Everyone Wants Copilot—but Is Your Data Ready: Microsoft Fabric Architecture with Walter Calcagno [MVP-MCT]

Everyone wants Copilot. Everyone wants AI agents. Everyone wants employees to ask natural-language questions and immediately receive trustworthy answers from enterprise data. But what happens when the underlying data is fragmented, business definitions conflict, governance is weak, and semantic models were never designed for AI? In this episode of M365 FM, Mirko Peters sits down with Walter E. Calcagno, Microsoft MVP in Data Platform, Microsoft Certified Trainer, data architect, author, educator, and co-founder of Data Consultants, to explore why successful enterprise AI begins long before the first prompt is written.Walter brings a perspective shaped by more than a decade working with data. His journey started in accounting, where managing increasingly complex financial information in spreadsheets pushed him toward SQL, databases, Power BI, big data, Spark, and eventually artificial intelligence. He also explains why creating original Spanish-language technical content has become an important part of his work and discusses the growing Spanish-speaking Microsoft data community.THE BUSINESS QUESTION COMES BEFORE THE TECHNOLOGYOne of the central themes of the conversation is deceptively simple: organizations do not really need more dashboards—they need better answers to business questions. Walter describes a progression that begins with understanding what happened, continues with understanding why it happened, moves toward predicting what will happen next, and ultimately asks what actions can influence the desired outcome. Business intelligence, statistical analysis, machine learning, and recommendation systems can all contribute to those questions, but the technology remains secondary to the business problem.This changes the way organizations should think about analytics. Instead of beginning with Power BI, Fabric, Python, or another technology and asking what can be built with it, organizations should begin with the decisions they need to make. Technologies will evolve and individual products may disappear, but the fundamental business questions remain.WHY MICROSOFT FABRIC CHANGES THE DATA PLATFORMWalter explains Microsoft Fabric as Microsoft's attempt to bring technologies that organizations previously assembled individually into a unified Software-as-a-Service data platform. Storage, movement, analytics, big-data processing, Power BI, Spark, machine learning, and other capabilities can operate within a more integrated environment rather than forcing organizations to assemble every component separately.A major part of that architecture is OneLake. Walter discusses how OneLake provides a common data foundation and how technologies such as Delta Lake help organizations work with large volumes of data while retaining structures and capabilities traditionally associated with databases. The objective is not simply to centralize technology. It is to make enterprise data easier to organize, process, analyze, and eventually expose to AI systems.The discussion then moves to one of the biggest misconceptions surrounding enterprise AI: if the data already exists somewhere, why not simply connect an LLM or Copilot directly to it?Walter argues that this skips essential architectural layers. AI needs context about what enterprise data actually means. Within Fabric, semantic models provide structured representations of business data. But large organizations frequently have many semantic models across departments and domains. Trying to solve that problem by creating one enormous semantic model is not necessarily the answer.This is where Walter highlights ontology as another important layer. Rather than forcing everything into a single semantic model, an ontology can describe relationships across models and provide a structure through which AI systems can navigate enterprise information. In Walter's view, semantic models combined with ontology models represent an increasingly important foundation for connecting enterprise data with LLMs while preserving meaning, relationships, and access controls.WHAT DOES “AI-READY DATA” ACTUALLY MEAN?Having data does not mean having AI-ready data. Walter uses the familiar medallion architecture to explain why data must progress through different levels of preparation. The first layer can preserve historical source data without attempting to solve every quality problem. A subsequent layer cleans and standardizes information, handles duplicates, establishes consistency, and prepares the data for broader analytical use. Additional layers can then prepare specific subsets of data for specific purposes such as business intelligence, machine learning, deep learning, or AI.Walter also challenges the idea that medallion architecture must always mean exactly three layers. The number of layers should follow the requirements of the architecture rather than the terminology used to describe it. What matters is that organizations understand what each stage is designed to accomplish.AI readiness also has an economic dimension. Sending unnecessary data into an LLM consumes tokens and increases cost. Preparing the correct data and metadata therefore becomes both an architectural and financial requirement.METADATA IS THE MAP YOUR AI NEEDSOne of Walter's strongest recommendations is to stop thinking about AI as a system that should continuously scan everything an organization owns. Instead, organizations should create a map that helps the model find the information it actually needs.That map is metadata.Walter compares this to traveling between cities. You do not drive through every small street looking for your destination. You use highways, then progressively smaller roads until you arrive at the exact location. Metadata provides a similar navigation structure for an LLM: first identifying where relevant information exists and then allowing the system to access the specific data required to answer the question.Without strong metadata, AI systems must work harder, consume more resources, and have fewer reliable signals about where trustworthy information resides.AI SHOULD NOT INVENT YOUR BUSINESS DEFINITIONSRevenue, customer, margin, active employee, sales, churn, and countless other terms can mean different things across departments. An LLM should not be expected to decide which definition is correct.Walter argues that organizations must explicitly define these concepts and provide the model with the appropriate context and instructions. The AI needs to learn how a KPI or business concept is defined within that specific organization. Allowing the model to make those decisions independently introduces unnecessary risk and can contribute to hallucinations or inconsistent answers.Reliable enterprise AI therefore depends not just on clean rows and columns, but on shared definitions and clearly communicated business meaning.THE FIVE DIMENSIONS OF AI DATA READINESSWalter introduces the five-dimensional assessment his team uses when organizations approach them asking to implement AI against enterprise data. Rather than immediately deploying an LLM, they first evaluate whether the organization is actually prepared for AI adoption.The assessment examines areas including governance, technology, internal organizational or political decision-making, human resources, and training/readiness. The objective is to create a snapshot of where the organization stands and identify what must be addressed before investing heavily in an AI solution.If governance is missing, governance may need to come first. If employees lack the skills required to use the technology effectively, training becomes the priority. Walter's point is practical: buying access to powerful AI does not create business value when the organizational foundations required to use it are absent.THE SEVEN-LAYER DATA ARCHITECTUREWalter also walks through the seven-layer architecture model described in his work. Three foundational layers span the overall data environment: governance, monitoring, and security. These capabilities should not be treated as isolated additions at the end of a project; they underpin the entire architecture.Above them are four functional areas. First are the data sources themselves, which can range from relational databases and NoSQL systems to APIs and other external sources. Next comes the engineering and movement of that data, including pipelines and streaming scenarios. The third functional area determines where the data will live, such as a database, warehouse, data mart, or lakehouse. Finally comes the analytics layer, where the information can be used by business intelligence, data science, machine learning, deep learning, or AI solutions.The architecture provides a framework for moving from operational data to actual decisions while maintaining control, security, and observability throughout the process.GOVERNANCE IS NOT OPTIONAL FOR ENTERPRISE AIGovernance has a direct impact on AI readiness because AI makes accessing information dramatically easier. That convenience becomes dangerous when the underlying access model is poorly governed.Walter uses sensitive employee information as an example. Working for the same organization does not mean every employee should have access to colleagues' salaries, addresses, phone numbers, or other private information. The same principle applies across enterprise datasets: people should have access to the information required for their responsibilities and decisions—not automatically to everything the organization possesses.An impressive AI interface built on top of uncontrolled data access is not a mature enterprise AI implementation. Governance establishes ownership, responsibilities, permissions, and boundaries that allow AI to operate safely against business information.Become a supporter of this podcast: https://www.spreaker.com/podcast/m365-fm-modern-work-security-and-productivity-with-microsoft-365--6704921/support.