Every organisation claims to be data-driven. Very few have the architecture of a modern data platform to back it up. The gap is rarely a shortage of data, it is a shortage of coherent infrastructure. Siloed warehouses, disconnected ingestion pipelines, governance tacked on as an afterthought, and AI capabilities bolted onto legacy stacks rather than built in from the foundation. The result is an enterprise that cannot move as fast as its data demands.
This article sets out a reference architecture for a modern data platform, using Microsoft Azure as the implementation example. It is intended as a practical blueprint: one that leadership can use to assess where their current state falls short, and one that technical teams can use as a starting point for platform design.
Why Existing Data Architectures Are Failing
The traditional architecture, a relational data warehouse fed by overnight batch ETL jobs, with a reporting layer sitting on top, was built for a world of structured, slow-moving data and periodic decision-making. That world no longer exists.
Organisations today generate data from operational systems, SaaS applications, IoT devices, event streams, and third-party APIs, often simultaneously and at high velocity. The result is that conventional stacks are increasingly unable to support the AI and real-time analytics workloads that leadership now expects. As one analysis puts it, enterprises that still rely on fragmented warehouse architectures struggle to support AI and real-time analytics, because modern data strategy requires a unified platform that converges storage, compute, governance, and AI under a single operational model.
The answer is not a bigger warehouse. It is a fundamentally different architecture.
The Core Principles of a Modern Data Platform
Before examining the Azure blueprint, it is worth establishing what a modern data platform is actually designed to achieve. Four principles underpin the architecture described in this article.
Unified storage with layered transformation. All data, structured, semi-structured, and unstructured, should flow into a single logical lake, then be progressively refined through defined layers before reaching consumers. This eliminates the replication and divergence that plague multi-system environments.
Separation of storage and compute. Modern platforms decouple where data lives from the resources used to process it. This allows compute to scale independently based on workload demand, preventing the cost and performance problems of tightly coupled legacy stacks.
Governance as infrastructure, not process. Data lineage, classification, access control, and quality monitoring must be embedded in the platform itself, not managed through manual policy documents and periodic audits.
AI-readiness by design. The platform should not require retrofitting for machine learning and generative AI workloads. Model training, inference pipelines, and retrieval-augmented generation (RAG) should be first-class citizens of the architecture from the outset.
The Azure Reference Architecture: Layer by Layer
The following describes a production-grade reference architecture built on Azure services, organised around five functional layers: ingestion, storage, processing, serving, and governance.
Layer 1, Data Ingestion
The ingestion layer is responsible for bringing data from its various sources into the platform, whether in batch or in real time.
For batch ingestion, Azure Data Factory provides a managed orchestration service capable of connecting to over 100 native data sources, including on-premises databases, cloud SaaS platforms, and file systems. Pipelines can be scheduled, event-triggered, or run on-demand, and support both full and incremental load patterns.
For real-time and streaming workloads, Azure Event Hubs functions as the platform’s primary event ingestion service. It is designed to ingest millions of events per second from any source, with native integration to Azure Stream Analytics for in-flight transformation, and to Azure Data Lake Storage for automatic capture and long-term retention. Organisations with existing Apache Kafka workloads benefit from Event Hubs’ protocol-level Kafka compatibility, which allows migration without changing producer or consumer applications.
For more complex transformation logic applied at ingestion time, schema enforcement, deduplication, enrichment, Azure Stream Analytics provides a SQL-based real-time processing engine that sits between Event Hubs and downstream storage, with native integration across the Azure analytics stack.
Layer 2, Unified Storage: OneLake and the Lakehouse Model
The storage layer is where modern architecture diverges most sharply from its predecessors. Rather than maintaining separate systems for raw files, structured warehouse tables, and analytical indexes, the modern approach consolidates everything into a single logical store.
In the Azure ecosystem, this is realised through Microsoft Fabric’s OneLake, a tenant-wide, unified data lake built on Azure Data Lake Storage Gen2. All Fabric workloads operate over OneLake, meaning that data written once is accessible to data engineering, analytics, data science, and business intelligence tooling without duplication or movement. The platform simplifies what was previously a complex web of resource groups, RBAC configurations, and regional redundancy decisions into a single SaaS experience.
Within OneLake, data is organised using the medallion architecture, a three-layer model that has become the de facto standard for lakehouse data organisation:
- Bronze layer holds raw data exactly as ingested, no transformation, no schema enforcement. This provides a complete audit trail and allows reprocessing from source if downstream logic changes.
- Silver layer contains cleaned, deduplicated, and schema-validated data, typically stored in Delta Lake or Parquet format. This is where data quality rules are applied and where most analytical queries are executed.
- Gold layer holds curated, business-ready datasets: aggregated views, domain-specific models, and feature stores for machine learning. Data in the Gold layer is designed for direct consumption by reporting tools, AI models, and downstream applications.
Fabric standardises on Delta Lake format across all layers, which means all Fabric compute engines can access and manipulate the same dataset without duplicating data. Delta Lake also provides ACID transaction guarantees, enabling reliable concurrent reads and writes on large datasets, something traditional data lakes could not support without custom solutions.
Layer 3, Processing
The processing layer transforms data between medallion layers and supports workloads ranging from large-scale batch transformation to interactive notebook-based exploration.
Azure Databricks is the primary processing engine for organisations running complex transformation pipelines, machine learning workloads, or large-scale data engineering at enterprise volumes. Built on Apache Spark, Databricks provides a unified analytics platform that enables organisations to modernise legacy architectures by combining ETL, data warehousing, and AI into a single, future-proof stack. Its Unity Catalog governance layer, native Delta Lake support, and deep Azure Active Directory integration make it well-suited to regulated enterprise environments.
For teams working within the Microsoft Fabric ecosystem, Fabric Spark notebooks and pipelines provide equivalent transformation capabilities with tighter integration to OneLake and Power BI, and a lower operational overhead for teams already invested in the Microsoft stack.
The architectural decision between Databricks and native Fabric compute typically comes down to the complexity of existing Spark workloads, the degree of multi-cloud or vendor-neutral positioning required, and the organisation’s preferred governance model.
Layer 4, Serving
The serving layer makes processed data available to its various consumers: business intelligence tools, operational applications, data science platforms, and AI systems.
Microsoft Fabric Lakehouse and Warehouse serve as the primary analytical endpoints. As described in Fabric’s documentation, the two components operate on the same underlying data in OneLake, data engineering and data science teams work in the Lakehouse, while analysts query via the Warehouse’s SQL endpoint, with full consistency and traceability between them. This interoperability eliminates the historical trade-off between flexibility and performance.
Power BI provides the primary visualisation and self-service reporting layer, with native Fabric integration enabling direct semantic model connections that update in near-real time as Gold layer data changes.
For AI and generative AI workloads, Azure AI Foundry, which reached general availability in 2025, provides a unified studio for building, evaluating, and deploying AI applications using both Microsoft-hosted and open-source models. The integration between AI Foundry and OneLake means that retrieval-augmented generation (RAG) applications can query enterprise data directly, without requiring a separate vector database or data pipeline.
Azure Synapse Analytics remains relevant for organisations with established Synapse investments or those requiring tight integration with on-premises SQL Server environments. For greenfield deployments, Microsoft Fabric’s unified model is generally the preferred direction.
Layer 5, Governance and Security
No modern data platform is complete without embedded governance. In the Azure architecture, this responsibility falls primarily to Microsoft Purview.
Purview’s Unified Catalog, which reached general availability in September 2024, provides AI-powered data discovery, lineage tracking, access control, and data quality management across the entire data estate, including Azure, Microsoft 365, on-premises systems, and external clouds such as AWS and Google Cloud Platform. The platform classifies sensitive data automatically using machine learning, applies information protection labels, and enforces data loss prevention policies across workloads.
For organisations with regulatory obligations, GDPR, HIPAA, ISO 27001, the EU AI Act, Purview’s built-in compliance frameworks significantly reduce the manual effort associated with audit preparation and policy enforcement. According to analysis from Refoundry, Gartner projects that by 2026, 20% of organisations will have formal data governance programmes that are business-centric and ROI-driven, up from less than 10% in 2021, a recognition that governance is no longer an IT concern but a board-level risk and value driver.
Hybrid and Multi-Cloud Considerations
Few enterprises operate in a single cloud. For organisations running workloads across Azure, AWS, or on-premises infrastructure, Azure Arc extends Azure management, governance, and services to external environments. This allows Azure Policy, Microsoft Defender for Cloud, and Purview scanning to apply consistently across the entire hybrid estate, a meaningful reduction in governance complexity for organisations managing distributed data environments.
What a Mature Platform Enables
A reference architecture of this kind is not the end goal, it is the foundation for the capabilities that create competitive advantage. Organisations operating on a mature modern data platform are able to:
- Build and deploy machine learning models on curated Gold layer data without bespoke data extraction pipelines
- Serve real-time decisioning in applications, fraud detection, personalisation, predictive maintenance, using streaming data processed through Event Hubs and Stream Analytics
- Enable business teams to access and query trusted data through governed self-service interfaces, without dependence on data engineering capacity for every analytical request
- Respond to AI Act and GDPR audit requirements in hours rather than weeks, because data lineage and classification are maintained automatically within the platform
- Onboard new data sources, domains, and workloads without architectural rework, because the foundation is designed to extend
The reference architecture described here is not a theoretical ideal; it reflects the patterns that enterprise data teams are deploying at scale today, and the direction that Microsoft’s platform investments are clearly heading. The convergence of storage, processing, governance, and AI under a unified lakehouse model represents a structural shift from the fragmented stacks that many organisations still operate.
The organisations that deliberately adopt this architecture, rather than accruing technical debt through incremental patching of legacy systems, will be best positioned to operationalise AI at enterprise scale. The infrastructure question is not separate from the AI strategy question. It is the foundation of it.
