AI architecture: query Power BI and Microsoft Fabric in natural language
Web App, Azure AI Foundry agent, MCP server, Power BI semantic model and Microsoft Fabric: an architecture based on our project for Alliance Marine, giving your teams conversational access to their data.
Review your architectureYour Power BI dashboards answer the questions someone anticipated. The others (“Which sales dropped this quarter, and in which regions?”) end up in the data team’s request queue. An AI agent connected to the data changes this: anyone asks a question in natural language and gets a quantified, sourced answer.
The answer still has to be right. A language model writing SQL on its own against raw tables easily gets a join or a definition wrong. Our architecture therefore relies on what the company has already validated: the Power BI semantic model, its measures and business rules, on top of Microsoft Fabric data.
The link between the agent and the data goes through an MCP (Model Context Protocol) server. It exposes a small set of well-defined tools to the agent, controls what can be read, and can later be reused by other agents or platforms.
The architecture, layer by layer
Five layers, each replaceable without rewriting the others.
- 01
Web App
Web application · Entra ID
The chat interface. Users sign in with their company account; an answer can display a table or a Power BI visual.
- 02
AI agent
Azure AI Foundry
The language model and its instructions. It understands the question, picks the tools to call and writes the answer.
- 03
MCP server
Model Context Protocol
The gateway to the data: a closed list of tools (find a measure, run a query, list dimensions).
- 04
Semantic model
Power BI · DAX
Validated business definitions: revenue, margin, scopes. The agent queries measures, not raw tables.
- 05
Data platform
Microsoft Fabric · OneLake
Centralised, historised and governed data, fed by existing pipelines.
Client case: Alliance Marine
For Alliance Marine, we designed and deployed a Web App, Azure AI Foundry and Power BI stack, with an MCP server enabling natural-language queries on data hosted in Microsoft Fabric.
This project was the basis for formalising the architecture described on this page: the design, security and cost-control choices below come from it.
Why use an MCP server?
An open standard
MCP is supported by a growing number of platforms, including Azure AI Foundry, Copilot Studio and Claude. Tools built once serve several agents.
Fine-grained control
The agent only reaches the exposed tools, with validated parameters. No free-form queries across the whole database.
More accurate answers
Tools query the semantic model’s measures: the agent gets the same figures as the Power BI reports instead of recomputing them its own way.
Data security and governance
Conversational access must not become a way around existing rules. Our recommendations:
- ✓ Authentication with Microsoft Entra ID, using existing company accounts
- ✓ User permissions enforced, including Power BI row-level security (RLS)
- ✓ Read-only MCP tools with controlled parameters
- ✓ Logging of questions, tools called and answers
- ✓ Services hosted in a European Azure region
FinOps: the cost of this architecture, under control
Aggregated answers
The MCP server returns results already computed by the semantic model, not thousands of rows: fewer tokens per question.
The right model at each step
A lightweight model to understand and route the request, a more powerful one only for a synthesis that justifies it.
A right-sized Fabric capacity
The load added by the agent is measured, then the capacity is adjusted, paused or reserved according to actual hours and volume.
A tracked cost per question
Resource tagging, budgets and alerts: you know what each use case costs and can make trade-offs.
Frequently asked questions about AI architecture
Can an AI agent be connected to Power BI?
Yes. An MCP server can expose tools to the agent that run DAX queries against a published Power BI semantic model. The agent thus gets figures consistent with existing reports.
Is Microsoft Fabric required for this architecture?
No, but it is the simplest case when the data is already there. The same principle works with another data warehouse: only the MCP server changes.
How is this different from Copilot in Power BI?
Copilot in Power BI helps explore a report or a model, inside the Power BI interface. Our architecture embeds the agent in your own application, combines several sources and tools, and leaves you in control of the model, the instructions and the costs.
Can the agent modify data?
Not in our reference architecture: MCP tools are read-only. Write actions can be added case by case, with user confirmation and logging.
Want to talk to your data?
Tell us about your data platform and your business questions: we will propose an architecture and an estimate of its running cost.
Discuss your project