A Design Pattern for a Governed Semantic Layer to Ensure Trustworthy Conversational AI

Organizations increasingly face a critical and costly problem: fragmented data across departmental silos leads to inconsistent business intelligence (BI) reports and, critically, undermines trust in new conversational AI tools. Existing AI prototypes often fail because they operate outside governed data ecosystems, generating “confidently wrong” answers that contradict official metrics. This fundamental lack of consistency and governance inhibits data-driven decision-making and slows the adoption of valuable Generative AI capabilities. This master's thesis addresses this gap by proposing, designing, and empirically validating a reusable architectural pattern for a Governed Semantic Layer that serves as the single source of truth for both traditional BI and modern Conversational AI. The design is founded on a Semantically-Bound Natural Language Query (NLQ) paradigm, which architecturally forces the AI agent to consume logic and governance policies exclusively from a central, code-based semantic model. This proactively embeds correctness, rather than relying on reactive post-query checks. The proposed solution was implemented as a proof-of-concept (PoC) in a corporate environment using Google Cloud, BigQuery, and Looker/LookML. The core contribution of this work is a validated, documented blueprint that transforms Conversational AI from a risky, isolated prototype into a reliable, enterprise-ready tool. The final design and its accompanying Change Management Protocol provide organizations with a concrete strategy for establishing trust by design in their integrated analytics platforms.