AI in Power BI: Copilot, AI Visuals and Predictive Analytics
Four stages: describe, diagnose, predict, prescribe.
AI in Power BI is a set of capabilities you can add to the reporting you already have: built-in AI visuals, Copilot, and predictive models from Azure Machine Learning or Fabric. You do not need a new platform. You need the right order of work and clean data underneath.
This guide covers all three, what each needs, and a six-step roadmap from descriptive reporting to predictive analytics. It replaces our earlier separate posts on AI analytics in Power BI, the transition from data analytics to AI, and predictive insights, and it is the Power BI companion to our broader guide to AI in data analytics.
Why Most Power BI Estates Stall at Descriptive Reporting
Descriptive reports answer what happened. They rarely say why it happened or what is likely next, so stakeholders stop asking forward-looking questions and start exporting data to other tools. The semantic models, pipelines and report layouts you built still have value. AI features extend them, which is why you can add them step by step instead of rebuilding.
AI in Power BI: Four Stages from Descriptive to Predictive
Think of analytics maturity as four stages. Each answers a different question and depends on the stage before it. Skipping ahead, for example training a model on ungoverned data, is the most common reason AI pilots stall.
| Stage | Question | Typical capability | Needs |
|---|---|---|---|
| Descriptive | What happened? | Dashboards, KPIs, standard visuals | Governed semantic model |
| Diagnostic | Why did it happen? | Key influencers, decomposition tree, anomaly detection | Clean dimensions and a date table |
| Predictive | What will happen? | Forecast, Azure ML models, AutoML in Dataflow Gen1 | History, labelled outcomes, validation |
| Prescriptive | What should we do? | Scored actions in a decision workflow, Copilot for follow-up | Owners, thresholds and feedback loop |
Built-In AI Visuals in Power BI
Microsoft groups four visuals as AI-powered: key influencers, decomposition tree, smart narrative and anomaly detection. They work on your existing semantic model, so they are the lowest-effort place to start.
Key influencers and decomposition tree
Key influencers shows the major contributors to a chosen result, such as which customer attributes go with churn. The decomposition tree breaks one measure down across dimensions in any order, and can use AI to suggest the next dimension to drill into. It needs a measure to analyse and one or more dimensions to explain by, and it suits root-cause exploration.
Smart narrative and anomaly detection
Smart narrative writes a plain-language summary of trends and takeaways next to a visual. Anomaly detection flags unexpected spikes and dips, with an expected range and optional explanations. It works only on line charts with a time axis, needs at least four data points, and is not supported with legends or multiple values.
Forecasting in the Analytics pane
If you have time data, the Analytics pane can add a forecast to a line chart, with a length and confidence interval you control. It is a simple projection, so use it for direction rather than planning.
Copilot in Power BI: What It Does and What It Needs
Copilot lets report authors draft pages, measures and summaries, and lets users ask questions in natural language. It has firm requirements: a paid Fabric capacity (F2 or higher) or Power BI Premium capacity (P1 or higher), the Copilot admin setting on, and a capacity in a supported region. Trial capacities and free SKUs are not supported.
Prepare the semantic model first
Microsoft is direct that an unprepared model gives generic or inaccurate answers. Use clear names, field descriptions, a star schema, and fewer unused objects and ambiguous relationships, and use the Prep data for AI features to add verified answers and AI instructions. These steps are the same ones that improve any self-service experience, so the work pays back twice.
Set realistic expectations
Copilot output can be inaccurate, and it is nondeterministic: the same prompt can return different answers. Roll it out in phases, explain when users should read the report instead of prompting, and teach them how to verify results. If you used the legacy Q&A feature, note that Microsoft has announced its deprecation for December 2026 and points users to Copilot.
Azure Machine Learning and Predictive Models in Power BI
When built-in visuals are not enough, you can bring in a model. There are four common routes, and they differ in effort and governance.
Azure Machine Learning in Power Query
You can apply a trained Azure Machine Learning model to your data by invoking it as a function in Power Query. The model creator must provide the schema files, written in Python, that describe the model inputs and outputs. This keeps predictions in a governed pipeline and produces prediction columns that flow into the semantic model.
AutoML in dataflows
Dataflow Gen1 creators can train binary prediction, general classification and regression models during data preparation, evaluate them, and apply them after refresh. Note that Dataflow Gen1 is now a legacy state with no new feature investment, and Microsoft recommends Dataflow Gen2 for Premium customers with Fabric access, so check the current position before you build new work on it.
Foundry Tools and Python or R
Azure AI services, now called Foundry Tools, can add text and image insights through Power Query functions, and they run on Fabric or Premium capacity. Python and R scripts suit smaller, batch use cases. Set a memory limit for AI workloads so they do not consume unexpected capacity units.
Fabric data science
For larger work, train models in Fabric notebooks, track them, and write predictions back to OneLake for Power BI to report on. Microsoft's reference architecture for churn prediction follows this pattern: train on historical data in a Spark pool, deploy with Azure Machine Learning, store the scores, and show risk in Power BI. Our MLflow in Microsoft Fabric guide covers tracking and deployment.
- AI in Power BI is added in stages, on the platform you already run.
- Start with built-in AI visuals, then add Copilot or a predictive model.
- Copilot needs Fabric F2+ or Premium P1+, and a prepared semantic model.
- Predictions should land as governed data with visible accuracy and freshness.
- Measure workflow outcomes against a baseline set before you deploy.
Choosing the Right AI Capability
| Capability | Answers | Needs | Watch for |
|---|---|---|---|
| Key influencers | Which factors drive a result | Built in; works on the semantic model | Shows association, not proof of cause |
| Decomposition tree | Where a number comes from, step by step | Built in; needs a measure and dimensions | Best for root-cause exploration |
| Smart narrative | A written summary of a visual or page | Built in | Check the wording against the data |
| Anomaly detection | Unexpected spikes and dips | Line chart with a time axis and at least four data points | Not available with legends or multiple values |
| Forecast | Where a line is heading | Analytics pane, line charts only | Simple statistical projection |
| Copilot | Natural language questions, report and DAX drafts | Fabric F2+ or Premium P1+, admin setting, prepared model | Outputs vary; verify them |
| Azure ML or Fabric model | Custom prediction such as churn | Trained and deployed model, capacity, governance | Needs data science skills and monitoring |
Match the capability to the decision, not the other way round. Many decisions need only a key influencers visual and a clear owner.
Data Readiness: The Precondition for AI in Power BI
Every capability above depends on the semantic model. A certified date table, consistent keys, governed measures and enough history are the baseline. As a rule of thumb, seasonal metrics need at least two full annual cycles before a forecast is credible, though the right amount depends on the metric. Our AI readiness guide and the data maturity scorecard help you check where you stand.
Governance: Making Predictions Trustworthy
People act on a prediction only if they trust it. For each predictive output, show what it is based on, how accurate it has been, and when it was last updated. A simple model performance card on the report page does this. Add access controls for sensitive scoring data, and review models on a schedule. The Power BI governance platform supports this at tenant level.
Roadmap: Six Steps from Descriptive to Predictive in Power BI
Step 1: Audit the semantic model
Check naming, relationships, descriptions and the date table. Fix issues here, because additional instructions cannot correct wrong relationships or aggregations.
Step 2: Activate built-in AI visuals
Use key influencers, decomposition tree, smart narrative and anomaly detection on one or two decisions, and note what users ask next.
Step 3: Pick one predictive use case
Choose a decision with history, a named owner and a measurable result, such as churn, demand or a revenue forecast.
Step 4: Build and validate the model
Train in Azure Machine Learning or Fabric, validate on held-out data, and agree the accuracy that is good enough to act on.
Step 5: Connect predictions to the semantic model
Land scores as governed tables or columns, refresh them on a schedule, and show their accuracy and last update.
Step 6: Design the decision workflow
Define who acts on a score, what they do and how the outcome is recorded, so the model can improve.
Where AI in Power BI Pays Off: Four Business Functions
These are typical uses, not guaranteed results. Finance teams can refresh forecasts as each period opens and flag unusual variances. Sales teams can score accounts for churn or expansion risk. Operations teams can surface exceptions in production or fulfilment metrics with anomaly detection. Marketing teams can analyse which segments respond to campaigns. In each case the value comes from the action taken, so name an owner first.
Measuring Return on AI in Power BI
Measure workflow outcomes rather than feature use. Set baselines before deployment for analyst hours spent on routine reporting, time from question to answer, and the share of users who self-serve. After rollout, compare the same measures over the same periods, and add one business outcome tied to the use case, such as retained revenue from a churn programme.
AI in Power BI FAQ
What does Copilot in Power BI need?
It needs a paid Fabric capacity (F2 or higher) or Power BI Premium capacity (P1 or higher). Trial capacities and free SKUs are not supported. The admin setting for Copilot must be on, and the capacity must be in a supported region.
Can Power BI do predictive analytics?
Yes, in three ways. The Analytics pane adds forecasts to line charts, Azure Machine Learning models and Dataflow Gen1 AutoML can add prediction columns during data preparation, and Fabric notebooks can train and score models that Power BI then reports on.
Which AI visuals are built into Power BI?
Key influencers, the decomposition tree, smart narrative and anomaly detection. They work on your existing semantic model. Anomaly detection only works on line charts with a time axis and needs at least four data points.
What happens to Power BI Q&A?
Microsoft has announced that the legacy Q&A natural language tool will be deprecated in December 2026 and recommends Power BI Copilot instead. If you rely on Q&A, plan the move now.
Can I trust Copilot answers?
Treat them as drafts. Microsoft notes that Copilot output can be inaccurate and is nondeterministic, so the same prompt can return different answers. Prepare the semantic model, set expectations with users, and verify important results.
Where should I start?
Audit your semantic model, switch on the built-in AI visuals for one or two decisions, and only then add Copilot or a predictive model. Each step builds on clean, governed data.
For delivery support, see our AI and machine learning services and Power BI consulting, or talk to our team. For the wider picture, read AI use cases in analytics.