Power BI
Dashboards answer questions somebody anticipated. Most questions people actually have were not anticipated, which is why they ask a person instead.
Dashboards answer questions somebody anticipated. Most questions people actually have were not anticipated, which is why they ask a person instead.
What Power BI holds, and why that matters here
Power BI holds the curated analytical view of an organisation — the measures, dimensions and reports that someone has already built and validated. Its weakness is not the analysis; it is that consuming it requires knowing which report exists and how to read it.
The result is a familiar pattern: dashboards are built, adoption is modest, and people ask an analyst instead. The analyst answers from the same data the dashboard shows.
- Datasets — the curated, validated models, which Microsoft now calls semantic models.
- Measures — the calculations already agreed as authoritative.
- Reports — the views that have been built.
- Row-level security — the rules governing who sees which slice.
- Refresh state — how current the data is, which matters more than most consumers realise.
Questions staff can ask of Power BI through Elbi
What was the figure last month?
A measure retrieved conversationally rather than by navigating a report.
How does this compare?
Period comparison using the agreed calculation.
Which is highest?
A ranked answer without opening anything.
Is that up or down?
Direction and magnitude in a sentence.
How current is this?
Refresh state, which prevents acting on stale figures.
Which report shows this?
Pointing to the right existing report rather than answering badly.
How a Power BI question is answered
- An employee asksThis is an internal use case. The person asking is a member of staff, signed in through your own identity provider, so the assistant knows who they are before anything is queried.
- The question meets the catalogueThe assistant works from an agreed list of semantic models and the measures in them, each with its Arabic and English names and the phrasings people actually use. "Revenue", "sales" and "المبيعات" can all point to the one measure finance owns.
- The query runs as that personPower BI can run a query under a specified user's effective identity, so the row-level security roles that apply to them filter the result. A regional manager asking about sales gets their region's figure, as they would in the report.
- Existing measures, not new mathsThe assistant asks the model for a measure it already has, sliced by the period or dimension in the question. It does not write new calculations, and it never goes past the model to the databases behind it.
- The answer carries its contextThe figure comes back with the period, the filter applied and the time of the model's last refresh, so nobody quotes last week's number as today's.
- Reports for anything biggerA breakdown, a chart or a question the catalogue does not cover is answered with a link to the right report, or passed to the BI team, rather than approximated.
Scope and control in Power BI
| Direction of access | Read-only queries against named semantic models. No publishing, editing, sharing or refresh triggering. |
|---|---|
| Models in scope | Only the models listed during deployment, usually the ones your BI team has already endorsed. Everything else in the tenant is out of reach. |
| Who can see what | Row-level security is applied through the asking person's identity. Power BI applies RLS to people who consume content as viewers; workspace admins, members and contributors see the whole model, which is why the identity used for queries is chosen carefully and tested. |
| Credentials | A service principal — an app registration in Microsoft Entra ID — created by your administrators, with access granted only to the workspaces in scope. You hold its secret or certificate and can rotate or delete it without us. |
| Where data goes | Figures are used to answer and are not stored: conversation content is destroyed within 24 hours, usually when the conversation ends, and no external AI service sees it. |
| If a refresh failed | The assistant gives the time of the last successful refresh and says the figure may be out of date. If Power BI cannot be reached, it says so instead of guessing. |
What this replaces
The value here is different from a system of record. The numbers are already computed and already agreed — the assistant is not calculating anything, it is retrieving a validated measure and phrasing it. That is both safer and more useful than letting an assistant compute figures itself.
It also preserves the single version of truth. When an assistant answers from the same measures the dashboards use, the answer matches the report. When it computes independently, it produces a second set of numbers, and the resulting arguments are worse than the original problem.
Row-level security matters here more than in most connections. Analytical models frequently contain the whole organisation's data with visibility controlled by rules, and the assistant must respect those rules exactly.
Who asks Power BI questions this way
Executives
Figures without navigating a report.
Sales management
Performance questions asked conversationally.
Operations
Current-state metrics.
Anyone with low dashboard adoption
Converting existing investment into answered questions.
Getting value from Power BI through conversation
Begin with one semantic model that is already trusted and widely used — typically sales or operations — and the measures people ask the BI team about most. If an analyst can list the questions they answer by opening the same report every week, that list is the first catalogue.
"How current is this?" belongs in the first release, not a later one. The most damaging thing a conversational answer can do is present a stale figure with confidence, and refresh time is already recorded for every model.
"Which report shows this?" is the third. It costs nothing to answer, it sends people to the work your analysts have already built, and it keeps the assistant honest about what it should and should not try to summarise.
What connecting Power BI involves
- Choose the modelsPick the semantic models in scope and confirm who owns each one. Only endorsed, maintained models belong in the first release.
- Write the measure catalogueFor each measure: its name in Arabic and English, what it means, which dimensions it can be sliced by, and which questions should map to it. The model's owner signs it off.
- Register the service principalYour Entra ID administrators create the app registration, a Power BI administrator allows service principals to use the APIs in tenant settings, and the principal is given access to the relevant workspaces only.
- Map identities to RLSConfirm that each person's sign-in resolves to the identity your RLS roles expect, and that no one's query runs with more rights than they have in Power BI.
- Test across rolesAsk the same questions as users in different roles and regions, and check each answer against the report they would normally open.
- Release to one teamStart with the team that asks the BI function most often, then widen the catalogue as new questions appear in the logs.
Personal data inside Power BI models
Semantic models often hold more personal data than their dashboards show: customer names in a sales table, employee IDs in an HR model, salaries behind an aggregate. Under the Saudi Personal Data Protection Law, a question answered at the level of region or month rarely raises an issue; one that drills down to a named individual can.
The catalogue is where that line is drawn. Measures are exposed at the level the questions need, dimensions that identify individuals are left out unless there is a clear business reason, and RLS covers the rest.
Power BI's own options, and where an assistant fits
Power BI already offers several ways to get a figure without building a report: the Q&A visual, Copilot on eligible capacity, email subscriptions, data alerts on dashboard tiles and the Power BI app in Microsoft Teams. For people who already work in Power BI, these are often the right answer.
A connected assistant is for the people who do not: the manager who asks a colleague, the team that works in Arabic, the question that arrives in the same conversation as others about unrelated systems. It uses the same models and the same measures, so the figures agree with the reports.
Where it goes wrong is scope. An assistant pointed at every model in the tenant, or allowed to compose its own calculations, produces a second set of numbers — which is the problem Power BI was bought to solve.
What BI teams see with Power BI connected
Common questions
No, and it should not. It retrieves the measures your analysts have already defined, so answers match your reports.
Yes. Analytical models often hold everything, with visibility governed by rules the assistant must follow exactly.
No. It answers the questions people ask instead of opening one.
No. The service principal belongs to your tenant, is created by your administrators and reaches only the workspaces you grant. Deleting it or its secret ends the connection.
Staff, in almost every case. Power BI content is normally internal, and the connection is set up for signed-in employees. Figures for customers belong in a system built to serve them.
The answer states the time of the last successful refresh and warns that the figure may be out of date, so nobody acts on it unknowingly.
Connect Power BI to Elbi
Your agreed measures, answered in a sentence for signed-in staff, in Arabic and English.