Connecting AI to Power BI: 5 Approaches Beyond Microsoft Copilot

Approach 1: Azure OpenAI Service — LLM-Powered Annotations and Anomaly Explanation

Azure OpenAI provides you with GPT-4o, GPT-4.1, and other advanced models within your Azure tenant. It includes enterprise security controls, private endpoints, and data residency guarantees.

The main use case for Power BI integration is automated insight generation. This allows the LLM to:

• Explain why a KPI changed
• Identify what drove an anomaly
• Suggest insights based on data patterns

Architecture pattern: A Fabric notebook or Azure Function runs on a schedule. This schedule can be hourly, daily, or triggered by a data refresh.

It queries the Power BI semantic model through the XMLA endpoint or reads from the Lakehouse.

Next, it creates a prompt using the relevant data context. For example, it might say, “Revenue dropped 12% week-over-week. Here are the top 10 contributing factors by dimension...” This prompt is sent to Azure OpenAI. The LLM's response is then written to a table.

Finally, the Power BI report displays this narrative in a text card or custom visual. It updates with each data refresh.

When to use: Real-time or near-real-time annotations inside dashboards. Anomaly explanation. Automated data storytelling for operational dashboards where users need context, not just numbers.

Limitations: Azure OpenAI context windows (128K tokens for GPT-4o) may not be sufficient for very large datasets without summarization. Cost at high inference volume can be significant — EPC Group recommends caching responses and only re-generating when the underlying data changes meaningfully.

Approach 2: Anthropic Claude via API — Long-Context Narrative Generation

Azure OpenAI is strong in short-form, structured outputs. In contrast, Anthropic's Claude models excel in long-context analysis and narrative writing.

Claude can handle over 200K tokens in a single prompt. This means you can input an entire quarter's worth of KPI data from all business units, regions, and product lines. In return, you will receive a coherent executive summary that rivals the work of a senior analyst.

Architecture pattern: A Python script runs as an Azure Function, Fabric notebook, or scheduled job. It extracts data from the Power BI semantic model and formats it as structured CSV or JSON. The script sends this data to the Claude API with a detailed system prompt that defines the report format and tone. Finally, it writes the resulting narrative to Azure Blob Storage or a database table.

Power BI displays the narrative, or it is delivered as a PDF attachment via Power Automate.

Use cases EPC Group has deployed:

• Quarterly Business Reviews (QBRs) — Claude generates 10–15 page narrative reports summarizing financial performance, operational KPIs, and strategic recommendations from Power BI data. Delivered to the board as polished PDFs.
• Healthcare compliance narratives — for AI governance in healthcare, Claude summarizes patient outcome dashboards into narratives that compliance officers review, with citations back to the source data.
• Multi-property hotel performance summaries — weekly narratives for hotel GMs comparing their property to portfolio averages, generated from the same Power BI semantic model that feeds the interactive dashboards.

Governance consideration: Claude is accessed through an API. This means that data leaves your Azure tenant and goes to Anthropic's infrastructure. For HIPAA or FedRAMP workloads, EPC Group sanitizes PII before sending data to the Claude API. All identifiable data is kept within Azure.

For non-regulated workloads, Anthropic's data retention policies ensure adequate protection. These policies include:

• No training on API data.

Approach 3: Python Visuals with Embedded ML Models

Power BI supports Python and R visuals that execute scripts when the visual renders. This opens the door to embedding machine learning models directly in the report — anomaly detection, clustering, forecasting, or classification displayed as matplotlib, seaborn, or plotly charts.

Architecture pattern: A Python visual gets a filtered subset of data from the Power BI semantic model. This data is based on the active slicers and filters. The script loads a pre-trained model, which can be a pickled scikit-learn model, an ONNX model, or a simple statistical algorithm. It then runs inference on the data and displays the result as a chart. The visual updates in real-time as users interact with the filters.

Production examples:

• Anomaly detection visual — Isolation Forest model flags outlier transactions in financial dashboards. Red dots on a scatter plot indicate anomalies, with tooltips showing the anomaly score.
• Customer segmentation — K-means clustering visual that segments customers by purchase behavior, displayed as a 2D scatter plot with PCA-reduced dimensions. Clicking a cluster filters the rest of the dashboard to that segment.
• Time-series decomposition — statsmodels seasonal decomposition showing trend, seasonal, and residual components of a KPI, helping analysts understand whether a change is structural or cyclical.

Limitations: Python visuals in the Power BI service operate in a sandboxed environment. They have no network access, a 5-minute timeout, and limited library availability. These visuals re-execute with every interaction, which can slow down reports.

For production, EPC Group uses Python visuals for display. The actual ML inference runs in a Fabric notebook or an Azure ML pipeline. This setup writes results to a table, allowing the visual to read the pre-computed output.

Approach 4: Microsoft Fabric Data Science Notebooks

Microsoft Fabric unifies data engineering, data science, and BI in a single platform. Fabric Data Science notebooks (Spark-based, supporting PySpark, Python, R, and Scala) can train and deploy ML models that write predictions directly to the Lakehouse — and Power BI reads from the same Lakehouse via DirectLake mode.

Architecture pattern: A Fabric notebook reads training data from the Lakehouse. It trains a model using MLflow for experiment tracking. The model is then registered in the Fabric ML model registry. Finally, a scoring pipeline is scheduled to write predictions to a Lakehouse Delta table.

The Power BI semantic model includes this table via DirectLake. This setup ensures that predictions appear in dashboards with near-zero latency and no data duplication.

Use cases:

• Demand forecasting — LightGBM or Prophet model trained on historical sales data, external signals (weather, events, economic indicators), and calendar features. Predictions written to Lakehouse and displayed alongside actuals in the Power BI sales dashboard.
• Customer churn prediction — classification model scoring active customers daily. Churn probability appears as a column in the customer dimension, enabling Power BI slicers like “show me high-risk customers with > $100K annual spend.”
• Supply chain risk scoring — model that scores suppliers based on lead time variability, quality defect rates, financial health indicators, and geopolitical risk factors. Risk scores surface in Power BI procurement dashboards.

Why this is EPC Group's preferred approach: Fabric notebooks maintain all data within the Microsoft security boundary. They use the same capacity as Power BI, eliminating separate Azure ML billing. Additionally, they support MLflow for model versioning and governance. Fabric notebooks also write directly to Lakehouse tables that Power BI accesses via DirectLake, providing the tightest integration available today.

Approach 5: Custom AI Visuals (Power BI Custom Visuals SDK)

The Power BI Custom Visuals SDK allows developers to build TypeScript/D3.js visuals that can call external APIs — including AI endpoints. This enables embedding AI directly into the visual interaction layer, where users interact with AI outputs as native Power BI elements.

Architecture pattern: A custom visual built with the pbiviz SDK receives data from the semantic model. It calls an Azure Function, which proxies to Azure OpenAI, Claude, or a custom ML endpoint. The visual then renders the AI response alongside the data visualization.

The Azure Function manages:

• Authentication
• Rate limiting
• Response caching

Examples EPC Group has built:

• AI-annotated chart visual — a line chart that automatically displays LLM-generated annotations at significant inflection points. Hover over a data point and see “Revenue increased 23% due to Black Friday promotions and new product launch in the Southeast region.”
• Natural language query panel — a custom visual that provides a chat-like interface within the Power BI report, routing questions to Azure OpenAI with the current filter context as grounding data. More flexible than the built-in Copilot Q&A because it can target custom models and include business-specific instructions.
• Predictive tooltip visual — hover over a customer or product and see a predictive overlay: “This customer has a 73% probability of churning in the next 90 days based on declining order frequency and support ticket volume.”

Limitations: Custom visuals require TypeScript development expertise and must pass organizational governance review before deployment. They also need certification if distributed via AppSource. For most organizations, EPC Group recommends approaches 1–4 first and reserves custom visuals for high-value use cases that justify the development investment.

Building a Multi-Model AI Strategy for Power BI

The common mistake is treating AI integration as a single-model problem — deploying Copilot and calling it done. Enterprise organizations benefit from a multi-model approach where different AI capabilities serve different analytics needs:

The orchestration layer determines which model responds to each request. This can be straightforward, using purpose-built pipelines where each use case has its own pipeline. Alternatively, it can be complex, using a semantic router that classifies intent and directs requests to the right model.

EPC Group begins with purpose-built pipelines. We add orchestration only when the number of AI touchpoints makes it necessary.

Frequently Asked Questions