Real-time business context for SMEs on the Microsoft stack

Most SMEs already run on Microsoft: Business Central or another ERP, SQL Server, Power BI and Microsoft 365. The lowest-operating-cost way to give people and AI agents live business context is to stay inside it: mirror the source systems into OneLake, define the metrics once in a Power BI semantic model, and let agents query that model through Fabric IQ.

Step 1: decide what “real time” means

NeedFreshnessExample
Operationalsecondsstock level at the order desk
Managementminutes to hourlytoday’s orders, open quotes, cash position
Analyticaldailymargins, customer segments, trends

Most SME questions are in the middle row. Mirroring gives near-real-time freshness for it without a streaming platform.

Step 2: capture changes from the source systems

SourcePathStatus
Business Central (online)Native mirroring into a Fabric mirrored database. The user picks companies and tables; Business Central sends an initial snapshot and then ongoing changes to OneLake as Delta Parquet. One source gives up to 1,000 tables and a minimum Business Central online version of 29.0.Prerelease. Partner sources disagree on dates: Microsoft Learn describes a production-ready preview in October 2026, the release roadmap lists preview in November 2026 and general availability in April 2027.
Business Central, interimAzure Data Factory’s Dynamics 365 connector, or OData and API pulls, delivering to OneLake or ADLS Gen2Available now, but scheduled batch rather than continuous
SQL Server 2016 to 2022Fabric mirroring using Change Data Capture; works on-premises, on Azure VMs and on other clouds, with an on-premises or virtual network data gateway behind a firewallGenerally available
SQL Server 2025Fabric mirroring using the change feed; on-premises only, not on Linux or Azure VMsGenerally available
Azure SQL Database, Azure SQL Managed Instance, Azure Cosmos DB, SnowflakeFabric mirroringSupported mirroring sources
[[./dataverseDataverse]] (Dynamics 365 customer apps, Power Apps)Link to Microsoft Fabric: a Delta Parquet replica in Dataverse storage, shown in OneLake as shortcuts, with no storage account or Spark pool to run

Power Apps that need live Business Central data can read it in place through Dataverse virtual tables, which do not copy the data; analytics should still use mirroring.

Mirroring means no pipelines to build or run for these sources. Business Central mirroring itself has no separate cost; the user pays for Fabric capacity.

Step 3: one store and one small capacity

  • OneLake holds the mirrored tables in an open format, so OneLake is the single store; no separate warehouse is needed at SME volumes.
  • Fabric capacity is the one running cost. The smallest F SKU is F2. Direct Lake mode needs a Fabric (F) or Power BI Premium (P) capacity; Import and DirectQuery work with Free and Pro licences.
  • Direct Lake guardrails — the SKU sets the maximum memory per Direct Lake semantic model. A model that exceeds it falls back to DirectQuery, which makes reports slower rather than failing them.
  • Replace Azure Synapse with this path for new work; Synapse is in maintenance mode.

Step 4: the semantic layer is the context

Fresh numbers are not context. Context is agreeing what “active customer”, “margin” and “on-time delivery” mean, so people and agents get the same answer.

  • Power BI semantic model — define each metric once as a measure (Power BI). Dashboards, Copilot and Fabric data agents all read it.
  • Fabric IQ ontology — adds business objects, relationships and rules on top. A first ontology can be generated from a Direct Lake Power BI semantic model; keys, time series and relationship bindings still need hand work. It references data where it lives and does not copy it.
  • Status — Fabric IQ is generally available; the ontology item and the MCP server are still in preview. Using the ontology as an MCP server needs an F2 or higher capacity.
  • Agent access — data agents, operations agents, Copilot Studio, Foundry IQ and MCP clients can use the same ontology and semantic models as governed, permissioned sources.

Compare with vendor-neutral options in semantic-layers-compared and Apache Ossie; they matter if the business later leaves the Microsoft stack, but they add something to run.

Step 5: delivery

  • Power BI reports and dashboards for the visible numbers
  • Copilot and Fabric data agents for ad-hoc questions, grounded in the semantic model
  • Teams and Microsoft 365 as the place people already work, through Copilot

What it takes beyond tooling

  • One named owner per source system who knows what the fields mean; this is usually the bottleneck
  • A short metric list to start: the 10 to 20 numbers the owner or manager checks weekly
  • Read-only agent access to the governed semantic model first, never to raw tables
  • Capacity sizing before rollout: test the real model size against the chosen F SKU’s guardrails

Sources