Ask a BI question and check what the answer measured

Example result.

North-region net revenue was £1,768,000 for 1 July–30 September 2026. Measure: Net Revenue (governed measure, not calculated on the fly). Filters: Region Name is “North”. Dates: Order Date 01/07/2026–30/09/2026. Reconciles to the board-pack visual. First answer used Revenue (Gross); corrected.

A measure is the calculation behind a figure, such as net revenue. The semantic model holds the data and calculations the BI tool works from. A fluent answer doesn’t tell you which ones it picked; the panel does.

1. Ask a question that resolves to one number

“How are we doing in the North?” lets the assistant pick a measure, a period and a grouping, all silently. A checkable question pins down four things: the measure, the slice, the period and the comparison.

What was [exact measure name] for [region or group] from [start date] to [end date], and how does that compare with the same period last year?
State the measure, date field, filters and source you used.

[exact measure name] is the name in your report’s field list, such as “Net Revenue”. [region or group] is the slice. Use real dates for [start date] and [end date], not “last quarter”.

Why this works: relative dates depend on the model’s date table and fiscal calendar, which may not match the calendar in your head. Real dates and an exact measure name remove two silent choices.

What you should see: one figure, one comparison and a sentence naming what it used. That sentence is a claim; step 2 checks it.

Microsoft lists anomaly detection, forecasting and key influencers as unsupported, so “why do sales drop every July?” is the wrong first question.

2. Open the panel that shows the working

This is the step people skip, and it’s the one that makes the number usable.

• Power BI: under the answer, expand How Copilot arrived at this to see the fields, measures and filters Copilot selected. Select the link beside a Data used item for detail. If you’ve turned on Fabric IQ in the Copilot pane while viewing a report in the service, expand the reasoning steps instead; they show the generated DAX query and its results.
• Tableau Pulse: open the metric’s insights exploration page to see its current value, the change against the prior period and the filters applied to the metric definition. If your site has Tableau Agent in Pulse, select the citation link at the end of each insight and Metric Sources to see which metrics were used.

Why this works: the panel is the product’s record of what it ran, not the model’s description of what it meant to run.

What you should see: a measure name, a list of filters and two dates. If the panel shows none of these, you can’t verify the answer, so don’t quote it.

3. Read back four things, then reconcile

Work down the panel and check each one against your question.

• The measure. Is it the one you asked for, or one with a similar name? Models collect Revenue, Revenue (Gross), Net Revenue and Revenue LY, and the assistant chooses between them by name and description.
• Whether it was calculated on the fly. Power BI Copilot can write DAX for calculations that aren’t in the model, such as growth or ratios. A governed measure and a calculation made up for your question five seconds ago deserve different levels of trust.
• The filters. Every one, including ones you didn’t ask for. Copilot data questions don’t apply the slicers or filters currently on the report page, so put the scope in the question.
• The date range. The start and end dates themselves, not a label.

Then compare the figure with a trusted report visual for the same measure, filters, dates and refresh time. Compare underlying values, not rounded labels.

Why this works: each of the four is a separate way a correct-looking number can answer a different question. Matching totals alone can hide two different scopes.

What you should see: four ticks and a figure that matches the reference, or a specific line that doesn’t.

Check the result

• The measure in the panel is the measure you named, letter for letter.
• Every filter in the panel is one you meant, and none you asked for is missing.
• The dates are the dates you asked for, on the date field you expected.
• The figure matches a trusted visual with the same scope and refresh time.

The usual failure: the assistant answers with a similarly named measure, such as gross revenue when you asked for net. The sentence still says “net revenue”, so only the panel gives it away. Re-ask, naming the thing it got wrong:

Use the [correct measure] measure instead. Keep the same filters and date range, and tell me which measure, filters and dates you used.

[correct measure] is the exact name from your field list.

If Power BI returns the identical answer, it may be serving it from cache: the same prompt on an unchanged model within 24 hours gets the same response, and clearing the chat doesn’t change that. Change the wording.

Fictional example. A homeware wholesaler’s finance manager asks the step 1 question in Power BI for net revenue, North region, 1 July to 30 September 2026. The answer:

Net revenue for the North region in Q3 2026 was £1,842,000,
up 6.1% on the same quarter last year.

The board pack says £1,768,000, about four per cent lower. She expands How Copilot arrived at this:

Measure used: Revenue (Gross)
Filters: Region Name is "North"; Order Date between 01/07/2026 and 30/09/2026
Comparison: Order Date between 01/07/2025 and 30/09/2025
Rows returned: 1

With Fabric IQ on, the reasoning step would show a query along these lines:

EVALUATE
ROW(
  "Revenue", CALCULATE(
    [Revenue (Gross)],
    'Region'[Region Name] = "North",
    DATESBETWEEN('Date'[Date], DATE(2026,7,1), DATE(2026,9,30))
  )
)

She uses the repair prompt with “Net Revenue”, gets £1,768,000, and it matches the board-pack visual. £1,842,000 − £74,000 of credits and returns = £1,768,000.

• The first line gave it away. She asked for net and got Revenue (Gross), the measure whose name best matched. The answer’s own sentence still said “net”.
• You don’t need to write DAX to read it. The measure in square brackets, the literal "North" and the two dates are legible to anyone. If the data says “Northern” and people say “North”, a filter on the wrong string returns a number rather than an error.
• Filters and dates were right, so the gap had one cause. Checking all four items is what lets you say that with confidence.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Sources

• Microsoft Learn: Copilot for Power BI overview
• Microsoft Learn: Ask data questions with Copilot and Fabric IQ
• Microsoft Learn: Prepare your data for AI
• Tableau Help: Explore metrics in Tableau Pulse
• Tableau Help: Set up Tableau Pulse
• Tableau Help: Ask questions with Tableau Agent in Pulse