Microsoft Foundry — formerly Azure AI Studio — is the unified platform that now underpins every AI capability in the Microsoft ecosystem. For Power BI users, it unlocks access to over 1,900 AI models, including GPT-5, Claude, Grok, Mistral, DeepSeek, and Meta Llama, all accessible directly from your Fabric workspace without provisioning separate Azure resources.
If you think Power BI AI starts and ends with Copilot, you are missing the bigger picture. Here are five concrete ways Foundry AI transforms what your Power BI reports can do.
What Is Microsoft Foundry?
Microsoft Foundry is the rebranded and unified Azure AI platform. It combines what used to be Azure AI Studio, Azure AI Foundry, and parts of Azure Cognitive Services into a single experience for building, deploying, and managing AI applications.
For Power BI users, the key detail is this: Foundry provides a model marketplace with 1,900+ models, a tool catalog with 1,400+ integrations, and pre-built connectors that work directly inside Fabric. You do not need to be an AI engineer to use it. The integration is designed for data analysts and BI developers.
Way 1: Foundry Tools in Fabric (Pre-Built, No Azure Resource Needed)
The simplest entry point is Foundry Tools in Fabric. These are pre-built AI models available directly inside your Fabric workspace. You do not need to provision an Azure OpenAI resource, manage API keys, or configure endpoints.
What is available today:
- Azure OpenAI Service: GPT-5.1, GPT-5, GPT-5-mini, GPT-4.1, GPT-4.1-mini, and text-embedding-ada-002
- Text Analytics: Sentiment analysis, key phrase extraction, PII detection, named entity recognition, entity linking, language detection, summarisation
- Azure Translator: Text translation and transliteration between languages and scripts
All three are billed against your Fabric capacity. There is no separate Azure bill. Usage is tracked through the Copilot Capacity (FCC) or your workspace's assigned capacity.
How to access them: Foundry Tools are available through SynapseML, the REST API, the Python SDK, and the AI Functions API in Fabric notebooks. For most Power BI users, the AI Functions path is the easiest.
Way 2: AI Functions — Nine Single-Line Commands for Data Transformation
Fabric AI Functions are the most underappreciated feature in the entire Power BI ecosystem. They let you run LLM-powered operations on your data with a single line of code.
The nine available functions:
- ai.analyze_sentiment — Detect positive, negative, mixed, or neutral sentiment in text
- ai.classify — Categorise text by custom labels you define
- ai.embed — Generate vector embeddings for similarity search and retrieval
- ai.extract — Pull out entities (names, locations, dates) with structured JSON Schema support
- ai.fix_grammar — Correct spelling, grammar, and punctuation
- ai.generate_response — Custom text generation with your own instructions, supports JSON output
- ai.similarity — Compare semantic meaning between text pairs, scored -1 to 1
- ai.summarize — Generate summaries with customisable instructions
- ai.translate — Translate text to another language
Here is what real usage looks like in a Fabric notebook:
from synapse.ml.ai import AIAnalyzeSentiment
df = spark.table("customer_feedback")
result = AIAnalyzeSentiment(inputCol="feedback_text", outputCol="sentiment").transform(df)
That is it. One line. No API keys, no endpoint configuration, no model selection. The default model is gpt-4.1-mini, but you can configure alternatives including Claude and Llama through the AI Functions configuration.
Multimodal support: AI Functions can process images, PDFs, and text files alongside your tabular data. Pass column_type="path" and the function reads the file directly from OneLake. Supported formats include JPG, PNG, GIF, WebP, PDF, MD, TXT, CSV, JSON, and XML.
Concurrency: AI Functions support up to 200 concurrent requests by default, which is tunable for large-scale processing.
Way 3: Custom Model Integration via SynapseML
When you need a model that is not in the pre-built Foundry Tools — or you need full control over model parameters — SynapseML is the bridge.
SynapseML is Microsoft's open-source library for scalable machine learning pipelines on Apache Spark. It integrates with Azure OpenAI, Azure AI Services, and the broader Foundry model catalog. You can call any model in the Foundry marketplace from a Fabric notebook, including models hosted by third parties.
Use cases where SynapseML matters:
- Custom fine-tuned models that you have trained on your own data
- Third-party models (Claude, Grok, Mistral) that you want to call from Fabric
- Complex multi-step pipelines where different models handle different stages
- Batch scoring of large datasets with models that do not have pre-built AI Function wrappers
The pattern is straightforward: provision your model endpoint in Foundry, then call it from SynapseML in your Fabric notebook. The output lands back in OneLake, where Power BI can consume it through Direct Lake mode.
Way 4: Multimodal AI — Images, PDFs, and Documents in Your Data
Most Power BI users think of AI as a text-processing tool. Foundry AI changes that with multimodal support.
You can now process images and documents alongside your tabular data. Practical examples:
- Product images: Classify product photos by category, extract text from labels, detect defects
- Invoices and receipts: Extract line items, totals, and vendor information from PDF invoices
- Customer emails: Analyse sentiment, extract action items, classify by topic
- Contracts: Identify key terms, extract dates and amounts, flag risk clauses
The workflow is simple: store files in OneLake, reference them by path in your AI Functions, and the LLM processes them alongside your structured data. Results land in a new column in your DataFrame, ready for Power BI visualisation.
This is genuinely new capability. Until recently, processing unstructured files inside Power BI required building a separate Azure Function or Logic App. Now it is a single function call in a Fabric notebook.
Way 5: Data Agents with Foundry Model Backing
Fabric Data Agents (formerly AI Skills) let you build conversational Q&A systems over your Fabric data. What makes them powerful is that they can be backed by any Foundry model, not just the default GPT option.
A Data Agent connected to your semantic model can answer questions like "What was our revenue growth in Q3 by product category?" by translating natural language into DAX, executing it against your model, and returning a structured answer.
The agent supports:
- NL2SQL — Natural language to SQL queries against lakehouses and warehouses
- NL2DAX — Natural language to DAX queries against semantic models
- NL2KQL — Natural language to KQL queries against eventhouses
- Microsoft Graph queries — For organisational data
When you connect a Foundry model like Claude or GPT-5 to your Data Agent, you get the model's reasoning capabilities applied to your specific data context. The agent respects your existing security model — it only returns data the querying user is authorised to see.
Data Agents can be published to Microsoft 365 Copilot, Teams, and Copilot Studio, making your data accessible from any workspace your users already inhabit.
Cost Considerations and Capacity Planning
Foundry Tools in Fabric are billed against your Fabric capacity in CU (Capacity Units). The cost depends on:
- Model used: GPT-5.1 is more expensive per token than gpt-4.1-mini
- Volume: AI Functions process rows in batches; 100K rows of sentiment analysis costs more than 1K
- Concurrency: Higher concurrency settings consume more CU
Recommendations for cost management:
- Start with gpt-4.1-mini for AI Functions. It is the default for a reason — fast, capable, and cost-effective.
- Use ai.stats to monitor token usage. The built-in progress bar shows cost estimates as functions run.
- Batch your processing. Run AI Functions in scheduled notebooks rather than on-demand for large datasets.
- Set up Fabric Copilot Capacity (FCC) if Copilot usage is heavy. It separates AI costs from general capacity.
- Consider Claude for long-context tasks. Foundry's model marketplace lets you pick the right model for each job. Claude's large context window can be more cost-effective than GPT for document-heavy workflows.
Getting Started
The fastest path to Foundry AI in Power BI:
- Ensure you have Fabric F2+ capacity
- Open a Fabric notebook in your workspace
- Import SynapseML and call an AI Function on your data
- Save the output to OneLake
- Build a Power BI report on top of it
You can go from zero to AI-enriched data in under an hour.
Conclusion
Microsoft Foundry AI is not a single feature — it is the foundation that makes every AI capability in Power BI and Fabric possible. From one-line AI Functions to custom model orchestration with SynapseML, the platform gives Power BI teams access to world-class AI without leaving their Fabric workspace.
The organisations that will get the most from Power BI in 2026 are not the ones with the most reports. They are the ones that enrich their data with AI before it ever hits a visual. Foundry makes that possible.
Ready to bring Foundry AI into your Power BI workflow? Book a discovery call with powerbi.ai and we will show you what is possible with your data.