Copilot in Fabric

What it is

Copilot in Fabric is the embedded AI assistant across every workload. It generates code (DAX measures, Spark notebooks, KQL queries, T-SQL stored procedures, M expressions), summaries (notebook narration, dashboard insights), and structure (semantic model relationships, report layouts). It runs under your identity, respects your permissions, and uses the data you can already access.

Where it shows up

• Power BI — generate visuals, DAX measures, and narratives from a model
• Notebooks — code generation in Python, Spark SQL, and KQL
• Warehouse — T-SQL generation and rewrites
• Data Factory — pipeline activity suggestions, Dataflow M generation
• KQL queryset / Real-Time Dashboard — KQL generation
• Data Agents — the Copilot framework that powers conversational answers

Tenant governance

Copilot has tenant- and capacity-level controls. The platform admin chooses which workspaces, users, or capacities can use Copilot. Cross-geo processing settings determine where prompts and responses are processed — important for regulated industries. Always review:

• Enable / disable per workload, per workspace
• Cross-geo processing and storing
• Tenant-level data exclusion lists
• Purview integration for audit logs of Copilot interactions

Prompt patterns that work

• Be specific about the dataset. "Write a DAX measure for YoY revenue using the Sales table and the Date dimension" beats "do YoY."
• Show shape, not just intent. Paste a sample of the source schema or a few rows of data.
• Ask for explanations. "Then explain what you did" — accelerates your learning and catches errors.
• Iterate. The first response is rarely the final answer. Critique it; Copilot refines.

Best practices

• Always review generated code. Copilot is a productivity tool, not a substitute for the engineer.
• Adopt patterns, not snippets. If Copilot keeps generating a useful template, save it as a notebook template or Tabular Editor script.
• Use Copilot to learn KQL. Excellent for translating SQL intent into idiomatic KQL.
• Don't paste sensitive data in prompts. Even with tenant controls, treat prompts the same way you'd treat a Teams message — assume it's logged.

Common pitfalls

Related items