Insurance: from claims intelligence to governed intervention
Insurance is where this architecture becomes practical very quickly. A carrier can use OneLake, mirroring, and shortcut-based ingestion to bring together policy administration, billing, claims, vendor, and event data without building yet another integration maze. At FabCon, Microsoft said mirroring for Oracle and SAP Datasphere are generally available, and in April it announced shortcut transformations as generally available for turning files referenced through OneLake shortcuts into Delta tables without building pipelines. For insurers still living with a mix of core systems, partner feeds, and file-based handoffs, that is not a small detail. It is what makes a shared substrate for agentic workflows more realistic.
From there, the right insurance model is not “one agent for everything.” It is a governed network of specialized agents grounded in a common business language. An ontology can define Policyholder, Policy, Claim, Coverage, Exposure, Adjuster, Repair Vendor, and Fraud Signal as shared business entities tied to live data. Graph is explicitly designed for relationship-heavy questions, which is exactly what fraud rings, claimant networks, vendor relationships, and catastrophe exposure chains look like in practice. A claims data agent could answer which catastrophe claims are drifting past service targets, which open files show reserve movement without corresponding progress, or which claims share suspicious counterparties. An operations agent can then monitor real-time data and recommend next actions when patterns emerge.
The more interesting change since FabCon is that the action layer is getting sharper. Ontology Rules with Fabric Activator let teams define conditions and actions on business entities in business language. Business Events in Real-Time Intelligence can generate events from notebooks and user data functions, then route them into alerts, workflows, or downstream automation. Microsoft’s post-FabCon Fabric IQ update also points toward more usable enterprise controls by adding Azure Private Link support and promising public MCP endpoints for ontology. For a carrier, that means the platform is moving closer to a state where “detect a claims exception,” “explain it in business terms,” and “route it into a governed workflow” can happen inside one connected environment rather than across disconnected tools.
Control and release discipline matter just as much as clever prompts in insurance. Current Fabric data agent guidance now covers publishing, sharing, versioning, and Git/deployment-pipeline-based ALM. Microsoft’s sharing guidance also says data agents honor underlying permissions, including row-level and column-level security, and that users only need Read permission on semantic models for query scenarios through a data agent. That is the kind of practical detail regulated industries care about, because the real production question is never only “can the agent answer?” It is “can we control who sees what, test changes safely, and promote updates without breaking trust?”
Wealth management: from advisor assistant to advisor infrastructure
Wealth management has a different rhythm, but it has the same underlying need: a governed household view across accounts, mandates, model portfolios, cash, service cases, meetings, and research. Here, the most important shift since FabCon may be distribution. Fabric data agents can be published into Microsoft 365 Copilot, where they appear in the Agent Store and can be used directly in Teams, and they can also be consumed from Copilot in Power BI alongside reports and semantic models. Power BI’s standalone Copilot experience can search across reports, semantic models, and Fabric data agents a user already has access to. That is a much stronger operating model than asking advisors and service teams to leave their workflow and hunt through a portfolio of separate dashboards.
Fabric IQ also adds two capabilities that matter disproportionately in wealth management. First, graph is built for relationship-heavy analysis, which is exactly what you need for households, trusts, beneficiaries, legal entities, referral networks, and concentration chains. Second, plan in Fabric IQ brings budgets, forecasts, and scenarios onto the same governed platform as analytics and AI, with support for shared semantic models and write-back to Fabric SQL databases. In practice, that means a wealth firm can start thinking beyond static book-of-business reporting and toward scenario-aware planning around advisor capacity, flows, fee revenue, retention, and client cash needs without leaving the Fabric estate.
That opens the door to a more credible advisor cockpit. A data agent can help identify households drifting from target allocation, clients with high idle cash and no outreach, or service exceptions likely to damage retention. An operations layer can monitor review-cycle breaches, suitability deadlines, onboarding bottlenecks, or transfer delays in real time. And the human-in-the-loop step is better now than it was even a year ago: translytical task flows are generally available in Power BI, Direct Lake in OneLake is generally available, and user data functions can be invoked from Power BI, pipelines, notebooks, and Activator rules. For wealth firms, that means the same platform can surface the exception, explain it in context, and pass a controlled action into a workflow without forcing a handoff into a separate stack.
What the post-FabCon updates really change
The biggest change since FabCon is not that Fabric suddenly solved agentic AI for financial services. It is that Microsoft is filling in the production details that serious firms look for before they move beyond pilots. Data agents are now positioned as generally available. ALM guidance now exists. Fabric is leaning harder into MCP-based interoperability. Ontology is becoming more operational. Planning has entered the picture. And the action surfaces in Power BI and Real-Time Intelligence are getting stronger. That combination is what turns “ask your data a question” into “ground an answer, watch the business state, and route the next action.”
The caution is still important. In current Microsoft documentation, several of the surrounding capabilities that make the full vision compelling remain in preview, including ontology, plan, graph, operations agents, and ontology rules. For insurance and wealth management, that should not be read as a reason to wait for perfection. It should be read as a design signal. Start with narrow, high-value workflows where evidence, permissions, escalation, and monitoring matter more than autonomy. In claims, that might be catastrophe triage or fraud escalation. In wealth management, it might be household review prioritization or service-risk detection. The first win is not autonomy. The first win is dependable, governed assistance.
Conclusion
The question is no longer whether Microsoft Fabric can host an agentic demo for financial services. That part is easy. The better question is whether Fabric can support the messy middle where regulated work actually lives: shared semantics, changing conditions, human approvals, audited actions, and iterative release discipline.
Since FabCon, the answer looks more credible than it did even a few weeks earlier. In insurance, that points toward claims, fraud, and service operations that move faster without shedding control. In wealth management, it points toward advisor intelligence that is closer to the workflow, more scenario-aware, and better grounded in the client and household context that actually matters. The opportunity now is to stop asking where to place a chatbot and start asking which governed workflow deserves the first real deployment.
