Your best people spend two hours a day being a reporting engine.

With one question, not a platform. Pick the recurring question that costs your team the most hours this month, and trace what answering it actually requires — which systems, which definitions, which exceptions. That trace tells you whether your gap is the warehouse, the semantic model or the agent, and it becomes the acceptance test for the whole engagement. Every deployment on this page began that way. (New to the category? Start with what a data agent is and how it works.)

The askers rather than the builders — leadership, operations and facility managers, finance and commercial teams. The people who today wait on an analyst or export to Excel. They work inside Microsoft 365 Copilot, so there is no new tool to adopt, while your data team keeps owning the semantic model underneath. (For how this differs from Fabric's builder-facing Copilot features, see Fabric Copilot vs Data Agent.)

Almost always the semantic model, not the language model. If the agent has to guess which table holds revenue, which of four date fields means June, or which department column is authoritative, it will guess. The fix is model-side: business naming, validated relationships, descriptions, verified answers and explicit AI instructions.

Usually, yes — and it is the first thing we look at. Cost tracks how much reasoning the agent must do before it can answer, so removing ambiguity from the model removes spend. On one deployment that took consumption from roughly 5,000 CU per question to about 1,000 CU with no capacity change. For how Fabric bills data agents and how to size capacity in the first place, see what Fabric data agents cost and how to size your capacity.

In practice, yes. Pointing an agent at raw operational tables produces ambiguity, slow answers and high compute cost. A Bronze/Silver/Gold warehouse in OneLake with a star-schema semantic model is what makes plain-English questions reliably answerable. Our Business Central → Fabric write-up shows exactly how that foundation is built.

Credentials live in Azure Key Vault, never in notebooks or code. Access is via Entra ID app registration with least-privilege permission sets. On the agent itself, the table is the unit of access — individual columns cannot be excluded — so table selection is the security boundary and is scoped deliberately against approved use cases.

Four things, in order: a semantic model readiness assessment of what you already have; the data warehouse and modelling work if the foundation is not there; the agent build — schema simplification, verified answers, AI instructions, scoping and Copilot publishing; then an enablement loop where we review real user questions and enrich the model weekly. Unlike general AI consulting, this is a data engineering engagement with an AI deliverable at the end — which is why the semantic model, not the prompt, is where we spend the time.

In-house works well when you already have a governed warehouse, a Power BI team fluent in semantic modelling, and someone who can own Fabric capacity management. Bring in help when the data foundation is the gap, when you are on your second failed agent pilot, or when you need the cost-per-query problem solved rather than diagnosed. Either way, insist the work is judged on model quality, not demo quality.

Overlapping, not identical. “AI agent for data analysis” and “data analysis agent” usually describe exactly this pattern — natural-language questions answered against a governed dataset. Broader agentic AI adds autonomous multi-step action: an agent that not only answers but triggers workflows. A Fabric data agent is deliberately narrower and read-only over selected tables, which is what makes it governable in an enterprise. See our explainer on AI agents for data analysis for where the category's accuracy limits actually sit.

The recurring loop is: a business user asks a question in Copilot; the agent resolves it against the semantic model and returns the answer with the steps it took; unanswerable or ambiguous questions are logged; our team reviews that log and enriches the model — a description, a relationship fix, a verified answer, an AI instruction. Over weeks, the share of questions answered first-time climbs and the compute cost per question falls.

Anything with an API or a supported connector. Published implementations include Dynamics 365 Business Central, QuickBooks and Xero, plus SQL, Azure and operational systems on the enterprise side.