AI-Assisted Fabric Data Modelling Case Study

The modelling problem

Start with the model the reports need

Moving a SQL Server warehouse to Fabric raised a modelling question: which parts of the existing design should carry forward?

Copying the tables would also copy old assumptions. The work centred on reviewed facts, dimensions and business definitions, with source evidence behind the redesign.

Microsoft Fabric

Source evidence

Let the data inform the proposal

SQL Server metadata described the structures. Profiling examined keys, missing values and record behaviour, giving the modelling system evidence to work from.

That evidence informed AI-assisted proposals for facts, dimensions and Silver and Gold structures. A proposed model still needed an accountable review before adoption.

Modelling decisions

Keep reporting meaning in human hands

Human review determined the grain, relationships and measures before adoption. Those decisions tied the model to reporting needs, rather than simply reproducing the source tables.

The Fabric foundation combined the warehouse redesign with accountable definitions. Silver and Gold structures supported the model; reporting definitions and ontology made its business meaning explicit.

Process → Architecture → Delivery

Our methodology

01DiscoverAgree the outcome, the decisions and the requirements.
02Current stateMap how the work and data flow today, with evidence.
03Root causesFind what drives the delay, rework or disagreeing numbers.
04ApproachAgree the target design, scope and measures of success.
05BuildBuild the process change, model or report the design calls for.
06OptimiseTest at real volumes; tune speed, cost and usability.
07ProductioniseRelease with managed deployment, monitoring and support.
08GovernSet owners, access and controls so it keeps working.

Swimlane process map: To-be: AI-assisted proposal with accountable review. The steps are listed below the diagram.

Illustrative reference processTo-be: AI-assisted proposal with accountable review AI drafts; it never approves. Evidence travels with each proposal and reviewers record their rationale before a design is built.

Steps in this diagram

Lanes: Notebooks (automated), AI model, Data architect, Business reviewer.

Reporting needs agreed
Scan metadata + profile data
Propose facts, dims, measures
Check against evidence
Decide grain, relationships, measures
Accepted?
Revise proposal
Build Silver / Gold
Definitions published

Full size

Architecture diagram: Evidence-led modelling pipeline: SQL Server to Fabric. The components are listed below the diagram.

Reference architectureEvidence-led modelling pipeline: SQL Server to Fabric Metadata and profiling are the evidence. A Foundry-hosted model drafts proposals, reviewers decide in a pull request with the evidence attached, and only accepted designs are built as Silver and Gold.

Components in this design

SQL Server (legacy warehouse)
Metadata scan (schemas, keys, FKs)
Data profiling (uniqueness, nulls, orphans)
AI proposal (facts, dims, Silver/Gold)
Evidence store (profiles + proposals)
Pull request review (grain, relationships, measures)
Review board (sample report on this page)
Mirror / copy (Bronze landing)
Silver + Gold (reviewed model)
Semantic model (definitions + ontology)

Full size

First page of the Data Model Review sample report

Delivery: interactive Power BI sample reportData Model Review Track every proposed fact and dimension from evidence to accepted design, so nothing reaches Gold without an accountable decision. Delivered with named owners, role-based access, release pipelines and a controlled way to change governed measures.

Full screen

Who it is for: Data architect, data owners, reporting leads.
Decision it supports: Approve, change or reject each proposal; prioritise profiling gaps.
Report pages: Review board

Reference design: how this pattern is typically built on Microsoft services. Component choices for a specific engagement depend on the client's environment and existing licences.

Microsoft icons are used under Microsoft's terms. Icon notices

Synthetic sample data, not client data.

Results in view

Legacy licence cost avoided

Client-reported result

Reported saving

AUD 40,000licence cost avoided

Saving amount · AUD · axis starts at zero

Saving reportedAUD 40,000

Approximately AUD 40,000 reported legacy licence cost avoided. Period not stated by the client.

Next step

Rebuilding a data model for Fabric?

Tell us what is happening. We agree priorities and scope before proposing any work.

Talk through your project The service: Microsoft Fabric Architecture & Data Engineering