Almost every enterprise we talk to has an AI ambition. Far fewer have the data foundation to deliver it.
The pattern is familiar. A generative AI pilot impresses in the demo, then stalls on the way to production. The data it needs is scattered, late, inconsistent or impossible to trace.
Modern data architecture is the usual answer, and it means different things depending on who you ask. A data engineer thinks pipelines and integration. Ask a governance lead and you'll hear about lineage. Most business users just want faster access to numbers they trust.
Search the term and almost everything you'll read is written by a platform vendor making the case for its own product. That's understandable. It still leaves a gap: a straight answer on what makes an architecture modern and how to build one. We build on both Microsoft Fabric and Databricks, so that's the view this guide takes.
| Expectation | The question it answers |
|---|---|
| Speed | How quickly can we onboard a new data source and deliver quality, usable data to the business? |
| Trust | Can a user, or an AI agent, understand where data came from, how it was transformed and who may see it, whether they reach it through a report or a conversation? |
| Reuse | Can every team use the same governed data and business definitions, without duplicating or contradicting logic? |
| Flexibility | Can we add new technologies, data sources and use cases without redesigning large parts of the architecture? |
| Economics | Can we scale storage, processing and consumption while keeping cost, operational overhead and complexity under control? |
| Generation | Typical era | Strengths | Where it falls short for AI |
|---|---|---|---|
| Enterprise data warehouse | 1990s to 2000s | Consistent, trusted reporting on structured data | Batch only, structured only, costly to scale and change |
| Data lake | 2010s | Cheap storage at scale for any data type | Weak governance and quality; hard for business users to use |
| Cloud data platform (separate warehouse and lake) | Mid-2010s onward | Elastic scale and managed services | Duplicate data and governance; lineage breaks between systems |
| Unified lakehouse platform | 2020s | One governed copy for BI and AI; open formats; real-time and unstructured support | Still needs deliberate design for governance, semantics and operations |
Many of the clients who come to us have already done significant modernization work, often a cloud migration. They are frequently surprised that it did not make them ready for AI. Moving to the cloud does not, on its own, give you:
Each of these has to be designed. Usually nobody notices the gaps until after an AI pilot. The demo was built on a clean extract someone pulled by hand; production needs live data from many systems that often do not agree on what a customer is.
An agent answering a customer question needs the customer record, their orders, their support tickets and maybe product data too. We regularly find the same customer stored under different IDs across ERP, CRM, finance and support systems. Resolving that is real work, and it's rarely in the AI budget.
An agent checking stock needs the number as it is now, not as it was at 2 a.m. Most legacy pipelines refresh overnight. That's fine for a morning report and a problem when a customer is asking an agent whether something's in stock. Not every real-time need means streaming everything into the analytics platform, though. Sometimes the right answer is for the agent to call the operational system directly. The architecture has to make that call on purpose.
Most governance setups can tell you who has access to a table. Far fewer can tell you which tables fed a model's answer last Tuesday, or whether an agent acting for a user respected that user's permissions. Once risk or compliance gets involved, those are the first questions they ask.
| Characteristic | What it means in practice | Expectations it serves |
|---|---|---|
| Unified, open storage | One governed copy of data in an open table format such as Delta Lake or Apache Iceberg, readable by many engines | Reuse, flexibility, economics |
| Separation of storage and compute | Storage and processing scale, and are paid for, independently | Economics, speed |
| Structured and unstructured data together | Tables, documents, images and logs managed and governed in the same foundation | Flexibility, trust |
| Batch and real time | Scheduled processing for history and model training; streaming or direct access for anything live | Speed |
| Governance by design | Access control, lineage, classification and quality built into the platform rather than documented afterwards | Trust |
| A shared semantic layer | Business metrics and definitions defined once and reused by reports, applications and AI agents | Reuse, trust |
| Metadata-driven and discoverable | A catalog that makes data findable and understandable across the business | Reuse, speed |
| AI-ready serving | Vector search, feature management and governed access paths for models and agents | Flexibility, trust |
| Engineered for operations | Version control, CI/CD, monitoring and data observability, so the platform is run as a product | Trust, economics |
| Pattern | What it is | Best fit |
|---|---|---|
| Lakehouse | One storage layer for analytics and AI, combining a lake's scale with a warehouse's transactions and governance | Teams that want a single foundation for both BI and AI |
| Data mesh | Data ownership handed to business domains, each running its data as a product under shared governance | Large enterprises where a central data team is the bottleneck and domain teams are mature |
| Data fabric | A metadata and integration layer over existing systems that makes their data usable without consolidating it | Organizations with heavy legacy they cannot replace but still need to use |
| Approach | How it works | Best when |
|---|---|---|
| ETL | Transforms data before loading it | The logic is heavy, or the destination has limited compute |
| ELT | Loads raw data first and transforms it where it lands | You are on elastic cloud compute, want to keep raw data and need transforms that are easy to audit |
| Approach | How it works | Best when |
|---|---|---|
| Batch | Processes data on a schedule | Large volumes where a few hours' delay costs nothing |
| Change data capture and mirroring | Replicates changes from operational databases continuously | Near-real-time copies of operational data without a custom streaming build |
| Streaming | Processes events the moment they arrive | Real-time inference, monitoring and operations |
| Design-stage choice | What it locks in |
|---|---|
| Storage format | Whether lineage tracks itself or someone documents it by hand |
| Access model | Whether an AI agent gets scoped, auditable permissions or blunt all-or-nothing access |
| Metadata approach | Whether data is findable across the business or trapped department by department |
| Dimension | Microsoft Fabric | Databricks |
|---|---|---|
| Delivery model | Software as a service, capacity-based pricing | Platform on your cloud with serverless options, consumption-based pricing |
| Storage | OneLake in Delta Parquet; shortcuts to external storage | Your cloud object storage in Delta Lake; Iceberg support through Unity Catalog |
| Ingestion and transformation | Data Factory pipelines, Dataflows Gen2, Spark notebooks, mirroring | Lakeflow Connect, Lakeflow Declarative Pipelines, Spark, Lakeflow Jobs |
| Real time | Real-Time Intelligence (Eventstream, Eventhouse) | Structured Streaming, Lakeflow pipelines |
| Governance | OneLake security, OneLake catalog, Microsoft Purview | Unity Catalog for access, lineage and discovery across data and AI |
| Semantic layer | Power BI semantic models with Direct Lake | Unity Catalog metric views; Power BI is also common |
| ML and AI | Fabric Data Science with MLflow, Fabric data agents, Copilot in Fabric, Azure AI Foundry | MLflow, Mosaic AI model serving, vector search and agent tooling, Genie |
| Consumption | Power BI built in; Microsoft 365 and Copilot integration | Databricks SQL, AI/BI dashboards, Genie; open connectors including Power BI |
| Best fit | Microsoft-centric, BI-led organizations wanting a managed platform | Engineering-heavy, large-scale ML, multi-cloud, openness a priority |
What data you have, where it sits, how it moves and who governs it.
Decide what the architecture has to do first: reporting, analytics, AI or operations.
Both sets of requirements in scope from day one.
Get something live, learn from it, then build the next layer on ground that's already proven.
If you're not sure which generation your architecture is in, find out before you plan a build. Our AI Data Readiness Assessment looks at your data architecture, governance and quality against the AI use cases you actually want to run, and gives you a prioritized list of what to fix first.
If you already know what you need, talk to a rocker. We design and build modern data architecture on Microsoft Fabric and Databricks, in stages that each pay back.