What happens when AI agents start making sense of financial and business data — not just displaying it?In this episode of M365.FM, Mirko Peters talks with Rishi Sapra Microsoft MVP about the architecture required to move beyond traditional dashboards and toward AI-powered, context-aware analytics with Microsoft Fabric, Power BI, Copilot Studio, semantic models, ontologies, and data agents.Organizations already have enormous amounts of information spread across ERP systems, Excel workbooks, Power BI reports, Microsoft Fabric, SharePoint, financial systems, budgets, forecasts, and operational applications.Adding AI on top of that data does not automatically mean the AI understands the business.What does “revenue” actually mean? Which definition of margin should an AI agent use? Which KPIs represent the official version of the truth? And how does an agent understand relationships between customers, products, regions, cost centers, contracts, and business processes?The answer increasingly lies in the context layer between raw data and AI. FROM SELF-SERVICE BI TO SELF-SERVICE AI Rishi explains his journey from financial modeling and Excel through Power Query and the early days of Power BI to today's Microsoft Fabric and AI ecosystem.Power BI helped bring business intelligence out of centralized IT departments and into the hands of business users.But the platform has also become significantly more sophisticated.Semantic models, DAX, lakehouses, OneLake, Direct Lake, Copilot Studio, data agents, ontologies, MCP-based tools, governance, and AI now create an architecture that can quickly become difficult for individual business users to understand.AI agents could change that relationship.Instead of requiring every business user to become a data engineer, BI developer, and AI engineer, agents could increasingly build and operate parts of the technical architecture while humans provide the business context. WHY BUSINESS CONTEXT MATTERS FOR AI One of the central questions of the episode is surprisingly simple:How well documented are your business processes and decisions?Much of an organization's real knowledge does not live in a database. It exists inside Excel formulas, Power BI measures, SharePoint files, business processes, documentation — and people's heads.For AI agents to produce meaningful business insights, organizations need to capture more than data.They need to capture context.Who is asking the question?What decisions does that person need to make?Which KPIs matter?Which business rules apply?What does a specific metric mean in that particular context?This leads to the concept of persona-driven insights: designing analytics around the decisions and questions of specific business users rather than simply exposing more data. DATA MODELS VS SEMANTIC MODELS VS ONTOLOGIES The conversation explores three increasingly important concepts in modern Microsoft analytics architecture.A data model structures the underlying data and relationships.A Power BI semantic model adds business logic, measures, calculations, relationships, and security — creating a governed analytical layer and a reliable source for KPIs.But AI often needs more.An ontology can describe business entities and relationships in a way that allows AI to reason about concepts such as customers, products, stores, employees, regions, contracts, revenue, and business processes.Semantic models help answer:“What is the number?”Ontologies and additional context can help AI investigate:“Why did the number change?”Together, these layers provide much stronger grounding for AI agents. MICROSOFT FABRIC AS THE DATA FOUNDATION FOR AI Microsoft Fabric plays a central role in this architecture.OneLake, Lakehouses, Delta tables, semantic models, Direct Lake, Fabric Data Agents, and integration with Copilot Studio can create a unified foundation for structured and unstructured organizational data.The episode also explains why Direct Lake matters.Instead of repeatedly importing and refreshing data into traditional Power BI semantic models, Direct Lake allows Power BI to work directly with data stored in Delta format while maintaining analytical performance.This can significantly simplify the path from enterprise data to analytics and AI. THE FIVE LAYERS OF AN ORGANIZATIONAL BRAIN Rishi describes an “organizational brain” built around five interconnected layers:Data — trusted enterprise information and source systems.Logic — DAX, SQL, Python, calculations, KPIs, and business rules.Tools — semantic models, APIs, MCP servers, applications, and other capabilities agents can use.Skills — instructions and business processes describing how agents should use those tools and interpret information.Governance — permissions, policies, security, controls, and rules governing what agents are allowed to do.The goal is not simply to give an LLM access to more data.The goal is to give AI a governed environment in which it understands which data, logic, tools, and processes should be used for a particular business question. AI AGENTS NEED DETERMINISTIC DATA Generative AI is powerful because it can reason flexibly.Financial reporting cannot rely entirely on flexibility.Revenue, margins, forecasts, costs, and other business metrics often require deterministic calculations and a governed source of truth.The episode explores why the future of enterprise AI may therefore depend on combining two worlds:Deterministic computing for trusted calculations and business logic.Generative AI for reasoning, interpretation, exploration, and natural-language interaction.Semantic models and Microsoft Fabric can provide the deterministic foundation while AI agents provide the flexible reasoning layer. FROM DASHBOARDS TO PERSONALIZED INTELLIGENCE Traditional dashboards tell users what happened.A dashboard might show that revenue decreased by seven percent. The user then needs to drill through dimensions, filters, reports, and datasets to understand why.AI agents can potentially perform much of this exploration automatically.But good storytelling still requires context.The most important number is not always the largest number. A business metric may need to be interpreted relative to revenue, budget, previous periods, organizational structure, or other factors.This is where persona-driven analytics becomes particularly important.The CFO, sales leader, and operational manager may all ask about the same KPI while requiring very different explanations and actions. COPILOT STUDIO AND THE NEXT GENERATION OF AGENTS The conversation also explores the evolution of Microsoft Copilot Studio from traditional topic- and knowledge-based chatbot experiences toward more capable agentic systems.Modern agents can potentially combine tools, skills, enterprise data, workflows, and reasoning.That additional capability also introduces additional risk.The more autonomy an agent receives, the more important grounding, permissions, governance, evaluation, and clearly defined instructions become.AI agents should not invent financial numbers or arbitrarily choose data sources.They need trusted semantic models, governed data, explicit skills, and clear guardrails. MAKING GOVERNANCE GREAT AGAIN Governance becomes even more important in an agentic organization.Instead of treating governance as a document that employees are expected to read, organizations can increasingly encode governance directly into the environments, policies, skills, and instructions used by AI agents.The discussion explores the idea of treating agents more like digital employees.They need defined responsibilities, approved tools, access permissions, business rules, performance expectations, and boundaries.Governance therefore becomes less about documentation and more about automated enforcement. THE OPERATING MODEL FOR FABRIC AND AI Finally, the episode examines how organizations can manage this architecture at scale.A Center of Excellence can provide the enablement layer while a hub-and-spoke model combines centralized governance with decentralized innovation.Certified enterprise data, semantic models, logic, governance policies, and reusable skills can live in governed hubs.Business teams can experiment within their own domains and promote successful assets into the governed enterprise layer.The result is neither completely centralized nor completely decentralized.It is a federated model designed to support both control and self-service. IN THIS EPISODE We discuss Microsoft Fabric, Power BI semantic models, data modeling, ontologies, OneLake, Lakehouses, Direct Lake, Delta tables, Fabric Data Agents, Microsoft Copilot Studio, AI agents, persona-driven insights, storytelling with data, deterministic computing, enterprise AI governance, Center of Excellence models, hub-and-spoke architectures, self-service BI, self-service AI, and the idea of building an organizational brain for AI.The bigger question is no longer simply:“How do we build better dashboards?”It is:“How do we give AI enough trusted business context to understand our organization — without allowing it to invent its own version of the truth?” Become a supporter of this podcast: https://www.spreaker.com/podcast/m365-fm-a-microsoft-mvp-podcast-by-mirko-peters--6704921/support.
