Microsoft’s September 28, 2026 data announcement focuses on making existing business data usable by Copilot and agents without replacing its definitions and ownership. The releases span generally available services, previews, and planned features, so a data-platform evaluation needs a capability-by-capability rollout check.
Business definitions become part of the AI context
Microsoft says Fabric IQ in Copilot Chat and Cowork is generally available and uses governed Power BI semantic models. IQ Sharing is in preview. SQL Server on Azure Local is generally available, while disconnected operations is a preview. Database Hub is in public preview; specialist database agents are described as coming soon.
The useful architectural distinction is between raw data and an agreed interpretation of it. A sales number can be mathematically correct yet answer the wrong business question if it uses bookings instead of recognized revenue. A semantic model gives an AI workflow a place to retrieve the same definitions used in reports. That reduces one source of ambiguity; it cannot make an unclear question unambiguous.
Sharing context does not settle permission design
The Copilot product announcement places governed data alongside delegated work and app creation. In an enterprise rollout, the difficult case is often a task that joins information from several owners.
Before enabling a shared workflow, test a real user who can access one dataset but not another. Verify that summaries, exports, and generated applications preserve that distinction. Cross-company sharing also needs a lifecycle: who can revoke access, how quickly revocation applies, and what happens to an already-created artifact.
Database assistance needs a separate mutation boundary
A database assistant can help investigate a slow query or summarize workload behavior without receiving permission to alter production state. Preserve that separation when evaluating forthcoming agents. An explanation, a proposed query, and an executed change should be distinguishable in the interface and audit record.
Start with an investigation workflow where an engineer reviews the result. Include misleading correlations and incomplete telemetry in the trial so the team can see whether the assistant reports uncertainty. The goal is better operational evidence, not automatic acceptance of a plausible diagnosis.
Evaluate a governed answer end to end
Choose a frequently disputed business question and define the expected source, metric, permission scope, and freshness. Compare the assistant’s answer with the canonical reporting path, then test an outdated document and a user without access. This provides a more useful adoption decision than a broad claim that company data is now AI-ready.