The agent can count delays. It still cannot cancel the flight.

The question

The story I keep hearing is that Power BI's next shape is a prompt-first Desktop, and that a chat pane good enough makes the semantic model optional. I don't buy it. If the stack keeps growing web apps, Copilot panes, and ontologies, what still has to be a named object before I let an agent near a decision?

The Next Evolution of Power BI — A Conversation with Mohammad Ali is the source. Ali is Partner Director for the Power BI product. His answer is not "type more." It is a split most teams blur: what is the answer, versus what the agent is allowed to do.

Two jobs sitting on the lake

Ali describes a nested picture, but the part I can use on a Monday is two jobs on OneLake.

The semantic model encapsulates facts and dimensions. Revenue has a definition. Tables relate on purpose. That is a mathematical exercise: delayed flights, check-ins, people on a plane. He says the semantic model is going through a rebirth because both humans and agents need that classic definition, and he wants the meaning closer to the lake so SQL and Python can hit the same object, not only a report. He also cites the installed base (about 20 million active semantic models) as the reason most customers are already in that middle stage.

An ontology is for a process. Seeing delays is not the same as cancelling a flight. The second object holds the decision matrix an agent would walk: rules, sequence, action. Ali's line is the one I will keep. The semantic model tells you the answer; the ontology tells the agent how to act. When people ask which to build, he usually says they want both.

I would still start in the middle. Ontology is preview. Jumping to a master process graph while "revenue" is still a vibe in a measure is how you get a confident agent with a wrong total.

Prompt-first is not the default path

Reza asked whether Desktop becomes a prompt-first window. Ali said maybe, later, and probably on Fabric Web Apps rather than JSON-shuffling the interactive report you already have. Prompt fatigue is real. He does not want that UI forced on people who do not find it natural.

That matches how I actually build. Nordic Boardroom still starts with gold tables and a semantic model, then a reader path from "are we okay?" to "who do we call." A prompt that rearranges visuals does not replace the grain of a KPI. Fabric Web Apps can skip PBIR and go straight to HTML, hosted on capacity, with authentication and DAX scaffolding included. Ali treats that as complementary. Conversation and custom apps sit next to reporting; they do not retire the model. Capacity-only, not Pro, so it is also not free.

Copilot-in-the-pane versus Claude-via-MCP is the same lesson in a different costume. Ali's path is to standardise on the MCP server internally so Power BI Copilot, M365, Foundry, and an external model hit one Fabric IQ door. Until that lands, I treat every extra pane as another client of the same definitions, not a second definition of revenue.

What I already refuse to blur

Orbit's weekly Write is a small version of the same split. Signals and project state are intake: what showed up, what is in the workshop. Slack Approve is the action. The essay can be rewritten; publication waits on a person. I would not let a generator both invent the claim and merge it.

Ledger is the limit I live with. A research cycle can name a Sharpe gap and a proposed trim. I still click the trades. No broker automation. That is process-shaped work I have not encoded, on purpose.

The cost of Ali's picture is coordination. Two objects means two kinds of change. Semantic edits are meaning. Ontology edits are permission to act. Mixing them in one prompt looks fast until you cannot tell whether the agent lied about the number or about the rule.

Takeaway

When the stack moves, I would not start by asking whether Desktop should open as a chat. I would ask whether "what is the answer?" has an object an agent can query without inventing grain, and whether "what may we do?" is a separate object I am willing to maintain. Most organisations will want both. Ontology is still preview. Skip the middle doll and the agent can count delays all day. It still should not cancel the flight.