Governed Semantics for Trustworthy AI: Bridging the Context Gap

Most enterprises have AI projects running on their data, but few actually trust the results. That disconnect between AI promise and AI reality sits at the heart of Do You Trust AI on Your Data? Bridging the "Context Gap" with Governed Semantics, a session featuring Vijay Ganesan, senior director of product management at Snowflake.

Vijay makes a compelling case that governed semantics for trustworthy AI isn't optional. It is the missing layer between your data warehouse and the conversational analytics tools your teams want to use. Without it, large language models (LLMs) hallucinate, invent metrics, and produce numbers no one can rely on. With it, you get consistent, accurate answers across every analytical surface, whether someone writes SQL or asks a question in plain English.

Close the context gap with governed semantics

Trustworthy AI on enterprise data depends on closing the context gap with governed semantics, not just adding an LLM on top of a warehouse. Conversational analytics often fails in enterprises because LLMs lack the unique enterprise business context needed for 100 percent accuracy. Governed semantics provides the missing layer between LLMs and warehouses so answers match enterprise definitions and rules, directly improving trust and readiness for large-scale rollout.

Vijay frames the core problem bluntly: "Hallucinations, incorrect numbers. The LLM is making up metrics. It's inventing assumptions. It's making wrong predictions. And it results in low trust."

The accuracy bar for analytics is unforgiving. As Vijay puts it, "If I'm a knowledge worker at an enterprise and I ask revenue numbers, well, that better be right. That better be 100 percent right. There is no approximation."

What you can do now to close the gap

Start by treating governed semantics as a prerequisite layer for AI on analytics, not an optional enhancement. Here are three practical steps to get moving:

1. Audit your current AI-on-data projects for accuracy issues. Where are people seeing hallucinations or incorrect numbers? Those pain points signal where semantic definitions are missing.
2. Identify your canonical metric definitions. What is revenue? What is an active customer? Document these in a centralized place, not scattered across BI tools.
3. Establish a single source of truth for business rules. If your count of accounts excludes partner accounts, that rule needs to live somewhere AI can access it.

Surface hidden business context for AI accuracy

The majority of analytics-critical meaning is hidden business context—definitions, relationships, rules, organizational changes, and tribal knowledge—that must be made explicit for AI to answer correctly. Vijay uses an iceberg analogy: LLMs only see 10 percent of what they need. The other 90 percent sits below the surface, invisible to the model.

That hidden 90 percent includes several layers:

• Business definitions cover canonical metrics like revenue, profit margin, and active customer counts.
• Relationships describe join keys between entities and hierarchies (product category, subcategory, item) that do not exist in your relational database.
• Business rules capture exclusions and conditions—only count completed orders, exclude returns.
• Organizational context reflects time-based changes, like a pricing model shift on July 1st. Tribal knowledge holds the corner cases that only Sarah in billing, for example, understands.

How to operationalize the hidden 90 percent

Encoding this context in a semantic model reduces AI ambiguity and aligns outputs to canonical enterprise logic. Here is a blueprint for approaching it:

• Map your metric definitions first. Start with the five to 10 metrics your executives care about most. Write down the exact calculation, the tables involved, and any exclusion rules.
• Document relationships that live outside the database. Join keys, hierarchies, and entity relationships often exist only in people's heads or in scattered documentation. Capture them in your semantic layer.
• Track organizational context changes. When pricing models shift, when territories realign, when product hierarchies change—these need to be reflected in your semantic model so AI knows which rules apply when.
• Interview your tribal knowledge holders. That person who knows all the billing edge cases? Their knowledge needs to be encoded, not just relied upon.

End multiple versions of truth with centralized semantic models

Centralizing semantic models breaks the cycle of multiple versions of truth created by tool-specific, proprietary semantic layers. When product says average revenue per active user is $229, sales says $340, and finance says $189, who do you believe? This problem has existed for decades, but AI amplifies it.

But, according to Vijay, a semantic model can become the universal interface for analytics. It sits between the data platform and consuming tools—AI, BI, and applications—so every client gets the exact same answer.

The measurable benefit is consistency: the same answer on any analytical surface. That means no more KPI disputes, no more reconciliation overhead, and no more governance headaches from conflicting definitions.

Steps to centralize your semantic layer

Moving from tool-specific semantic models to a centralized approach takes planning. Here is a practical path forward:

1. Inventory your existing semantic models. Every BI tool you use has one. Count them, document what they contain, and identify overlaps and conflicts.
2. Choose a centralization strategy. You can build semantic models in your data platform (like Snowflake Semantic Views), use a dedicated semantic layer tool, or adopt open standards like Open Semantic Interchange (OSI) for vendor-neutral interoperability.
3. Start with high-impact metrics. Do not try to centralize everything at once. Pick the metrics that cause the most disputes across teams and centralize those first.
4. Plan for iteration. Vijay emphasizes that getting semantics right takes cycles of trying, tuning, and learning. Automation tools like Semantic View Autopilot can reduce maintenance burden by learning from schema, query history, and user interactions.

Key takeaways about building a governed semantics layer

• Treat governed semantics as a prerequisite for AI on analytics, not an afterthought.
• Encode the 90 percent of hidden context—definitions, relationships, rules, organizational changes, and tribal knowledge—in a semantic model.
• Centralize semantic models to eliminate multiple versions of truth and ensure consistent answers across all tools and interfaces.

Governed semantics for trustworthy AI isn't a nice-to-have. It's the foundation that makes conversational analytics actually work at enterprise scale. When you put a solid semantic foundation on top of your data model, you can trust AI on your data. Without it, you are stuck with hallucinations, inconsistent numbers, and low adoption.

Ready to dig deeper into how semantic models can transform your AI-on-data strategy? Watch the full session from Domo to hear Vijay walk through the details, including how Snowflake Semantic Views integrate with Domo and how Open Semantic Interchange is tackling vendor lock-in concerns.