Ask a question the way you'd ask a colleague, and get the number back from the reports you already have. Claude reads your existing Power BI. Nothing new gets built, and there's no copy of your warehouse to maintain.
Sales for Q3 so far (1 July–18 September) are £1.84M across 412 invoices, counted on invoice date.
That is 11% ahead of the same period last year, which finished at £1.66M. The quarter isn't closed, so this figure keeps moving until the end of September.
Nobody here is short of dashboards. What people are short of is a fast route through them. So the question goes into a chat message instead, and somebody on the BI team stops what they were doing.
One figure prompts three follow-ups. Each one is another pass through the report, or another message.
Time that should go into the next model goes into questions the current model can already answer.
When a question costs fifteen minutes and somebody else's attention, the marginal one doesn't get asked. That's often the one worth asking.
A screen recording from a live Power BI environment. Same reports, same permissions, no test data. Watch what happens between the question and the answer: Claude finds the report, reads the semantic model, writes the query, then tells you what the number does and doesn't cover.
Everything you see happens inside Claude. Power BI is never opened, and the report itself is never modified. Claude is reading the same semantic model the dashboard reads.
The connection runs on Microsoft's Remote Power BI MCP Server. Microsoft runs it. There is no middleware for your team to host, no copy of your warehouse, and no new semantic model to maintain.
Microsoft ships a Remote Power BI MCP Server. It gets enabled in your Fabric tenant settings by whoever already administers Power BI. That's a setting, not a build. We review the environment with them first and confirm what needs to be true before anything else starts.
An Entra ID app registration handles the connection, with delegated Power BI permissions and OAuth configured against your tenant. We test the sign-in flow properly rather than assuming it, because how it behaves depends on how your Microsoft estate is already set up.
Where your Claude organisation supports it, the Power BI connector is configured centrally. Individual users set nothing up. They sign in with the Microsoft account they already have, and their existing Claude and Power BI access carries over unchanged.
Claude reads the semantic models sitting behind your current reports. There is no new consolidated model to design, no second version of the truth to keep in sync, and your reports stay in the workspaces they're in today.
A question goes through MCP to the relevant semantic model. Power BI evaluates it under that person's permissions and Row-Level Security, then returns the result to Claude, which explains it. Claude is working from the most recent successfully refreshed data.
Swipe the diagram to see the whole path
This is usually the question that decides whether a project like this goes ahead, so it's worth being precise. Claude has no standing access to your data. Every question is evaluated as the person who typed it.
Claude inherits the Power BI access the person already has. If you can't open a report today, you can't ask about it tomorrow. Take someone's access away and their answers stop, with no separate step to remember.
RLS is enforced by Power BI when it evaluates the query, not by Claude afterwards. A regional manager asking a national question gets their region, exactly as they would in the report.
Each person authenticates with their own Microsoft Entra account. There is no shared service account with a view of everything, and no credential to pass around.
It reads. It cannot edit a report, change a model, or administer Power BI Service. We start here on purpose, and the security conversation is shorter for it.
Nothing on this list needs a migration, a new licence tier for your warehouse, or a training programme. The reports stay where they are.
No report to find, no filters to set. The recording covers a question, an answer and a follow-up in two minutes and twenty-two seconds. You can time it yourself.
Questions that used to arrive as chat messages get answered from the model directly. The ones that genuinely need an analyst still reach one.
Second and third questions are where most of the value hides, and they are exactly the ones people currently skip.
In the recording, Claude flags unprompted that a quarter-to-date figure is not a completed quarter. A dashboard tile shows that same figure with no such warning, and people quote it in meetings.
It reads your existing semantic models. No second model to reconcile, and no shadow reporting layer quietly drifting away from the real one.
Your reports, your permissions, your Microsoft directory. The only new habit is typing the question instead of clicking your way to it.
We'd rather you heard this now than found it in week three. These are real boundaries of the setup, not caveats in small print.
The upside of a narrow first phase is that there is very little that can go wrong. Editing and administration stay a later decision, taken on evidence rather than on hope.
The integration in that recording is running today across four dashboard areas, for a client whose reports sit on ERP and cloud data. Executives and operational managers use it against live company numbers. The answers get checked against the reports, and they hold up.
Which means the awkward parts, the tenant settings, the Entra registration, the OAuth behaviour that differs from one Microsoft estate to the next, are problems we've worked through once already. That is usually where a first rollout stalls.
What they ask it, day to day, is unglamorous and useful:
Inventory reporting
Inventory reporting
Company and sales reporting
Manager views, filtered by RLS
We don't switch on an AI assistant for a whole company on day one. Each phase ends with something you can check, and nothing moves forward until it holds.
We review your Power BI estate, confirm the admin access we need, enable the Remote Power BI MCP Server and prepare the Entra registration. Ends when the plumbing is ready for connection testing.
Claude is connected to the approved semantic models and tested against questions from your real dashboards: inventory, sales, customer spend, backorders. We confirm permissions and RLS still apply through the new route.
Testing runs on your live environment rather than a sanitised copy, because you need to check answers against numbers you already trust. We test across roles, including a restricted manager view.
Access extends to the remaining approved users. You get documentation covering the setup, the access process and the key configuration points, plus a walkthrough with your team.
Your reports and semantic models stay exactly where they are. A question is sent to the relevant model, Power BI evaluates it under the asker's permissions, and the result comes back to be explained. There is no copy of your warehouse sitting somewhere else, and nothing is written back.
No. This connects to the models behind your current reports. Building a new consolidated model is a common and expensive first instinct, and it isn't required here.
The same thing that stops them today. Power BI applies their permissions and Row-Level Security when it evaluates the query, before any result is returned. Claude has no access of its own to fall back on.
Your Power BI or Fabric administrator enables the tenant setting, and someone with Entra ID rights helps with the app registration. After that it's mostly us, plus a few of your people to test answers against reports they already know well.
You could. The tenant setting and the app registration are documented, and if you have a Fabric administrator and someone comfortable in Entra ID, it's a reasonable internal project. What you'd be paying us for is having already hit the parts that aren't documented, and knowing what to validate before anyone is let loose on live data.
As current as your last successful refresh. Claude reads the most recent refreshed data in Power BI. It isn't querying your source systems directly, and it can't see whether a scheduled refresh failed.
Not in this setup, which is read-only on purpose. Editing reports and administering Power BI Service is a separate piece of work using the Fabric tooling, and it's a decision better taken once the reading side is proven.
Start with a small group. In the live deployment it began with the owner and a handful of operational and sales managers. Once the connection and the permissions behave, extending it to more approved users is an access change rather than a new project.
Then we'll tell you, and we'd suggest fixing that first. Claude reads your model as it is: if a measure is defined wrongly, the answer inherits the error. This works best on top of modelling that's already trusted.
The most useful thing we can do is spend half an hour looking at what your Power BI estate actually contains, and give you a straight read on what is involved and where the effort actually sits.