Ask the warehouse a question in plain English. Get the answer and the query that made it.

The question is simple. Which region fell last month, and was it price or volume? The answer is three days away, because the dashboard does not have that cut and the analyst has a queue.

A model that writes SQL will write it with total confidence, against the wrong table, with the wrong join, and return a number that looks exactly like a right one.

What the business wants is not a chatbot. It is the number, fast, and proof of where it came from. Without the query, an answer from a model is hearsay.

One question, one query, one answer, and the query printed under it.

• The agent writes a single query against a curated layer: the tables, joins and metric definitions your data team has blessed, not the whole warehouse.
• It runs read-only inside your account, with row limits and a time budget. It cannot write, cannot see what its role cannot see and cannot run twice without saying so.
• The answer comes with the query, the row count and the tables it touched. An analyst can rerun it in one click. Every panel on Grid Intelligence prints its Snowflake query ID for the same reason.
• Metric definitions live in one place. Revenue means the same thing to the agent as it does to the finance report, because both read the same definition.
• When a question cannot be answered from the curated layer, the agent says so and names what is missing, instead of improvising a join.

A question set your own analysts wrote, with the answers they already trust, is the exam.

• Before launch, the agent answers a few hundred real questions from your team, and every answer is compared with the number the analyst got. Disagreements are reviewed one by one; most turn out to be a definition, and the definition gets fixed.
• We plant questions it must refuse: a metric not in the layer, a table its role cannot see, a question that needs a join nobody has approved. Each must end in “I cannot answer that from here”, not a number.
• On Grid Intelligence, Snowflake Cortex writes one query, runs it read-only inside the account and answers in two sentences, behind guardrails. The same pattern holds for a sales or finance warehouse.
• Verdict hands the model a structured evidence package and an instruction to translate it, not to speculate. The analytics agent is held to the same rule.

• Manufacturing

  AI for manufacturers, where the count has to be right.

• Financial services

  AI for financial services, built to pass an audit.

• Marketing

  AI for marketing, where one wrong number costs more than it saves.

• Data & Analytics
• Generative AI