When the Agent Builds the Model: Governing Agentic Analytics in Power BI and Fabric
Underneath the demos in Microsoft's Build 2026 analytics announcements sits a detail that will decide whether agent-built BI can be governed at all: Agent Skills for Power BI works through Power BI Projects (PBIP), the file-based format that lives comfortably in source control. The agent doesn't click around a canvas; it writes artifacts a human can review as a diff and reject before they ship.
That detail deserves more attention than the headline capabilities, because the headline capabilities are broad and they are arriving fast.
What Microsoft announced
Two previews. Agent Skills for Power BI lets you describe what you need in plain language (or hand the agent a screenshot) and have it construct semantic models from your Fabric data, generate report pages, and iterate on visuals. It extends the Power BI Modeling MCP released in November 2025, which covered the semantic layer, into an end-to-end workflow that includes reports. It works on existing solutions as well as new ones, the output is aligned to design best practices according to Microsoft, and rollout starts the week after the announcement.
The second preview, Fabric Apps on Semantic Models, goes further: AI agents building and deploying Fabric-native web apps on top of semantic models, powered by the open-source Rayfin SDK. The framing is that enterprise analytics and operational applications have lived in separate worlds, and that a shared semantic model can now serve both, with Fabric supplying the data foundation, security and governance underneath. The example scenarios Microsoft picks are telling for finance readers: inventory management, financial planning, staffing coordination.
The model is code now, and that changes who can say no
When a human analyst builds a report in Power BI Desktop and publishes it, governance mostly happens after the fact: someone notices a number is off, or an access review finds a workspace nobody owns. A PBIP-based agent workflow inverts that. Every change the agent makes is a file diff, which means you can require a pull request, run automated checks against it, and refuse to deploy anything that fails.
At Syngenta, where I work on a Fabric lakehouse with CI/CD for BI assets, the pipeline effectively is the governance layer: row-level security and data quality standards run through it, and so do deployments. An agent that emits PBIP files joins that pipeline like any other contributor, subject to the same gates. An agent that only operated through a UI could not be governed this way.
Format determines governability, and of everything in the announcement, that is the part I would weight most heavily.
Review effort moves from syntax to semantics
Agent output will be syntactically clean almost by definition; the DAX will parse and the visuals will render. The risk profile in finance is different: plausible but wrong. Imagine a controller who asks an agent for "net settlement expense by cost centre". To build that measure, the agent has to choose a fact table, decide how accruals and reversals are treated, pick the point of currency conversion, and resolve what "net" means in this context. Any combination of those choices produces a chart that looks finished. Usually only one of them ties to the general ledger. After years spent on settlement expense allocation and variance analysis, my first review question for an agent-built model would be whether it ties out, not whether it runs.
The encouraging part is that this check automates well. A handful of key measures reconciled against a certified source (the general ledger, an audited dataset) as a merge gate catches the worst of it without slowing anyone down.
Fabric Apps widen the surface
The announcement says the cost and complexity of building applications is collapsing, and that nearly anything describable can be generated, with full control over UI, business logic and integrations with external systems via APIs. Take that seriously and the consequence is obvious: the number of applications in your tenant is about to grow much faster than your capacity to track them.
The governance inherited from the semantic model is real, and it is the strongest argument for building these apps in Fabric rather than beside it. But it covers the data layer. The app layer is generated code: business logic that may reinterpret governed numbers, and outbound API calls that move data somewhere else entirely. Shadow IT used to mean an Excel file with a macro; in this model it ships a web app on certified data, which is both better (the data is governed) and worse (the reach is bigger). Someone also has to own each of these apps twelve months from now, when the person who prompted it into existence has moved on to another role.
Controls worth having before the previews reach your tenant
None of this argues against the previews; it argues for sequencing. The controls below are standard analytics engineering, applied with less slack than a human-paced team could afford to leave.
| Failure mode | Control that catches it |
|---|---|
| Plausible but wrong measure logic | Reconcile key measures against a certified source before merge |
| Model edits that quietly break RLS | RLS test cases running in CI on every PBIP change |
| Duplicate models and report sprawl | Deployment gate that enforces reuse of certified models |
| Generated apps calling external systems | Code review of app integrations; least-privilege connections |
| No trail of what the agent changed | PBIP in git; agent commits reviewed like human ones |
A team that has these in place can let an agent move fast, because the pipeline does the distrusting on its behalf. A team that doesn't will be approving agent output by eyeballing rendered reports, which is exactly the review method that already fails with human-built ones.
Where this leaves a finance BI team
The productivity claim is credible. A faster path from Fabric data to a production-ready report, plus agents that can polish existing reports, lands squarely on work that BI developers in finance sink a great deal of time into: layout, boilerplate modeling, visual iteration. The hours that frees up are best reinvested in definitions and controls, since those are now the binding constraint on quality.
The full announcement, Building in the Agentic Era with Power BI and Fabric, is on the Power BI updates blog, with the on-demand Build session and the Fabric Apps documentation linked from it. Worth reading before the preview shows up in your tenant: the agents are arriving either way, and the difference between teams will be whether the pipeline they arrive into has gates.
Facing a similar challenge?
📅 Book a Free Call