams build

• A weekday 7 am schedule runs the KPI query with run_query, an AI agent with a fixed schema writes the headline, three movers and one risk, and Slack gets the post; an if step on the revenue delta also texts the finance lead when the drop crosses your threshold.
• A Slack @-mention ("how many signups last week by plan?") has an agent write a read-only statement from the view schema in its prompt, an approval shows the SQL to an analyst, and on approve the rows come back summarized in the thread.
• A 2 am schedule counts null customer ids and duplicate orders, and a switch routes a clean result to a one-line all-good post and anything else to an AI summary plus a Linear issue with the failing counts.

How it connects

Snowflake connects with key-pair authentication, the method its SQL API recommends now that password logins are being phased out. Enter the account identifier (the subdomain of your Snowflake URL), the username, and the RSA private key in PKCS#8 PEM format whose public key is registered on that user with ALTER USER … SET RSA_PUBLIC_KEY. Set a default warehouse (Snowflake will not run SQL without one) plus optional database, schema and role, any of which a step can override. Statements use ? placeholders with typed bindings so values are never interpolated, and a query that outlasts the 45-second timeout returns a statement handle to page with a web request. The key is stored encrypted and never shown again; see managing connections.

Related

• Schedule trigger
• AI agent step
• Approval step
• Managing connections
• Postgres