Introduction
AI-assisted development is rapidly changing how data teams build and deploy analytics solutions.
Modern coding agents can understand schemas, generate implementation plans, create code, and accelerate workflows that traditionally require significant manual effort.
To evaluate these capabilities in a practical setting, I used Snowflake CoCo (Cortex Code) to build an end-to-end analytics solution involving data integration, semantic modeling, Snowflake Intelligence, and a Streamlit-based application. The objective was to combine data from multiple source tables, create a semantic layer that supports natural-language analytics, connect it to Snowflake Intelligence, and expose the solution through a user-friendly application.
What made the exercise particularly interesting was not just the speed of development, but the insight it provided into where AI can meaningfully accelerate engineering workтАФand where human expertise remains indispensable.
What is Snowflake CoCo?
Snowflake CoCo (Cortex Code) is SnowflakeтАЩs native AI coding agent. Unlike general-purpose coding assistants, CoCo operates within the Snowflake ecosystem and has awareness of the userтАЩs environment, including schemas, objects, and permissions.
This contextual awareness allows CoCo to generate platform-specific implementations, recommend development approaches, and assist with tasks ranging from SQL development and semantic modeling to application creation. Rather than acting as a standalone chatbot, CoCo functions as an AI development assistant embedded within the Snowflake environment.
Getting Started with Snowflake CoCo
Snowflake CoCo can be accessed through both the Snowsight interface and the command-line interface (CLI). While Snowsight is useful for exploration and experimentation, the CLI provides additional flexibility for development workflows.
Installation is straight forward:
curl -LsS https://ai.snowflake.com/static/cc-scripts/install.sh | sh
Configuration is managed through the Snowflake connection file:
default_connection_name = "DEMO"
[connections.DEMO]
account = "BRITHI-XXXXX"
user = "your_username"
password = "your_PAT"
role = "YOUR_ROLE"
Launch:
cortex -c DEMO
As an initial validation step, I asked CoCo a simple question: тАШWhat databases do I have access to?тАЩ The response accurately reflected the available objects and permissions, confirming that the agent had contextual awareness of the environment.
The Three Prompts That Built the Solution:
The objective was to build a complete analytics workflow consisting of:
- A unified data layer created from multiple source tables
- A semantic view capable of supporting natural-language analytics
- A Snowflake Intelligence integration
- A Streamlit application for end-user interaction
Step 1: Creating a Unified Analytics Layer
Prompt:
I have sales transactions in BRONZE_TRANSACTIONS, customer data in BRONZE_CUSTOMERS, and product information in BRONZE_PRODUCTS.
Join these into a Silver table called SILVER_SALES_UNIFIED. Normalize date formats, handle null revenue values with 0 defaults, and add a SOURCE_LOADED_AT timestamp.
Before execution, CoCo presented a detailed implementation plan outlining the objects to be created, assumptions being made, and transformations to be applied. This review step proved valuable because it provided transparency before any code was executed.
The generated solution included a Dynamic Table, three-way joins across source datasets, data type standardization, null handling logic, and audit metadata. More importantly, it demonstrated how planning and implementation could be accelerated without sacrificing visibility into what was being created.
Step 2: Building the Semantic Layer
Prompt:
Create a semantic view called SV_SALES_ANALYTICS over SILVER_SALES_UNIFIED. Support questions such as тАШTotal revenue by region last quarterтАЩ, тАШTop 10 products by units soldтАЩ, and тАШMonth-over-month customer growthтАЩ. Then configure a Snowflake Intelligence data source using it.
This stage highlighted an important distinction between generating code and defining business meaning. Semantic modeling requires decisions about measures, dimensions, grain, time hierarchies, and reporting logic.
CoCo generated the semantic view, documented its assumptions, and identified areas where additional business clarification was required. In this case, questions around multi-currency aggregation required human input before implementation could proceed.
This was an important reminder that while AI can accelerate implementation, business definitions and reporting logic remain human responsibilities.
Step 3: Creating the User Experience
Prompt:
Create a Streamlit app connecting to the Snowflake Intelligence agent on SV_SALES_ANALYTICS. Users type natural-language questions, view results as tables, and see visualizations.
CoCo generated the Streamlit application, including the user interface, API integration, response handling, and visualization components. The application was deployed within Snowsight and provided an intuitive conversational analytics experience.
What stood out was the reduction in effort required to move from raw data to a functioning analytics application. Activities that would typically span multiple development stages were completed through a guided AI-assisted workflow.
Lessons Learned
The speed of delivery was impressive, but it also raised an important question: what role does the developer play when much of the implementation is generated by AI?
One lesson became immediately clear: prompt quality directly impacts output quality. Early experiments with broad instructions produced technically valid but generic results that required significant rework. More precise prompts consistently produced better outcomes.
Another lesson involved reviewing generated plans. In one case, a semantic modeling decision resulted in a discrepancy between analytics outputs and an existing dashboard. The issue was ultimately traced back to an assumption that had been approved without sufficient validation. The experience reinforced that AI-generated plans still require careful review and understanding.
Similarly, while CoCo was highly effective at helping troubleshoot issues, it could only assist after a problem had been identified and described. Human awareness, validation, and troubleshooting skills remain critical.
Where AI Coding Agents Still Depend on Human Expertise
Business context cannot be inferred from schemas alone. AI agents can understand structures and relationships, but they do not inherently understand organizational nuances, historical data quality issues, evolving metric definitions, or business-specific reporting requirements.
Semantic modeling remains a business decision. Questions involving revenue definitions, returns processing, reporting granularity, and currency conversion require domain expertise that cannot be derived solely from metadata.
Effective troubleshooting also depends on understanding how systems work. AI can assist with debugging, but practitioners still need the ability to identify root causes, validate assumptions, and frame problems accurately.
Key Risks and Governance Considerations
Organizations adopting AI coding agents should also consider several important risks.
Overconfidence in generated outputs: AI-generated recommendations often appear authoritative, even when assumptions may be incorrect for a particular business context.
Hallucinated references: Generated code should still be validated. Plausible-looking object names, columns, or implementation approaches may not always align with reality.
Prompt injection and security considerations: Organizations should evaluate how AI agents interact with unstructured or sensitive data sources.
Skill atrophy: As AI handles more implementation work, teams should continue developing foundational technical skills to maintain engineering depth and troubleshooting capability.
Governance and responsible data usage: Access controls alone are not sufficient. Organizations should ensure AI-generated solutions align with privacy, compliance, and business requirements.
Conclusion
Snowflake CoCo demonstrated how AI-assisted development can significantly accelerate the creation of data pipelines, semantic models, and analytics applications. The productivity benefits are real, particularly for repetitive implementation tasks.
At the same time, the experience highlighted that successful adoption depends on more than technical capability. Business context, architectural judgment, governance, validation, and domain expertise remain critical. The most effective use of AI coding agents is not to replace developers, but to enable them to focus on higher-value activities such as design, decision-making, and problem solving.
As AI-assisted development continues to evolve, organizations that balance automation with strong engineering and governance practices will be best positioned to realize its full potential.