Microsoft has rolled out a series of innovations for Power BI and Microsoft Fabric at Build 2026 and the June release under the heading “agentic era of analytics”. Much of this is preview and irrelevant for SMEs for the time being. Two things are explicitly not – and they have little to do with the marketing term. We have reviewed the announcements against the official product documentation and sorted by what you should use immediately, what you can evaluate and what you can safely ignore.
Implement immediately: the waste paper basket for workspaces
The most important innovation is the most unspectacular. Item Recovery is generally available (GA): Deleted fabric objects – reports, semantic models, lakehouses – can be recovered via the portal or via REST API, with a configurable storage of up to 90 days. Without additional costs.
Microsoft closes Fabric’s most painful governance gap. “Someone cleaned up the workspace” has been a case for the backup, which usually did not exist. If you take only one thing from this article: Configure the retention period before you need it.
Also GA and immediately useful for grown models: DAX User-Defined Functions. Reusable, typed DAX functions with optional parameters – the end of the copied measures, which multiply over years in slightly different variants.
Evaluate: Reports arising from prompts
Behind “Agent Skills for Power BI” (Preview) is not the button in the portal you expect. It’s a bundle you can install in your own AI coding wizard: instruction sets for semantic modeling and report creation, plus tools that can read schemas, run DAX, edit models, and remotely control Power BI Desktop.
The really interesting point is a structural one: The work happens in the PBIP/PBIR text files. This makes versioning in Git, Code Review and CI/CD for Power BI artifacts realistic for the first time – something that has been painfully missing in BI projects for years. The price: You need a developer tool chain (CLI, Node.js). For the department, this is nothing, for a BI team with a development background.
Related to this is Skills for Fabric – instruction sets published by Microsoft under an MIT license on GitHub that teach AI tools how Fabric actually works (which APIs, which query syntax, which auth). Support includes GitHub Copilot, VS Code, Claude Code, Cursor and Windsurf. Important to understand: These skills are Knowledge, not executive service.
Our recommendation: Try in a dev workspace with its own service principle – not on the production environment, and not with the rights of the administrator.
Evaluate: alarms become events
Activator can now publish (preview) structured business events if a condition applies in a Power BI report, real-time dashboard, KQL query, or warehouse SQL query. The difference to the previous “alarm by e-mail” is fundamental: A notification becomes an event with a defined scheme, which any number of recipients can subscribe to – Teams, Power Automate, a notebook, a copy job – without anyone having to touch the source. And every published event automatically ends up in the eventhouse, so that the question can be answered afterwards: How often has this actually happened?
The typical SME case: A threshold in the report – inventory, contribution margin, open items – generates an event that triggers a notification and triggers an automatic follow-up process.
A catch you need to know: The processing of business events has been billed via the normal capacity model since June. On a small F2 or F4 capacity, a too talkative rule can consume noticeable capacity, which is then missing from other workloads.
First check, then grab: real-time data from the ERP
The new Connector Mirrored Database (Change Feed) (Preview) brings insertions, changes and deletions from a mirrored database into an event stream in real time. It supports Azure SQL Database, Cosmos DB, Snowflake and Open Mirroring.
Before you evaluate this, a pragmatic note: If your source is a local SQL Server, an Azure SQL DB or a Managed Instance – and this is the norm in SMEs – then CDC support in the copy job is the better way. This is already widely available, while the Change Feed connector is still preview, does not allow table selection and incurs additional costs (the Delta Change Data Feed is a paid extension and increases memory usage).
Ignore: Azure HorizonDB
The new PostgreSQL-compatible database with built-in vector search is making headlines. For German SMEs, it is not an issue in July 2026, and this can be precisely justified: it is a preview, there is no region in Germany or Western Europe (As of July 2026: US regions, Sweden, Australia), there are no customer-managed keys, no cross-region replicas and a fixed backup storage of seven days. For personal data, this is a clear rejection. Observe – not evaluate.
And the licensing issue
The co-pilot functions in Fabric require at least one F2 capacity (or P-SKU) and do not run on trial capacities. For customers with a capacity within the EU Data Boundary, the processing on the data protection side is clean – no processing outside the region is necessary. In sovereign clouds, Copilot is not available.
Conclusion
The agent functions are the more exciting part of the announcement, but the benefits lie elsewhere in July 2026: the trash you configure today and the DAX functions that declutter your grown model. This is less exciting – and pays off immediately.
If you are reorganizing your BI landscape anyway: We show on our page to Business Intelligence & ETLhow we connect data sources and set up dashboards.
All data verified against the official Microsoft product documentation (Microsoft Learn). Preview functions may change in the short term. Status: 14 July 2026.
Would you like to implement this in your company? We support you pragmatically – from the idea to the operation.