Logo - Keyrus
  • Playbook
  • Services
    Advisory & License Sales
    Data Management
    Business Intelligence
    Cloud
    Training
  • Insights
  • Partners
  • Careers
  • About us
    What sets us apart
    Company purpose
    Innovation & Technologies
    Committed Keyrus
    Regulatory compliance
    Investors
    Management team
    Brands
    Locations
  • Contact UsJoin us
Expert opinion

6 min read

How to design AI-ready data models for Microsoft Fabric

Adam Walker, General Manager at Keyrus

Artificial Intelligence thrives on data, not just any data, but well-structured, context-rich, and trustworthy data. As organisations rush to embed AI into decision-making, analytics, and operations, one truth becomes clear: you cannot retrofit AI-readiness onto poor data models.

If you’re building in Microsoft Fabric, the opportunity, and more so the responsibility, is to design data models from the outset that are AI-ready by design, not AI-ready by accident. Let me explore what that actually means in practice.

1. Start with semantics, not just structure

A common mistake is to focus solely on the schema design including the tables, columns, and keys and not embedding semantic meaning into the model. AI systems, especially large language models and machine learning pipelines, perform better when data carries clear business context.

  • Adopt a common business vocabulary. Use a governed business glossary so that “customer,” “order,” or “margin” mean the same thing everywhere. In Microsoft Fabric, this aligns naturally with Lakehouse and OneLake data cataloguing, where you can tag datasets with business terms.

  • Model entities, not just tables. Treat your data objects as real-world entities with relationships, attributes, and lifecycle states. This makes feature engineering for AI far more straightforward later.

Many might say thatsemantics slow down delivery however I would argue that skipping semantics creates a technical debt trap. AI models trained on ambiguous data yield ambiguous results, the path to get there and the output may be fast but inherently flawed.

2. Prioritise data quality and provenance

AI amplifies the quality of your inputs be they good or bad. An AI-ready data model must enforce lineage, data quality, and trust signals.

  • Capture metadata exhaustively. Use Microsoft Purview or the built-in Fabric data catalog to register source systems, refresh cadence, owners, and data quality metrics.

  • Implement data contracts and validation rules. Treat upstream schema stability as non-negotiable. Breaks in contracts derail machine learning pipelines that depend on predictable structures.

  • Track lineage through pipelines. Microsoft Fabric Data Factory and Microsoft Fabric Dataflows Gen2 can be configured to write lineage metadata, which is critical for auditing model training sets.

Look at your data, as a practical test if you can’t answer, “Where did this field come from and when was it last verified?” your model isn’t AI-ready.

3. Embrace a lakehouse-native design

Microsoft Fabric uses a lakehouse architecture which combines data lake scalability with data warehouse structure. Designing for AI means embracing that duality:

  • Use the medallion architecture. Land raw data in bronze, clean it in silver, and curate business-ready data in gold. This clean separation lets AI workloads choose the right layer for their purpose.

  • Leverage Delta Lake tables. Delta’s ACID transactions and time-travel capabilities make your data trustworthy and reproducible, these are two essentials for AI model training and auditing.

  • Unify storage in OneLake. Storing all data in OneLake ensures discoverability and consistency across teams, which is vital when scaling AI across departments.

This approach prevents the common anti-pattern where AI projects spin up siloed copies of data outside governed environments.

4. Engineer for feature readiness

An AI-ready model isn’t just about storing data, it’s about readying data for machine learning features.

  • Design for temporal context. Include timestamps, versioning, and slowly changing dimensions. AI needs to know how facts change over time.

  • Normalise where needed, denormalise where helpful. Normalisation improves consistency, but wide denormalised tables often accelerate feature extraction.

  • Build reusable feature tables. Use Microsoft Fabric Real-Time Intelligence or the Microsoft Fabric Machine Learning experience to persist and share engineered features across models.

In short: don’t just model for today’s BI reports, model for tomorrow’s feature pipelines.

5. Govern with AI in mind

Data governance is often treated as a compliance checkbox. But for AI, governance is foundational. Models inherit the biases, gaps, and access patterns of their data.

  • Enforce access controls and security roles. AI workloads often cut across business domains; strong role-based security prevents accidental data leakage.

  • Curate training datasets with bias checks. Data scientists can’t detect bias if your model doesn’t carry demographic or domain context metadata.

  • Version datasets just like code. Use Git integration with Fabric notebooks or pipelines to ensure reproducible ML experiments.

A well-governed model ensures that AI outcomes are defensible, not just accurate.

6. Think evolution, not perfection

Finally, avoid the trap of trying to design the “perfect” AI-ready model from day one. AI needs agility as much as structure.

Design your model so that:

  • New data sources can be onboarded quickly

  • Schema evolution is non-disruptive

  • Lineage, quality rules, and semantics can grow incrementally

The key is to build governed adaptability: the ability to change rapidly without breaking trust.

Closing Thoughts

Designing AI-ready data models for Microsoft Fabric is less about technology tricks and more about mindset. It requires blending traditional data modeling discipline, semantics, normalisation, and quality all with a forward-looking posture: anticipating how AI will consume, interpret, and depend on your data.

Do it right, and you give your organisation a competitive edge: AI systems that are accurate because your data is trustworthy, fast because your pipelines are clean, and scalable because your architecture was designed for tomorrow, not yesterday.

Let us help you build AI systems that are reliable, explainable, and future-ready. Contact us at sales@keyrus.co.za to begin your AI journey with confidence.

Start your AI journey with a reputable partner
Related Articles
  • Event

    Calling all Tableau users: October 2, 2025

  • Success story

    Expert Data Architecture for SA Home Loans

  • Expert opinion

    Responsible AI begins with responsible data engineering

  • Expert opinion

    The power and shortfalls of Large Language Models (LLMs)

  • Event

    Keyrus as Premier Sponsor of Anaplan Connect Cape Town 2025 

Logo - Keyrus
Durban

10 Flanders Drive Mt Edgecombe 4302 Kwazulu Natal

Phone:+27 87 350 8860