Copilot and agents over the semantic model, a data science workload with notebooks and MLflow, real-time triggers. The four layers of AI in Fabric.
Microsoft Fabric supports AI and machine learning workloads in four layers that share one copy of the data in OneLake: Copilot, available from an F2 capacity, which writes DAX, builds report pages and answers plain-English questions over a Power BI semantic model; Fabric data agents and Fabric IQ, which let an agent reason across several semantic models through an ontology, in Teams or Copilot Studio, and take an action; the data science workload, where notebooks in Python or Spark train custom models with MLflow experiment tracking, AutoML and SynapseML, and score them back into the lakehouse; and Real-Time Intelligence with Activator, which raises a decision when a model's output or a KPI crosses a threshold. For a mid-market business the practical order is the reverse of the glamour: a governed semantic model with descriptions first, Copilot and agents second, triggers third, and a custom model only when a question needs one, such as demand forecasting, churn or anomaly detection. All four run on the same capacity, so there is no separate AI platform to buy. This post sets out each layer, what it needs underneath, and which to use first.
Layer one: Copilot over the semantic model
Copilot in Power BI Desktop writes DAX measures and report pages from a prompt; Copilot in the Power BI service answers questions and summarises reports; Copilot in notebooks and dataflows writes code and transformations. It needs a workspace on an F2 or larger Fabric capacity or Premium Per User, and it needs a semantic model with the business's names, descriptions and synonyms, because it answers from the model's metadata. A model without descriptions produces confident wrong answers. Is Power BI Copilot worth turning on, and what it needs first is the practical test.
Layer two: data agents and Fabric IQ
A Fabric data agent is configured over one or more semantic models, lakehouses or warehouses, given instructions and example questions, and then answers in natural language, in Fabric, in Teams or through Copilot Studio where it can be combined with other agents and take actions. Fabric IQ adds an ontology above the models, customers have orders, orders have lines, so an agent can reason across finance and sales rather than one table. This is where AI stops being a chat window and starts being a step in a decision. Fabric IQ and Copilot on your semantic model and an AI agent needs decision rights, not just intelligence cover the preparation and the governance.
Layer three: the data science workload
Notebooks in Python, PySpark or R over the lakehouse, with MLflow built in for experiment tracking and a model registry, AutoML for the common cases, SynapseML for Spark-scale training, and the ability to call Azure OpenAI and Azure AI services from a notebook. Models are scored in a notebook or pipeline and the predictions written back to a Delta table in OneLake, where Power BI reads them like any other column. The usual mid-market uses are demand forecasting, customer churn, late-payment risk and anomaly detection on operational data. Data science without a data scientist is the honest view of what a small team can run.
Layer four: real-time intelligence and triggers
Eventstreams land streaming data, KQL databases hold it for time-series queries, and Activator watches a value, a KPI, a model score, a threshold, and sends a Teams message, an email, or a Power Automate flow when it is crossed. It is what turns a forecast or an anomaly score into an action rather than a chart. How can decision intelligence be built on Microsoft Fabric and Power BI? sets out the loop.
What all four need underneath
A governed lakehouse with the data landed once and reconciled, a gold star schema, a semantic model with definitions and descriptions, and a named owner. AI amplifies whatever it is given, and most mid-market AI projects fail on undefined measures and ungoverned data rather than on models. How do you build a governed data foundation before AI? describes the foundation and the AI data readiness checker gives a first read in ten minutes.
Which to use first, and what it costs
Start with layer one on a well-described model, because it is included in the capacity you already have from F2 at £138 a month reserved and needs no new skills. Add a data agent over the finance and sales models when people ask cross-domain questions. Add Activator triggers for the three decisions with a clear threshold. Build a custom model only when a question needs a prediction rather than a report, and run it in the data science workload on the same capacity rather than buying a separate platform. Our AI and analytics page describes how we sequence it and what an AI analytics project costs puts figures on each step.
Related reading
Fabric IQ and Copilot on your semantic model: a conversational analytics guide
How can decision intelligence be built on Microsoft Fabric and Power BI?
What is Microsoft Fabric? What it replaces and how it sits with your Azure estate
Data science without a data scientist
Simon Devine
Managing Director
Part of the Hopton Analytics team, delivering governed analytics programmes for UK mid-market organisations.



