Case study

Store revenue, executive questions and forecasting on one set of definitions

Retail

A retailer with Microsoft Fabric in production wanted forecasting, Copilot questions and anomaly detection on its data. We made the definitions consistent first, then built the AI work on them and kept them that way.

Read the engagement brief(opens in a new tab)
Period
9 months
Company
500–1,000 people
A schematic: the separate reports this engagement started from, the one consolidated reporting layer that replaced them, and the technologies underneath.

AI on the definitions the reports use

The retailer’s Fabric platform was in daily use, with semantic models behind its reporting and store revenue already streaming in. Some definitions in those models had drifted from what the business had agreed, and an assistant or a forecast reading them would have inherited the drift. So the definitions were reviewed and realigned before any AI was prepared, and a fixed review rhythm keeps them aligned.

AI use cases, each on data already in the platform
3
stages: three in sequence, stewardship alongside
4
review cadences: monthly and quarterly
2
months of engagement
9

No outcome figures are published for this engagement: nothing exists that would let you check them, so what was built is described in full instead. Anonymised by sector and size, because the confidentiality that protects this client would protect you too. Sector and outcomes are real; identifying detail is not.

Before and after

Before and after the engagement

Drifting

  • Semantic models in daily useSome definitions drifted from what was agreed
  • KPIs not checked with their ownersNo schedule for catching drift
  • AI ideas without a governed baseEach would have needed its own data

Realigned

  • Definitions realignedReports and AI read the same figures
  • Monthly and quarterly reviewsDrift caught on a schedule
  • 3 AI use cases on the semantic layerForecasting, Copilot questions, anomaly detection
From a semantic layer drifting without review to one that is realigned, reviewed on a schedule and read by the AI work.

The engagement in full

Client
A retailer of 500 to 1,000 people, with Microsoft Fabric in production, semantic models over Direct Lake and a store-revenue stream.
Situation
Some definitions in those models had drifted from what the business had agreed, and nobody was checking. An assistant or a forecast reading them would have inherited the drift, and an assistant that answers inconsistently is harder to correct than a report that does.
Mandate
Review the semantic layer and make its definitions consistent, prepare AI use cases on it rather than beside it, and steward the platform’s architecture.
What we built
  • FoundationA review of the semantic models and the KPIs together, with findings ranked by priority and 3 AI use cases identified.
  • ActivationThe drifted definitions realigned, and the executive model prepared so Copilot answers in the same terms the executives use.
  • AccelerationDemand forecasting, Copilot questions and anomaly detection prepared as extensions of the semantic layer rather than beside it, each on data the business already trusts.
  • StewardshipA monthly semantic model review and a quarterly KPI review with the business, with runbooks and monitoring throughout.

Foundation: the review

The review looked at the models and the KPIs together, because a drifted definition shows up as a KPI its owner no longer recognises. Semantic models in production were read measure by measure, KPIs were checked with their business owners against how the business steers, and the findings were written up and ranked by priority: where definitions had drifted, and 3 AI use cases worth preparing.

Activation: the executive model for Copilot

With the definitions realigned, the executive model was prepared so Copilot answers in the same terms the executives use: each measure described in plain terms through annotations, the words executives use mapped to the model as dimension synonyms, and simple row-level security for that audience.

The 4 stages

AI was only prepared once the definitions it reads had been realigned in the Activation Sprint.

  1. FoundationReview and findings
  2. Activation SprintDefinitions realigned
  3. AccelerationAI use cases prepared
  4. StewardshipReviews on a schedule

Stewardship: the review rhythm

The rhythm set: a model review monthly, a KPI review quarterly, runbooks and monitoring throughout.

  1. Semantic model reviewMonthly
  2. KPI reviewQuarterly, with the business
  3. Runbooks and monitoringOngoing

The architecture

Sources: Store revenue stream · Source systems. Fabric platform: Lakehouse. Semantic layer: Other semantic models · Executive model. What reads it: Power BI reports · Demand forecasting · Copilot questions · Anomaly detection
The reports and the 3 AI use cases read the same reviewed semantic layer. Nothing was added beside it with data of its own.

Architecture decisions

DecisionRejected alternativeWhy
Realign definitions before any AIStarting the AI work straight awayAn assistant reading drifted definitions answers inconsistently and is harder to correct than a report.
AI as extensions of the semantic layerAI use cases as separate projectsForecasting and Copilot read governed data that already exists, with no data of their own.
Anomaly detection on the store-revenue streamA new feed built for AIThe stream was already in place, so the detection reads data the business already trusts.
Copilot on the executive model firstCopilot on every model at onceOne prepared model gives reliable answers before the scope widens.

What the client has today

  • A written review of its semantic models and KPIs, with findings ranked by priority.
  • Drifted definitions realigned, so reports and AI read the same figures.
  • An executive model prepared for Copilot questions.
  • Demand forecasting and anomaly detection on the store-revenue stream, prepared on governed data.
  • Monthly model reviews, quarterly KPI reviews, runbooks and monitoring.
Where it left them
Copilot, the forecast and the store-revenue checks read the same definitions as the reports, and the monthly review keeps them aligned.
Stack
  • Microsoft Fabric
  • Power BI semantic models
  • Direct Lake
  • Copilot in Power BI
  • Fabric AI Functions
  • Activator
All case studies

The other engagements

  • Retail

    A Fabric foundation for a retailer’s finance and sales

    A mid-size retailer of 50 to 250 people that needed reporting that finance and sales could both trust, built on Microsoft Fabric.

    • Microsoft Fabric
    • OneLake
    • Power BI
    • DAX
    • SQL
    • Dataflows Gen2

  • Industrial manufacturing

    Foundation then Activation then Acceleration

    Governed reporting across a production network

    An industrial manufacturer reporting across a production network of sites, each with its own view of what was running.

    • Microsoft Fabric
    • OneLake
    • OneLake Security
    • Direct Lake
    • Spark
    • MES
    • OPC UA
    • Power BI
    • DAX
    • Azure Key Vault
    • Purview
    • Azure DevOps

  • Hospitality and real estate

    Foundation then Activation

    A CFO cockpit over a group that closed its books separately

    A holding group across hospitality and real estate, with entities that each closed their own books and a CFO who had to answer for all of them.

    • Microsoft Fabric
    • OneLake
    • Direct Lake
    • Databricks Mirrored Catalog
    • Fabric Data Factory
    • Power BI
    • DAX
    • Calculation groups
    • Row-level security
    • Azure DevOps
    • Azure Key Vault
    • Purview

  • Multi-brand retail

    Foundation then Activation

    One governed model across a multi-brand retail group

    A retail group running several brands, each with its own systems, its own reporting habits and its own answer to what a margin is.

    • Microsoft Fabric
    • Direct Lake
    • Databricks
    • Eventstream
    • KQL Database
    • Power BI
    • DAX
    • SAP
    • Salesforce
    • Azure DevOps
All case studies

Free consultation

Know where your data stands in an hour.

A free 60-minute consultation. No pitch, no invoice. You’ll leave knowing whether your data is ready for AI, and whether we’re a good fit.

Book your free consultation(opens in a new tab)