Power BI DAX with AI – Write DAX with Copilot

Feature updates worth knowing

A lot of Copilot coverage still describes the 2024 launch state. Here’s what’s actually different now.

How to access DAX Copilot

There are three distinct surfaces. They’re not interchangeable — each has a different scope.

1. DAX query view (the main one for formula work)

In Power BI Desktop, switch to the DAX query view using the icon on the left rail. Press Ctrl + I or click the Copilot button to open inline Copilot. This is where you write and refine measures conversationally. It operates on the current semantic model only — Copilot sees your model’s metadata but nothing outside it.

2. Model view — measure descriptions

Select any measure in the Data pane, find the Description field in Model view properties, and click Create with Copilot. This generates a documentation string rather than a new formula. Run it after every formula edit to keep documentation accurate.

3. Web modeling in the Power BI service (June 2026, Preview)

Open a semantic model in the Power BI service and use the Copilot assistant to analyze and modify the model without opening Desktop. Currently covers schema improvements, relationship creation, and DAX measure generation. The desktop model and the service model stay in sync.

Prompts that reliably work — with their output

These are the prompt patterns that produce clean, useful DAX. The model still needs correct table and column names to match your semantic model; if Copilot guesses wrong, follow up with a correction prompt rather than starting over.

VAR Curr = [Total Sales]
VAR Prior = CALCULATE([Total Sales],
    DATEADD('Date'[Date],-1,YEAR))
RETURN
    IF(ISBLANK(Prior)||Prior=0,
        BLANK(),
        DIVIDE(Curr-Prior,Prior))

Top 5 Customers = 
RANKX(
    ALL(Customer[CustomerName]),
    [Total Revenue],
    ,DESC,Dense
) <= 5

3M Avg Sales =
AVERAGEX(
    DATESINPERIOD(
        'Date'[Date],
        LASTDATE('Date'[Date]),
        -3, MONTH
    ),
    [Total Sales]
)

-- Paste your measure in the query
-- window, then send this prompt.
-- Copilot rewrites it using VAR/RETURN
-- pattern, removes redundant CALCULATE
-- wrappers, and adds inline comments.
// Review the diff before accepting —
// Copilot may rename variables.

-- Copilot returns a natural language
-- explanation with one concrete DAX
-- example per function showing how
-- the filter context changes.
-- No generated measure to add —
-- this is a learning interaction.

% of Total Sales =
DIVIDE(
    [Total Sales],
    CALCULATE(
        [Total Sales],
        ALLSELECTED(Product[ProductName])
    )
)

What Copilot won’t handle, and where it degrades

Copilot works from model metadata alone. It doesn’t scan your actual data rows (except a min/max sample in Import mode). If your model is too large, or your metadata is thin, results degrade before they fail.

Source: Power BI Copilot community documentation and Microsoft Learn.

Model quality determines Copilot quality

Copilot uses your model’s metadata as its entire context — there is no external knowledge base it falls back on for your data. That means the naming conventions and descriptions you put into the model directly control how accurate the output is.

• Name columns and tables for humans, not for databases. CustomerName and OrderDate produce better results than cust_nm or ord_dt.
• Write descriptions for every measure. Only the first 200 characters are used. Start with what the measure does, not what it’s called. Use the Copilot description-generation feature to build this library faster.
• Mark trusted measures in workspace settings so Copilot knows which implementations to prefer when multiple measures calculate similar things.
• Keep relationships clean and explicit. Ambiguous or missing relationships produce ambiguous DAX. Copilot in web modeling can now flag these and suggest fixes from the service without opening Desktop.
• Use star schemas. Complex snowflakes and many-to-many relationships produce more errors. Copilot behaves predictably on well-modeled data.
• Always review before publishing. Copilot generates from metadata, not from running your data. A formula can be syntactically correct and logically wrong for your specific context.

FAQ – Power BI DAX with AI

Documentation & learning