One Copilot question across two Power BI models needs a shared contract

A customer wants to ask one question about data held in two Power BI reports. The first instinct is to search for a connector that can query both. A recent practitioner discussion exposes the deeper problem: the assistant must know whether the two models describe the same customers, dates and measures before it combines their results.

Microsoft's Fabric ontology guidance describes combining concepts from multiple semantic models; the capability is marked preview and should be tested in the intended tenant. A cross-model ontology can help express relationships. It does not make conflicting business definitions disappear.

Resolve four meanings before adding AI

First, define entity identity. Does Customer mean an account, a billing entity or an individual? Which keys are stable, and how are duplicates or mergers handled? Second, define measure meaning. “Revenue” may be gross in one model and net in another. Third, define time. One model may use order date and another posting date. Fourth, define security. A user allowed to see one model may not be allowed to see all rows in the other.

Write these rules as a shared contract. The assistant should cite which model supplied each number and refuse to calculate a joined answer when the keys or definitions cannot be reconciled safely. A confident combined answer with no provenance is a reporting defect.

Build a small answer test set

Choose ten real questions across the two domains. For each, record the expected entities, measures, filters, source model and whether a combined answer is actually valid. Include two deliberately unanswerable questions, such as a join with no trusted key or a metric that differs by definition.

Test the assistant against those cases with representative permissions. Inspect the generated query or evidence path where the product exposes it. The pass condition is not only a correct number. It is a correct explanation of which definitions were used and what was left unresolved.

Plan for model drift

When a semantic model changes a measure, relationship or field name, rerun the answer set. Assign an owner to approve any new cross-model mapping. Without that control, an AI feature can turn an ordinary model change into a large, quiet reporting error.

Microsoft's Fabric data agent ontology guidance can inform the implementation, but the buying question is operational: who owns the shared meaning and tests it after every change? If nobody does, start with a smaller, single-domain assistant until that ownership exists.

Adovo’s view

An AI answer can sound unified while the underlying business definitions still disagree. Fluency is not reconciliation.

Before asking one assistant to cross two models, decide who owns shared concepts such as customer, revenue and time. Record when the models can be joined and when the honest answer is that they cannot. An ontology can make an approved relationship explicit and reviewable; it cannot approve that relationship on the business's behalf.