Agentic AI in Action Part 32 Shows That One Name Has Two Definitions and a CoCo Skill Chooses the Correct One
Snowflake CoCo uses a markdown skill and semantic views to choose between Finance and Sales net_revenue definitions, returning 48,100 or 61,800.
Key Takeaways
- The article shows how “net_revenue” can have two valid definitions: Finance’s recognized revenue returns 48,100, while Sales’ net bookings returns 61,800.
- Snowflake semantic views keep each team’s governed metric logic separate, preventing the AI agent from improvising calculations from raw tables.
- A CoCo skill markdown file instructs the agent to ask which definition the user means, then run the matching semantic view query.
- The pattern demonstrates how agentic AI can handle ambiguous business terms by pairing governed data models with explicit clarification rules.
Meta Description: A Snowflake CoCo skill shows how two net_revenue definitions can return 48,100 or 61,800, using semantic views and one markdown instruction file.
Tags: Snowflake, CoCo, semantic views, net revenue, agentic AI
Featured Image: Abstract friendly illustration of two branching paths converging into one highlighted result, no people, no logos, no text.
The core answer: ask first, then use the governed view
For analytics teams using Snowflake CoCo, a plain-language request for net_revenue can hide two correct business definitions. In a post published in Towards AI on Medium, a Snowflake CoCo skill demonstrates how to surface that choice, then return either 48,100 from the Finance view or 61,800 from the Sales view. The demonstration uses capabilities available in CoCo today, while Snowflake’s Business Ontology, designed for this kind of ambiguity, is in private preview. The post also notes the setup script and skill file can be accessed from the blog.
The same metric name, two different tables
Finance and Sales both use the name net_revenue, but they mean different things. Finance means recognized revenue after refunds and discounts, with intercompany invoices and test accounts removed. Sales means bookings net of discounts given at signing. If an agent has no guidance, it may choose one calculation and not explain the choice.
The sample setup starts with a database, one schema for each team, a small warehouse, and one table for each team. FINANCE.INVOICES records one row per invoice, with refunds and discounts recorded against it, plus one intercompany row and one test account row that should never count as revenue. SALES.BOOKINGS records one row per signed deal, with the discount given at signing. The rows are fixed, so the later numbers can be checked.
Two semantic views make the gap visible
The demo creates REVENUE_SV on INVOICES for Finance and BOOKINGS_SV on BOOKINGS for Sales. Both expose a metric called net_revenue, but neither view is wrong; they answer different questions under the same name. REVENUE_SV subtracts refunds and discounts and leaves out the intercompany and test account rows. BOOKINGS_SV uses the booking amount minus the discount given at signing. With the sample data, the two numbers differ by 13,700.
A markdown skill file tells CoCo how to behave
A skill is a markdown file of instructions that CoCo reads and follows when a request matches it. The file used here is saved as SKILL.md from the repo and added through the CoCo window in Snowsight. To load it, open CoCo in a Snowsight workspace, click upload skill file, browse to SKILL.md, then type / in the message box and check that net-revenue-definitions appears in the list.
The skill file has four parts. The front matter gives the skill a name and a description, and CoCo matches the description against the prompt to decide when to load it. The table records the two approved definitions and the semantic view behind each one. The workflow tells CoCo to ask which team the user means, run the matching SEMANTIC_VIEW query, and answer with the number, the definition and its source. The rules stop it from calculating net revenue from the raw tables. The full file in the repo also covers breakdowns by region and month.
What happened when the question was asked
When the user asks “What was the net revenue”, CoCo asks which definition is meant, runs a SEMANTIC_VIEW query, and returns 48,100 labelled governed. That is the Finance team’s number to look for with the sample data. Answering Sales should return 61,800 from the bookings view. The result is useful because the number comes from a governed metric, not from a calculation CoCo improvised, and the answer keeps the definition next to the value.
Why the author says readers shaped this work
The post also shares a personal note: the author was named Winner in the AI Excellence category for Asia Pacific and Japan at the Snowflake Community Awards 2026. The work behind it includes writing the book Ultimate Snowflake Cortex AI for Generative AI Applications, publishing an agentic series of blogs with code on GitHub, building multiple marketplace apps, winning a hackathon, and helping build the internal practice. The award recognizes that work, but the post says it also belongs to people who read the posts, tried the code and wrote back with questions, because their feedback shaped much of what was written next. It links to the announcement and the awards video.
The post is “Agentic AI in Action- Part 32: One Name, Two Definitions, and a CoCo Skill That Picks the Right One,” originally published in Towards AI on Medium. The author’s stated focus is hands-on, implementation-focused perspectives on Generative & Agentic AI, LLMs, Snowflake and Cortex AI, translating advanced capabilities into practical, real-world analytics use cases. Readers are invited to follow the author on LinkedIn and Medium for more insights.
What teams can take from this pattern
When one business term has two meanings, an agent needs both definitions written down and a rule for choosing between them. In this example, two semantic views hold the governed numbers, and one skill file records both definitions, tells CoCo to ask which one the user means, and points it at the right view. Business Ontology is the managed version of the same idea, with owners, approvals and an audit trail. A skill like this is a simple way to start while that capability is in private preview.
Comments (0)