AI In Power BI (Ranked By What It Saves You)

The Power BI AI feature that gets demoed is building a report page from a prompt. It is the fourth most useful of the four things AI does in Power BI, and I have never once needed it on a Monday.

The one that pays is the commentary. You build a report once and then somebody rewrites the same variance narrative every month for four years. That is the recurring cost, and it is the one worth attacking first.

So this page runs in that order, ranked by what each one gives back rather than by how well it demos. Commentary, then DAX, then the semantic model, then the report page. Before any of it, the question of whether you can open Copilot at all, because a lot of finance teams cannot. Checked against Microsoft’s documentation on 29 August 2026.

First, can you open it at all

Copilot in Power BI runs on organizational capacity, not on your seat. You need a paid Fabric capacity of F2 or higher, or Power BI Premium at P1 or higher. Microsoft puts it plainly: a Power BI Pro or Premium Per User license alone isn’t sufficient.

Pro is what most finance teams get issued. So for a good share of the people reading this, the Copilot button is missing or greyed out and no amount of prompting fixes it. Trial capacities and free SKUs do not count either, which catches out anyone who spun up a Fabric trial to test exactly this.

Three smaller things that waste an afternoon if you do not know them. Copilot is on by default now, so if it is missing your admin turned it off rather than never turning it on. After you buy or scale capacity it can take up to 24 hours to appear. And sovereign clouds are not supported at all, with a list of regions still waiting on the standalone experience.

If none of that is available to you, skip to the section on working without capacity. Everything in it is free or already on your machine.

Commentary, the one that pays every month

Ranked first because it is the only item on this list you pay for twelve times a year. The dashboard is a one-off build. The two paragraphs explaining why gross margin moved 140 basis points are a job somebody does again every close, usually at 9pm, usually the same person.

The pattern that works is to hand the model the variance table and your own commentary from last month as the style reference, then let it draft and you edit. It gets the arithmetic right and the emphasis wrong, which is the correct division of labour, because the emphasis is the part that needed you. I walk through the whole loop in automating Power BI commentary with Claude.

What no model does is tell you that wages ran over on hours because you approved overtime in week three to cover the audit. Feed it the Power BI variance analysis and it writes the what. The why is still the job.

DAX, where it is strongest and gets demoed least

Time intelligence, running totals, the same-store comparison that has to exclude locations opened mid-year. These are the measures that eat an afternoon and they are exactly the shape a language model is good at, because the pattern is well documented and the mistake is visible the moment you check one row.

Copilot writes DAX in the DAX query view. So does Claude or ChatGPT in a browser tab, and on a Pro license that is your only route. Describe the model the way you would to a new analyst, table names, the grain of the fact table, which columns are dates, then ask for the measure and an explanation of what it does. More in DAX for finance.

The better version stops the guessing entirely. The Power BI MCP lets Claude read your semantic model directly, so it writes measures against the columns that are really there rather than the ones it assumed. That is the difference between a measure you spend ten minutes fixing and one that runs first time.

The semantic model, and the prep that decides the rest

Third on the list, and the one that decides whether the first two work at all. Microsoft is unusually blunt about it. Model owners have to prep their semantic models first, and without that prep Copilot can produce results that are generic, inaccurate or misleading. That is their word, misleading, on their own documentation page.

In finance terms: if your columns are still called Column1, three tables each have a field named Date with no relationship between them, and half your measures have no description, Copilot will hand you a confident number that ties to nothing. The prep is naming things properly, writing measure descriptions, and hiding the columns nobody should be dragging onto a canvas. It is the least interesting work on this page and it is the work that decides whether any of the rest lands. Building fast Power BI models covers the mechanics.

Copilot can now help with the model itself, flagging unclear structures and inconsistent naming and refining measures, though that experience is still in preview. Worth trying on a model you already know well, so you can judge whether the advice is any good.

Building the report page, the one from the demos

Last, and the weakest of the four. Copilot will draft report pages from a prompt. It puts plausible visuals on a canvas in a plausible arrangement, and then you spend longer fixing the hierarchy than you would have spent building it, because the thing a finance report has to do is lead with the number the room came for.

Use it as a starting canvas rather than a finished page and it earns its place, particularly on a report you would otherwise have put off for a fortnight. The craft of the thing is still yours, and finance dashboard design is where I keep the rules I hold myself to.

Where the generated version does hold up is when somebody hands you a semantic model you have never opened and asks what is in it. A drafted page plus a model summary is a faster orientation than clicking through 40 tables.

Four things that cost me time

Copilot caches its answers. Ask the same prompt against an unchanged model inside a rolling 24 hours and you get the same response back, and clearing the chat does not reset it. So when you tweak a prompt and the answer does not move, suspect the cache before you rewrite the prompt a third time. Change the wording meaningfully or refresh the underlying data.

Prompts run to 10,000 characters, which is far more room than people use. Paste the whole definition of the measure you want, the naming convention, the fiscal calendar quirk. Length is not the constraint here.

Copilot consumption bills to your capacity as background operations, and anything it triggers downstream, DAX queries, refreshes, subscription emails, bills separately through the normal path. If you share capacity with another team, your experiments show up on somebody else’s report. The Fabric Capacity Metrics app has a 30-day breakdown if you need to prove it either way.

And half of what gets announced is preview. The report Copilot pane is generally available. The standalone full-screen experience, the app-scoped one and the modelling assistant are all preview as of this week, which matters when you are the person who promised it in a steering meeting.

What to do with no capacity

Most of the value on this page survives without Copilot, because three of the four items never needed it.

Write the DAX in Claude or ChatGPT and paste it in. Write the commentary the same way, from an exported variance table. Do the model prep by hand, which was always going to be by hand. The only item you genuinely lose is the drafted report page, and that was the one ranked fourth.

Two routes worth setting up if you are in this position. The Power BI workflows I run through Claude covers the practical version, and building a dashboard with Claude Code is how far it goes when you let it write the files rather than the formulas.

The rest of the reporting shelf

The dashboards worth copying

Three builds, three audiences. The CFO dashboard is the one that gets opened in the meeting, the automated budget dashboard is the one that saves the most hours, and financial dashboards is the wider set if you are starting from nothing.

The model underneath

Power Pivot is where the relationships live and Power Query is how the data arrives, and a report that is slow or wrong is nearly always one of those two rather than the visuals. AI in Power Pivot covers the model-side version of everything above.

Going beyond the built-in visuals

When the standard chart set will not say what you need, custom Power BI visuals built with Claude is the route that does not need a developer, and Python in Power BI handles the statistical work the visuals gallery has never covered.

Forecasting and the wider tools

Power BI forecasting is free and better than people expect for a first pass. If you are comparing platforms rather than building, Tableau and Looker Studio both have their cases, and self-service analytics is the organizational version of the whole argument.

Where to go next

If the report is the end of a process rather than the start, finance automation is where the rest of that month lives. AI in Excel covers the work before the data ever reaches Power BI, AI tools for finance has the tool-by-tool verdicts, and AI agents for finance is the version where the refresh and the commentary happen without you opening anything. And if you came here because someone mentioned Copilot, setting up Power BI Copilot has the Fabric capacity it needs and what an F2 costs before you go asking for budget. Once that capacity exists, a Fabric data agent puts a chat window on the same semantic model, so the one-off questions stop landing on you.