Is your Semantic Model truly AI Ready for Data Managers

Key Takeaways

• Data Model Preparation: Explicitly configure tabular models for AI by hiding sensitive PII, establishing clear synonyms, and providing specific instructions for date logic and status fields.
• Defining Business Questions: Business questions, like policy counts by status, should be defined before building the dimensional model to ensure data quality and utility.
• The “What” vs. The “Why”: A semantic model explains what has occurred in the data but lacks context for why, necessitating the integration of knowledge management.
• Schema Simplification and Verification: To avoid AI confusion and prevent hallucinations, developers should bind verified data visualisations to specific trigger phrases for accurate responses to key business questions.
• Automated Dashboard Generation: Once simplified and governed, the semantic layer enables AI tools like Microsoft Copilot to autonomously create executive dashboards and reports from a simple text prompt.
• Strict Measure Definitions: To prevent calculation errors, organisations must explicitly define custom mathematical measures (e.g., average deductibles or loss ratios) rather than allowing the AI to autonomously aggregate or sum integer fields.
• Context-Assisted Graphs (CAG): A CAG enhances traditional RAG by linking qualitative context, like decision actors and evidence, directly to transactional events, creating a traceable “why narrative” for business decisions.
• Anomaly Detection and Architectural Decoupling: Semantic frameworks improve anomaly detection through a combination of AI techniques for clustering irregular data, while organisations should decouple semantic rules from execution tools to prevent vendor lock-in and ensure architectural flexibility.
• Iterative Implementation: Organisations should begin with specific domain-level use cases, like an automated FNOL triage process, to showcase immediate business value instead of trying to create a comprehensive enterprise semantic layer right away.

Webinar Details

Title: Is your Semantic Model truly AI Ready for Data Managers
Date: 2026-03-05
Presenter: Howard Diesel
Meetup Group: African Data Management Community
Write-up Author: Howard Diesel

Why Prepare Data Models for AI?

Preparing tabular data models, such as those in Power BI, for integration with artificial intelligence requires explicit configuration. Organisations must define which schema elements AI should access, deliberately excluding Personally Identifiable Information (PII) and sensitive terms.

A foundational requirement is establishing unambiguous synonyms; for example, mapping “policy ID” to “policy number” ensures accurate natural language processing by tools such as Microsoft Copilot. Furthermore, administrators must provide explicit instructions to the AI regarding date logic, status fields, and table-specific guidelines.

Additionally, a critical, often-overlooked practice in dimensional modelling is defining the specific business questions the model must answer before construction. Formulating queries—such as policy counts by status or jurisdiction—establishes a framework that dictates subsequent data quality requirements and guarantees the model’s operational utility.

Figure 1 Designing a Semantic Layer for AI Model Readiness

Figure 2 From BI to AI: Building an AI-Ready Semantic Layer

Figure 3 Premium Analysis

Figure 4 AI Data Schema (Exclude)

Figure 5 Synonyms

What is the Importance of Semantic Models, Knowledge Management, and the “What” vs. “Why”?

A fundamental distinction exists between semantic models and knowledge management systems. While a semantic model functions as an ontology detailing “what” has occurred within a dataset, it lacks the contextual capacity to explain “why.” Bridging this gap requires integrating knowledge management to provide comprehensive analytical insights. Furthermore, implementing a robust semantic layer is imperative for data governance.

This layer intercepts AI queries, restricting large language models (LLMs) from executing raw SQL commands directly against foundational databases, thereby mitigating data security risks. Governance is applied in practice by binding specific data visualisations to predefined, verified business questions. Consequently, when an AI tool encounters a standard query, it retrieves a validated visualisation rather than generating an unverified response, ensuring the accuracy and consistency of enterprise reporting.

Figure 6 The “What” and “Why”

Figure 7 Generic Large Language Models: Fail without Enterprise Context

Figure 8 Prep Data for AI: Verified Answers (Preview)

Why Simplify Schemas and Validate Answers?

Constructing an effective semantic layer necessitates the meticulous simplification of underlying data schemas. Complex structures, such as intricate snowflake architectures and opaque naming conventions, impede AI comprehension. It is necessary to eliminate technical keys and system identifiers that do not serve business intelligence purposes to streamline the model.

To further prevent AI hallucinations, developers must establish explicit synonyms and trigger phrases. These trigger phrases map standardised business queries to pre-verified answers and visualisations.

Rather than immediately attempting to construct a highly complex, enterprise-wide semantic model, industry practice suggests initiating development with localised, domain-specific models. By systematically applying domain-level definitions and verified visuals, organisations can sequentially integrate these models into a broader enterprise metadata repository, thereby maintaining accuracy and semantic integrity.

Figure 9 Bridging the Gap Requires a Highly Optimised Semantic Layer

Figure 10 The InsuranceCo Semantic Model Data Blueprint

Figure 11 Verified Answers – Trigger Phrase

Figure 12 Step 2: Bind Specific Trigger Phrases to Verified Answers

Figure 13 DEMO: Generating Complete Report Pages Instantly

How to use Copilot to Generate Reports and Dashboards?

When schemas are adequately simplified and governed, generative AI tools such as Microsoft Copilot can autonomously generate complex reporting dashboards. By inputting a prompt that includes predefined business questions and requested data fields, the AI can construct comprehensive report pages, calculating metrics such as rolling twelve-month trends and loss ratios.

To ensure the integrity of these automated outputs, strict parameter governance is required. System administrators can restrict AI access to specific factual records or sensitive fields directly within the semantic layer, providing stronger security than standard visual-layer hiding.

Additionally, administrators can program up to 10,000 distinct instructions to guide the AI’s behaviour. These directives manage measure prioritisation, resolve terminological ambiguities, and actively restrict the use of certain graphical representations, such as pie charts, ensuring adherence to corporate reporting standards.

Figure 14 DEMO: Interrogating the Data with Governed Context

Figure 15 Prompting CoPilot: “How many active policies do we have”

Figure 16 Verified Answers – Visuals

Figure 17 Using a Prompt to create a new Report with CoPilot in Power BI

Figure 18 All tables: Properties: Data Models: Measures

Figure 19 Prep Data for AI – Simplify the Data Schema

How to implement Measures, Diagnostics, and NotebookLM Integrations?

A critical protocol in preparing data for AI involves the explicit definition of mathematical measures. Standard data manipulation techniques, such as allowing an AI to aggregate integer fields autonomously, often result in calculation errors. Consequently, complex operations—including average deductibles or year-over-year loss ratios—must be strictly defined as distinct mathematical measures.

Once the semantic framework is finalised, the analytical outputs can be integrated with advanced AI processing tools such as NotebookLM. By extracting executive summaries from Business Intelligence platforms, organisations can utilise NotebookLM to automatically generate strategic artefacts, including infographics and formal presentations.

Furthermore, this integration facilitates internal training initiatives; the system can generate interactive flashcards based on organisational data dictionaries to support corporate data literacy programs. Finally, this structured data processing enhances the identification of statistical anomalies, allowing analysts to efficiently isolate significant data fluctuations and emerging risk profiles.

Figure 20 Using an Executive Summary Output from Power BI

Figure 21 Using the Executive Summary Generated to Create a Visual with NotebookLM

Figure 22 The InsuranceCo Semantic Layer: Data Model Blueprint

Figure 23 Chasing Anomalies and Identifying Model Errors

What are the challenges of implementing an Enterprise Semantic Layers?

The administrative burden of constructing semantic models can be mitigated through automation technologies. The introduction of Model Context Protocol (MCP) servers enables direct connections between Large Language Models (LLMs) and Business Intelligence environments.

This architecture enables users to manipulate reports, hide measures, and format data dynamically via natural language prompts. Despite these technological advancements, the development of a unified, enterprise-wide semantic layer remains a formidable challenge. Many large-scale implementations fail to demonstrate immediate return on investment, leading to organisational friction.

The consensus among data architecture experts is that success depends on rigorous, upfront definitional work. By systematically mapping subject-predicate relationships and clearly defining business requirements at the inception of a project, the corresponding data model naturally aligns with organisational objectives, creating a structurally sound foundation for subsequent enterprise scaling.

Figure 24 Finding Anomalies

What are Context Graphs, and How do they Help Navigate the “Why”?

To effectively answer why a specific business event occurred, organisations must move beyond traditional semantic models and implement a “context graph”. Utilising a “context mesh” architecture, the system generates qualitative bundles that are directly associated with specific data transactions. This framework meticulously records the core components of organisational decision-making: the responsible actors, evaluated evidence, contextual signals, and the explicit corporate policies that guided the final determination.

By binding this directed graph database to existing reporting infrastructures, end-users can trace data points to their underlying operational rationale. This mechanism provides a comprehensive “why narrative” for both automated and manual decisions. Consequently, when AI agents flag or escalate an event—such as an insurance claim denial—auditors are provided with a transparent, highly traceable record of the situational context, ensuring algorithmic and operational accountability.

Figure 25 The Insurance Co Context Mesh: Bridging the “What” and the “Why”

Figure 26 The Context Mesh Operationalises Institutional Memory and Logic

Figure 27 Examining the Context Mesh

Figure 28 The “Why” Summary

How to Identify Anomalies, and Why is it Important to Decouple Data Architecture?

Advanced semantic frameworks significantly enhance the detection of systemic anomalies. By employing an ensemble of distinct artificial intelligence methodologies, systems can effectively cluster irregular data points and generate comprehensive risk scores, thereby minimising both false positives and false negatives in automated triage processes.

Model integrity can be rigorously validated by injecting synthetic anomalies into training datasets; failure to detect these insertions indicates a need for immediate recalibration. Strategically, organisations are advised to decouple the storage of semantic rules and contextual data from specific execution applications.

Given the rapid evolution of Business Intelligence and AI software, tightly coupling logic to a singular proprietary vendor introduces substantial long-term risk. Although the industry currently lacks a definitive open standard for semantic interoperability, storing context rules independently enables organisations to fluidly integrate with various analytical tools, thereby maintaining architectural flexibility.

Figure 29 Table Tools: Data: “Bundle”

How to Implement Semantic Architecture, and What are RAG and CAG?

Implementation of semantic architectures should proceed iteratively, focusing initially on highly defined use cases, such as automated triage for First Notice of Loss (FNOL) procedures, rather than pursuing enterprise-wide integration immediately. A crucial conceptual advancement in this domain is the transition from Retrieval-Augmented Generation (RAG) to Context-Assisted Graphs (CAG).

While traditional RAG methodologies are limited to retrieving documents based on general queries, CAG directly associates qualitative metadata—including the specific actors involved and the temporal application of legal policies—to concrete transactional events documented within a fact table. By directly linking deterministic inputs and contextual signals to historical business decisions, the CAG methodology provides Large Language Models with a precise, auditable rationale. This advancement ensures that enterprise AI systems deliver transparent, accurate, and contextually grounded insights.

Figure 30 Mapping the Mesh into a Canonical, Relational Data Model

Figure 31 Context Bundles in Action: Triage and Escalation