Snowflake Intelligence and how it works in practice

What is Snowflake Intelligence

Snowflake Intelligence is an agentic application built directly into Snowflake and has been production-ready since November. In short: a business user asks a question in natural language, the agent understands the company domain, and returns an answer.

The key difference compared to anything you know from chatbots: the agent works simultaneously with structured and unstructured data. Tables sit next to emails, contracts, meeting notes, or call center transcripts.

How is it different from a dashboard

Snowflake Intelligence does not replace dashboards. It solves the part of the work dashboards were never designed for.

For analysts → for everyone. No SQL knowledge. No waiting for a Jira ticket. Ask a question in everyday language and get an answer in seconds. Instead of answering tickets all day, the data team can focus on work that actually matters.

Static → conversational. A dashboard shows what someone prepared in advance. An agent answers questions the dashboard was never prepared for.

Only numbers → numbers plus context. A dashboard shows that NPS (Net Promoter Score, a standard customer satisfaction metric) dropped. The agent explains why — because it also searches complaints and exit interviews and returns context.

What makes up the agent

For all of this to work reliably, Snowflake combines five components. Each has one clear role.

Cortex Agent is the orchestrator. It receives a question, splits it into subtasks, and decides which tool to use for each part. It plans, calls tools, and returns the final answer together with an explanation of how it reached it.

Cortex Analyst translates natural language into SQL. A question like “How many clients did we acquire in Q1 2026?” becomes a query over analytical tables, and the answer comes back with an automatically generated chart.

Cortex Search provides semantic search over unstructured text — emails, complaints, contracts, meeting notes. It finds not only keyword matches but meaning.

Custom tools are custom functions the agent can call. Typical examples include sending results by email, sending notifications to Teams, creating a ticket in a ticketing system, or triggering a dbt job that recalculates data.

The semantic layer is the ‘map of the company in business language’ — logical tables, dimensions, metrics, and synonyms. Without it, the agent would have to guess what ‘active customer’ or ‘churn rate’ means. With the semantic layer, it understands.

How it looks in practice

Imagine you’re a branch network manager. You open a chat and ask: “How is the Brno branch performing?” The agent responds with numbers, a chart, and quotes pulled from customer complaints — all in a single answer.

That’s exactly the scenario we built at Revolt BI using a fictional retail bank, CentralBank Digital (CZ, SK, AT). The agent combines structured data — branches, clients, transactions, loans, NPS, and churn — with unstructured sources: call center transcripts, complaints, exit interviews, and internal process documents.

What it can show:

1. A classic metric. “Which branches have the lowest NPS?” — Cortex Analyst generates SQL, returns the numbers, and automatically adds a chart:

2. A combined query. Now the interesting one: “The Brno–Královo Pole branch has the lowest NPS. What do complaints and customer call transcripts from this branch say?” The agent splits the task into two parts. Cortex Analyst extracts the numbers and Cortex Search finds the context.

The result is a complete answer that would take an analyst an hour to prepare — with numbers and context in one place. And the agent returns it in seconds. No ticket. No waiting.

And this is exactly the shift. It is not a better dashboard. It is the first tool that handles both ‘why is this happening?’ and ‘what should we do about it?’ in a single step.

Sharing results with the team — Artifacts

Without Artifacts, every result remained a one-off output. Now you can reopen it, refresh it, or share it.

Since March 2026, Snowflake added the Artifacts feature to Snowflake Intelligence (currently in preview). Every result can be saved as an Artifact — a persistent object that remembers a chart or table.

Importantly, sharing respects existing permissions: a colleague only sees the data they already have access to in Snowflake. And when they ask follow-up questions, the agent remembers the original context.

What is needed to build an agent in your company

Snowflake Intelligence is not a service you buy today and run tomorrow. It is an additional layer on top of a well-prepared data foundation. Key steps:

1. Have data in Snowflake. Structured data in tables, unstructured data (emails, documents, transcripts) as text.
2. Build a semantic layer for the domain the agent should cover — table definitions, relationships, metrics, and synonyms. This is the most important step of all — answer quality depends on it.
3. Create a Cortex Search index over unstructured sources. Snowflake also provides multilingual embedding models — a practical option for Czech-language data.
4. Build the agent in Snowsight UI or through SQL — connect the semantic layer, search service, add instructions, and sample questions.
5. Set up permissions. The agent inherits existing RBAC rules. Every user only sees what they are allowed to see in the underlying data — no parallel permissions world.
6. Iterate on quality. The first version of an agent is never perfect. Synonyms, sample values, metric settings, and validated query formulations are gradually improved.

When Snowflake Intelligence makes sense

Snowflake Intelligence fits companies where:

• data already is (or soon will be) in Snowflake and primary sources of truth are clearly defined,
• there is repeated frustration that business users wait for the BI team for questions they would prefer to answer themselves,
• real questions require a combination of numbers and context — complaints, emails, contracts,
• governance is a hard requirement and solutions must inherit existing access policies.

But good data foundations are required. Without a semantic layer, without clear ownership of business definitions, and with inconsistent data quality, the agent answers unreliably — and every wrong answer reduces the trust it needs for adoption.

Key Takeaways

Snowflake Intelligence does not bypass the data team. On the contrary, it shifts them from reactive ‘build me a report’ work into a role where they define the semantic layer, data quality, and agent boundaries. Dashboards stay where they are useful — recurring overviews and operational management. But questions like ‘why’ and ‘what should we do’ will increasingly end with the agent, not in an analyst’s inbox.

Daša Evelin Kiss

Senior Data BI & Data Engineer with hands-on experience in analytics, data engineering, and business reporting gained in large organizations.

---

Skilled in Snowflake, dbt, SQL, and Power BI, with practical experience building and managing data pipelines and dashboards. Also focused on Snowflake Intelligence and developing AI agents for data analytics. Strong emphasis on data quality and data-driven decision-making.

Thinking about an agent on top of your Snowflake data?

Thinking about an agent on top of your Snowflake data? Just reach out or schedule a short consultation — we’ll review your data together and show you what an agent in your domain could look like.