The promise of GenAI-powered analytics is compelling: business users ask questions in natural language and receive accurate insights instantly. However, a critical trust issue is emerging as organizations rush to implement “ask-your-data” experiences—how do you ensure these conversational BI systems deliver reliable, consistent results every time?
Without proper foundations, the same business question can yield different answers, destroying user confidence in AI-generated insights. Our recent webinar with partners at Distillery demonstrated how semantic layers solve this trust crisis by grounding GenAI experiences in consistent business logic.
The demo wasn’t just technically impressive; it represented a fundamental shift in how we can deliver on the promise of conversational business intelligence. Let me share the key insights and explain why this approach addresses the trust issue that has been plaguing “ask-your-data” implementations.
The Rise of Conversational BI and the Trust Problem
Organizations rush to implement natural language query (NLQ) capabilities, drawn by the compelling vision of business users simply asking questions and getting immediate answers. The appeal is obvious; instead of waiting for reports or mastering complex BI tools, users could ask, “What is the name and square footage for each warehouse in Fairview?” and receive instant, accurate responses.
But without proper data foundations, these conversational BI experiences consistently fail to deliver reliable results.
As I mentioned during the webinar, “Natural language queries and agents and chatbots really just make this lack of trust or lack of consistency of access to your data even worse.”
The core issue lies in how large language models interpret business questions. Without structured semantic guidance, the same question asked twice can yield different results. Promising GenAI analytics projects lose credibility with business users because of this inconsistency problem.
Why Shared Definitions Matter in Natural Language Experiences
Years of working with enterprise customers have proven that the foundation of trusted conversational BI isn’t the AI itself; it’s establishing consistent, business-friendly definitions of what your data means.
This is where semantic layers become absolutely critical. A semantic layer creates a “business representation of data”—it allows users to ask questions using terminology they already understand while ensuring that concepts like “revenue,” “customer,” or “warehouse capacity” have consistent definitions across every tool and AI agent.
Without this semantic foundation, LLMs struggle to understand business context, often generating queries against hundreds or thousands of physical database tables. The result is poor performance, inconsistent answers, and ultimately, failed implementations.
How AtScale’s Semantic Layer Improves NLQ Accuracy and Consistency
At AtScale, we’ve built our universal semantic layer around five core services that directly address the trust problem in conversational BI:
Our Metric Store creates logical views of physical data through semantic models. What I find powerful about our approach is that business analysts can build these models visually, while analytics engineers can code them using our open-source Semantic Modeling Language (SML). This flexibility has been crucial for adoption.
The Semantic Query Engine is where the magic happens for GenAI applications. It translates business-friendly natural language queries into optimized physical queries.
As I emphasized in the webinar: “When you pair a semantic layer with an LLM, you can be assured that every single question gets answered the same way every single time.”
Our Automatic Query Optimization ensures that conversational BI experiences respond in seconds, not minutes, by intelligently rewriting queries to access pre-built aggregates. This performance is critical—users abandon systems that can’t keep up with conversational speed.
Governance capabilities ensure that data access policies work consistently, whether users are accessing data through traditional BI tools or AI agents. This consistency has been a game-changer for enterprise deployments.
Finally, Universal Consumption means semantic models work with any tool—Excel, Power BI, Tableau, Python notebooks, and now LLMs through protocols like our MCP server.
What makes this approach so effective for LLMs is that we present each semantic model as a single virtual table. The LLM doesn’t need to navigate complex database schemas or figure out table relationships—we’ve done that work in the semantic layer.
Enabling Trusted GenAI Responses with Governed Semantics
Our integration with the Model Context Protocol (MCP) represents what I see as the future of LLM-data integration. MCP acts as “JDBC for LLMs,” providing a standardized way for AI agents to access external data and metadata.
The AtScale MCP server exposes our semantic models as tools that LLMs can use to answer business questions. What excites me about this approach is how it ensures:
- Consistency: Every query against the same semantic model returns identical results
- Accuracy: Business logic is embedded in the semantic layer, not dependent on LLM interpretation
- Performance: Queries are automatically optimized for interactive response times
- Governance: Data access policies apply uniformly across all AI interactions
Real-World Example: NLQ with AtScale in Action
The demonstration by Emmanuel Fernández and Francisco Maurici from Distillery perfectly illustrated the potential of this approach. Their “Distill Genie” assistant integrates our MCP server to enable natural language queries through Slack and Google Meet—precisely the kind of multi-modal experience I believe users want.
What impressed me most was the consistency. Users could ask the same business question through Slack chat, voice commands in video meetings, or direct application interfaces, and receive identical results every time. The system provided rich visualizations, LLM-generated insights, and exportable reports while maintaining conversation context for follow-up questions.
Francisco’s comment resonated with me: “I can promise you, for me it was extremely easy to integrate the AtScale MCP server with my code. And this was made in a few days without not much effort.”
This ease of integration is exactly what we aimed for when designing the MCP server.
Voice-Powered Data Access: MCP Enables Trusted “Talking to Your Data” Experiences
The Google Meet integration in the Distillery demo illustrates a fundamental shift in enterprise data access, the ability to ask complex business questions using voice commands while maintaining the same accuracy and governance as traditional BI tools.
The technical implementation demonstrates the power of MCP as a standardized protocol for LLM data integration. Distillery integrated wake word detection (Porcupine), voice-to-text conversion (Whisper), and text-to-speech responses, but the critical foundation remains the AtScale MCP server providing semantic layer access. This architecture ensures that voice-based queries receive the same trusted, consistent results as text-based interactions.
The business impact centers on meeting productivity and decision-making speed. When teams can get immediate, accurate answers to data questions during discussions, it eliminates the typical “let me get back to you on that” delays. Questions like “What are the total sales for the Western region this quarter?” get answered in real-time while maintaining full data governance and accuracy through the semantic layer.
What makes this approach particularly valuable is the consistent user experience across modalities. Whether users ask questions through Slack text, voice commands in meetings, or traditional BI interfaces, they receive identical results because all interactions flow through the same semantic models. This consistency builds user trust in the system’s reliability.
The ease of extending existing MCP integrations to new interfaces demonstrates the protocol’s flexibility. Distillery reused their entire backend architecture—the same supervisor agent, the same AtScale MCP server connection, and the same data formatting capabilities. This reusability proves that once you establish semantic layer foundations, you can expose enterprise data through virtually any user interface without rebuilding business logic.
From a governance perspective, voice-enabled data access maintains the same security and access controls as traditional methods. Users only access data they’re authorized to see, and all queries are logged and auditable. The semantic layer ensures that business rules and calculations remain consistent regardless of how users interact with the system.
This voice integration represents more than a novel interface—it demonstrates how proper semantic foundations enable organizations to innovate on user experience while maintaining enterprise-grade data governance and accuracy.
Key Takeaways for Implementing Reliable Ask-Your-Data Interfaces
Based on customer implementations and partnerships like the one with Distillery, here’s my advice for organizations wanting to implement trustworthy conversational BI:
Start with Semantic Modeling: Before implementing any natural language query capabilities, establish clear, consistent definitions of business metrics and dimensions. This foundation is critical for both accuracy and user adoption.
Choose Universal Solutions: Your semantic layer should be independent of specific data platforms or BI tools. Too many organizations lock themselves into vendor-specific approaches that limit future flexibility.
Implement Comprehensive Governance: Data access policies and security controls must work consistently across traditional BI tools and new AI agents. Governance can’t be an afterthought.
Optimize for Performance: Conversational BI experiences must respond quickly enough for interactive use. Users will abandon systems that make them wait.
Support Multiple Consumption Methods: Enable various interfaces while maintaining consistency. The future is multi-modal—chat, voice, traditional BI tools, and custom applications should all work seamlessly.
Provide Transparency: Allow users to understand how their questions were interpreted and what data sources were accessed. This transparency is crucial for building confidence in AI-generated insights.
The combination of AtScale’s semantic layer with GenAI capabilities represents the next evolution in enterprise analytics. By grounding LLMs in consistent, governed semantic models, we can finally deliver on the promise of reliable “ask-your-data” experiences that business users will trust and adopt.
SHARE
Case Study: Vodafone Portugal Modernizes Data Analytics