Snowflake CoCo: How To Make AI Work Across Your Data Workflow

Part 1 of the Snowflake Changes That Matter Series

Analysts often ask themselves the same questions:

1. Why are these reports not matching?

   
   

2. Where is that data located again?

   
   

3. What does this view do?

   
   

4. Why is my code not running?

These are not edge cases. They are the daily reality of analytics work, from data validation and report creation to SQL generation and one-off questions from business users.

When a C-Suite client asks for a number by the end of the day, the pressure is on. You need to locate the data, write the query, validate the result, and do it fast. The answer used to be doing all of that manually, relying on your own experience and whoever was sitting next to you.

Now, for Snowflake teams, one answer is Snowflake CoCo, Snowflake’s AI coding agent formerly known as Cortex Code.

Like any LLM, Snowflake CoCo promises faster data discovery, better natural language answers, and more efficient development workflows. It is a useful new feature, but it is still an LLM. That means it can sometimes give you the wrong answer and still charge you credits.

So, how can you make Snowflake CoCo work for you and your team?

Start With the Right Context

Snowflake CoCo is built to increase developer productivity across the full data workflow, from data engineering and pipeline development to data science and analytics. At a basic level, teams can use it to generate SQL, debug code, document existing logic, or answer questions about unfamiliar objects in their Snowflake environment.

The key is context. The more information you give Snowflake CoCo upfront, the better the output. That means going beyond the immediate ask and giving it the full picture: the table DDL, the downstream goal, business definitions, or an example of the expected output.

The difference between a vague prompt and a well-contextualized one is often the difference between a starting point and something you can actually run.

Snowflake CoCo lives in Snowsight, but the Cortex Code CLI also runs directly inside your local IDE terminal. That means engineers do not have to leave their primary development environment to do Snowflake-heavy lifting.

Use Snowflake CoCo for Governance, Not Just SQL

One of Snowflake CoCo’s more underappreciated use cases is routine governance work.

Scanning for PII, generating data classification SQL, and surfacing privilege gaps used to require hours of manual effort. Snowflake CoCo can help kick off that work with a single prompt.

For organizations with established governance standards, custom skills can encode your specific naming conventions, classification profiles, and compliance requirements directly into Snowflake CoCo’s workflows. That means Snowflake CoCo is not just generating generic SQL. It is working within your organization’s guardrails.

That said, SQL compilation errors happen. The output is a strong first draft, not a finished product. Always have a data engineer validate the logic before any classification logic is applied to production data.

The time savings are real. Human sign-off is still required.

Build Toward a Stronger Semantic Layer

For teams ready to go deeper, one of the highest-value investments is using Snowflake CoCo to accelerate your semantic layer build.

The semantic layer is the foundation for more accurate AI-generated answers in Cortex Analyst and Snowflake’s broader agentic workflows. Without it, these tools are working from raw table structures and trying to infer your business logic. With it, they can use defined metrics, relationships, and business context to generate more reliable answers.

In one Snowflake Builders Blog example, semantic layer work supported text-to-SQL accuracy of 90% or higher (Wells, Snowflake Builders Blog, April 2026). That does not mean every team will see the same result, but it reinforces the broader point: semantic context matters.

If your organization already has data models and metrics spread across BI tools like Power BI, Tableau, or Looker, Snowflake CoCo can help you rationalize that sprawl rather than starting from scratch.

If the full build feels overwhelming, start with the 20% of your data that answers 80% of your team’s questions. From there, use Snowflake CoCo to help generate the semantic views themselves.

More on building the semantic layer in a future post.

Watch the Snowflake Roadmap

If you are not ready for a full semantic layer investment, there are still plenty of entry points.

At the June 2026 Snowflake Summit, Snowflake announced a Skill Catalog with generally available, pre-built skills for data engineering and AI workflows. Teams can also build their own custom skills to help Snowflake CoCo answer questions using repeatable, standardized procedures.

Snowflake is also expanding where teams can use Snowflake CoCo and how it can support broader workflows. Other Summit updates worth noting for your roadmap include:

• Expanded support for desktop and mobile applications, with integrations across tools such as Slack, Excel, and VS Code.

  
  

• Autonomous workflow execution through native AI services called Cloud Agents and Automations.

  
  

• A Skill Catalog that enables reusable, shareable workflows across your organization.

  
  

• A Snowflake CoCo plug-in available for Anthropic’s Claude Code.

Manage Snowflake CoCo Costs Before the Bill Arrives

This is the part most teams skip until the bill arrives.

Snowflake CoCo is billed by token consumption. Every prompt, every response, and every tool call in an agentic workflow is metered.

One thing that surprises people: longer sessions may be cheaper per turn than shorter sessions. Snowflake caches context that does not change between turns, billing repeated context at roughly 10% of the normal input rate (Reddy, Medium, April 2026). Breaking a workflow into many short sessions can cost more than running one longer conversation.

Spending limits are necessary.

Snowflake supports separate rolling 24-hour per-user limits for Snowflake CoCo CLI and Snowsight, and budgets can be used to track shared spend by team or cost center. Start with a conservative daily cap and adjust based on usage patterns.

If you are rolling this out to a broader team, Snowflake also provides dedicated usage views to track AI consumption by function, model, user, and query. You can build alerts or automated controls on top of them.

If you are not monitoring usage, the invoice will do the teaching for you.

Where to Start

Next time you are staring down one of those familiar questions — why are these reports not matching, where is that data, why is my code not running — remember that the answer does not have to start with you.

Give Snowflake CoCo context first, even if that just means pointing it to the right view before you ask. Set spending limits before rolling it out to additional users. Build toward a semantic layer as your longer-term foundation.

Snowflake CoCo is a genuinely useful tool for the kind of work analysts and engineers do every day. But like any LLM, it works best when you have done the work to make its operating environment as clear and governed as possible.

Start narrow, prove the value, then expand.