Last updated: October 23, 2026
By ZapAI Team
TL;DR: Copilot in Power BI, the AI assistant in Microsoft’s Power Platform analytics tool, needs a paid Fabric capacity (F2 or higher) or Power BI Premium (P1 or higher). Pro or Premium Per User alone won’t do it, though a small Fabric Copilot capacity covers those users. A typical request costs about two cents in capacity units. The bigger cost is a messy semantic model, so fix the model first.
Copilot in Power BI lets users ask questions, build report pages, and summarize data in plain English. It requires F2+ Fabric or P1+ Premium capacity and bills token usage against that capacity.
The most common Copilot in Power BI support ticket we see isn’t about bad answers. It’s the button that never shows up. Someone on a Pro license opens a report, reads that Copilot is built into Power BI now, and finds nothing. Nothing at all. They aren’t doing anything wrong, because Copilot runs on capacity rather than on seats, and that single rule, buried three clicks deep in Microsoft’s documentation, explains most of the confusion we untangle in licensing calls.
Under the hood, Power BI’s Copilot features run on Microsoft Fabric capacity. This guide covers what each Copilot surface does, what you need to switch it on, what it costs in real capacity units, and the model prep that decides whether people trust the answers. The figures were checked against Microsoft Learn on October 23, 2026. A lot of pages ranking for this topic still quote 2024 numbers. We’ll point out where.
What Is Copilot in Power BI, and Which One Do You Need?
Copilot in Power BI is Microsoft’s generative AI assistant for reports and semantic models. It answers data questions, drafts report pages, writes DAX, and summarizes visuals. Five related experiences carry similar names, though, and each one is licensed and billed differently.
That naming problem is real. In licensing workshops we hear Copilot in Power BI, Microsoft 365 Copilot, and Fabric data agents used as if they were one product. They aren’t. Different licenses. Different meters. The table sorts them out.
| Experience | What it does | What you need | Status (Oct 2026) |
|---|---|---|---|
| Report Copilot pane | Summarizes a report, answers questions about it, drafts pages and narrative visuals | Report in a workspace on F2+ or P1+ capacity | Generally available |
| Standalone Power BI Copilot | Searches across reports and models you can access from the left navigation | Tenant setting plus a Fabric Copilot capacity or eligible workspace | Preview |
| Copilot in Power BI apps | Answers questions inside a published app | App content on eligible capacity | Preview |
| Microsoft 365 Copilot with Power BI data | Pulls Power BI insights into Copilot Chat, Teams, and Office apps | Microsoft 365 Copilot license, plus normal Power BI access | Rolled out worldwide from late August 2026 |
| Fabric data agent | A configurable Q&A agent over up to 5 data sources, usable in Teams or Copilot Studio | F2+ or P1+ capacity | Generally available |
Statuses come from Microsoft’s Copilot for Power BI overview (updated August 2026) and the Fabric data agent documentation. Most teams start with the report pane. It’s generally available, it’s scoped to one report, and nothing about it is exotic.
Data agents are a different animal. Microsoft notes that Fabric copilots come preconfigured, while a data agent lets you write its instructions, pick its sources, and add example queries. The trade? More setup, and answers capped at 25 rows by 25 columns.
What Do You Need to Turn On Copilot in Power BI?
You need a paid Fabric capacity of F2 or higher, or Power BI Premium P1 or higher, the Azure OpenAI tenant setting switched on, and a supported region. Trial capacities, free SKUs, and sovereign clouds don’t qualify.
Microsoft’s wording on licenses is blunt. A Power BI Pro or Premium Per User license alone isn’t sufficient, because Copilot requires organizational capacity. That’s the missing button again.
Short list. Each item still trips someone up.
- The tenant setting Users can use Copilot and other features powered by Azure OpenAI. It’s on by default, but some admins switched it off early on and never revisited it.
- Outside the US or the EU Data Boundary? Copilot stays off until an admin allows data to be processed outside your geographic region. UK, Canadian, and Australian tenants hit this one.
- Bought or scaled a capacity this morning? Microsoft says Copilot can take up to 24 hours to notice.
- Editing rights matter too. Read access gets you summaries and answers, while building a narrative visual or a new report page needs edit or build permission.
- Prompts cap at 10,000 characters on every surface. Generous, honestly.
Seen a forum answer saying you need F64? That was true once. Microsoft opened Copilot to all paid SKUs from F2 in 2025, and every current Learn page we checked says F2 or P1. The Copilot in Fabric overview also lists which regions map to US or EU Azure OpenAI datacenters, useful if your tenant sits somewhere in between.
Can Pro and PPU Teams Use Copilot Without Moving to Fabric?
Yes, through a Fabric Copilot capacity. You buy a small F SKU, designate it, and assign users. Their Copilot usage from Pro workspaces, PPU workspaces, and Power BI Desktop then bills to that capacity.
No ranking page we reviewed explains this route, and it’s the cheapest path for a typical mid-market Power BI shop. Per the Fabric Copilot capacity documentation (updated October 20, 2026), the designated capacity must be at least F2 or P1. Your reports stay where they are. Only the Copilot meter moves.
An F2 runs about $263 a month on US pay-as-you-go, at 2 capacity units times $0.18 per CU hour times 730 hours. For a team of 40 analysts on Pro, that’s a pilot budget, not a platform migration.
Four limits worth knowing before you build a plan around it. It only works in your tenant’s home region. Each user gets one Fabric Copilot capacity, and the most recently created assignment wins. Embedded workspaces aren’t supported. And it doesn’t cover Fabric AI functions, only Copilot and data agents.
Is a full move to Fabric capacity the better deal? Sometimes, once report viewers number in the hundreds. That’s a license break-even question, and we worked through it in our Fabric vs Power BI Premium comparison.
How Much Does Copilot in Power BI Cost?
Billing runs on tokens, converted to capacity units. Input costs 100 CU seconds per 1,000 tokens, cached input 10, and output 400. A typical request lands around 400 CU seconds, about two cents at US pay-as-you-go rates.
Those rates come from Microsoft’s Copilot consumption documentation, updated May 2026. Microsoft adds that 1,000 tokens is roughly 750 words and that caching happens automatically. The dollar figures in the table are our own arithmetic at $0.18 per CU hour, the Central US pay-as-you-go rate in the Azure retail prices API.
| Scenario | Capacity used | Approximate cost |
|---|---|---|
| Microsoft’s example request, 2,000 input and 500 output tokens | 400 CU seconds | $0.02 |
| Same request with 1,500 input tokens served from cache | 265 CU seconds | $0.013 |
| Report summary, 4,000 input and 1,500 output tokens (our assumption) | 1,000 CU seconds | $0.05 |
| Daily ceiling of an F2 if Copilot ran nothing but the example request | About 432 requests | $8.64 a day of capacity |
| Daily ceiling of an F64, per Microsoft | Over 13,824 requests | Part of the F64 fee |
Output tokens cost four times input. That one ratio explains which features are expensive. A quick question with a one-line answer is cheap. A narrative visual or an emailed subscription summary that writes three paragraphs for 300 recipients isn’t, and Microsoft itself recommends piloting with a small group before switching on summaries for hundreds of users.
Watch for stale numbers. When Copilot billing launched in March 2024, Microsoft charged 400 CU seconds per 1,000 input tokens and 1,200 per 1,000 output. Many guides, and plenty of AI chatbot answers, still quote those. Current rates are a quarter of the old input price and a third of the old output price. Budgets built on the 2024 numbers come out roughly three times too high. Odd problem. Overestimating cost sounds harmless, but we’ve seen a perfectly reasonable Copilot pilot get shelved in a steering meeting because someone pasted the launch-era rates into a spreadsheet and multiplied by every employee in the company.
Microsoft notes that consumption rates can change at any time. Check the page before you finalize a budget.
Will Copilot Slow Down Your Refreshes?
It can, on a small capacity. Copilot runs as a background operation smoothed over 24 hours, so it rarely spikes alone, but it shares the same capacity units as refreshes, dataflows, and report queries.
Smoothing is the reason a heavy Copilot day doesn’t throttle anyone at 10 a.m. It’s also the reason a heavy Copilot week can. Fabric starts throttling when a capacity has used up its next 10 minutes of capacity units, and a pile of smoothed background usage counts toward that. On an F2 or F4 already busy with overnight refreshes, a curious sales team can tip it over.
The Copilot meter isn’t the whole bill either. DAX queries that Copilot generates, and anything it triggers downstream, bill separately as normal Power BI operations. Watch the Copilot in Fabric operation in the Fabric Capacity Metrics app, which keeps 30 days of history.
Our rollout pattern is dull on purpose. Start with one security group, usually finance or sales ops, on a Fabric Copilot capacity separate from production refreshes. Watch two weeks of metrics. Then widen. Isolation costs a little more. Throttled month-end refreshes cost a lot more. Ask any controller.
How Do You Make a Semantic Model Copilot-Ready?
Clean names, descriptions, and relationships first. Then use Microsoft’s three Prep data for AI tools (AI data schema, verified answers, and AI instructions), and mark the model Approved for Copilot so users don’t see a quality warning.
Metadata is all it reads. Not your intentions. Microsoft’s Copilot privacy and security page says it uses table names, column names, and model properties to build DAX, and the Q&A setting has to be on for it to answer data questions at all. A column called Amt_2 sitting next to Amount is a coin flip for Copilot. It’s a coin flip for new analysts too.
In the model reviews we run before a Copilot rollout, the problems are almost always the same handful.
- Duplicate or cryptic field names, like three different date columns with no hint which drives the calendar.
- Technical columns left visible. Keys, sort helpers, and staging fields should be hidden.
- Measures without descriptions. Copilot can draft them for you, but someone who knows the business has to check each one.
- Missing synonyms. Your team says bookings, and the model says SalesOrderValue.
- Relationships that technically work but give wrong totals when sliced by a second dimension.
Fix those, then move on to the three tools in Microsoft’s Prepare your data for AI guide. AI data schema limits which tables and fields Copilot considers. Verified answers pin a specific visual to a common question, so the answer to what was revenue last quarter is the one finance signed off. AI instructions give business context in plain language, like fiscal year starts in July, or active customer means an order in the last 12 months.
Order matters. Instructions can’t rescue a confusing model, and we’ve watched teams write two pages of AI instructions to paper over a relationship bug that took ten minutes to fix. Model first. Instructions second.
Last step, mark the model Approved for Copilot (the setting used to be called prepped for AI). It can take up to an hour to apply, or a full day for models feeding dozens of reports. Without it, the standalone experience shows users a low-quality warning, which is a fast way to lose their trust before the first answer.
A Readiness Checklist Before You Switch It On
Run through this list before opening Copilot to anyone outside the BI team. Every item maps to a setting or a document above.
- Tenant setting for Azure OpenAI features enabled, scoped to a pilot security group.
- Cross-geo processing reviewed if your tenant is outside the US or EU Data Boundary. Get legal or compliance sign-off rather than just flipping it.
- F2+ or P1+ capacity in place, or a Fabric Copilot capacity assigned to Pro and PPU users. Then wait 24 hours.
- Q&A enabled on each target semantic model.
- Names, descriptions, hidden columns, and synonyms cleaned up.
- AI data schema trimmed, verified answers for your top 10 to 20 questions, short AI instructions.
- Approved for Copilot set on the models you want people using.
- Row-level security roles tested with Copilot prompts. Don’t assume. Check that a regional manager can’t coax another region’s numbers out of a summary.
- A Capacity Metrics app review booked for two weeks after launch.
Looks long. It isn’t. On a well-kept model most of this is a day of work, and on a model that five people have patched over six years without a naming standard it’s closer to a week, most of which goes to arguing about which of three revenue measures is the real one.
Who Should Wait on Copilot in Power BI?
Teams with untidy models, no spare capacity, or a mostly view-only audience should wait. Teams with clean models, analysts drowning in ad hoc requests, and room for an F2 should start a pilot now.
Bias disclosed. We get paid to set this up, so here’s the case against it. If your reports are mostly static PDFs exported for executives, Copilot adds little. If your semantic models have no naming standard, Copilot will repeat that mess back to users with confidence. Loudly. And if your capacity already runs hot at month-end, adding Copilot to it before splitting the workloads is asking for a throttling incident.
Where does it pay off? Finance and sales ops teams with a well-built model and a queue of what-if questions. Analysts use Copilot to draft DAX and report pages faster, and business users get quick summaries without waiting a week for a new visual. That second group is where the time savings show up. Measure them in your pilot rather than trusting anyone’s headline percentage, ours included.
Still on the fence about Power BI’s place in the wider Microsoft data stack? Our explainer on how Power BI supports better decisions covers the basics, and the Microsoft Fabric overview explains the capacity model Copilot runs on.
Wrapping Up
Cheaper than most guides suggest, and pickier than most teams expect. That’s Copilot in Power BI in 2026. F2 capacity is the floor, not F64. A Fabric Copilot capacity brings Pro and PPU users along without a migration, and a typical request costs around two cents. Model quality decides whether any of it is worth the money.
ZapAI is a Microsoft Dynamics 365, Power Platform, Fabric, and AI consulting firm based in Irving, Texas, working with US businesses. Our Power BI consulting team runs Copilot readiness reviews that cover the semantic model, capacity sizing, tenant settings, and a two-week pilot plan, so you know what Copilot will cost and whether people will trust it before it reaches the whole company.
What Power BI Teams Ask About Copilot
Is Copilot in Power BI free?
No separate license fee, but it isn’t free. Copilot consumes capacity units on an F2+ or P1+ capacity you already pay for, so heavy use either crowds out other workloads or pushes you to a bigger SKU.
Why don’t I see Copilot with my Power BI Pro license?
Pro and PPU licenses alone don’t include it. The report has to live on F2+ or P1+ capacity, or your admin has to assign you to a Fabric Copilot capacity. If that’s already done, check the tenant setting and give a new capacity 24 hours.
Do we really need F64 for Copilot?
Not anymore. Microsoft dropped the F64 requirement in 2025, and current documentation says F2 or P1. F64 still matters for a different reason, since it’s the size where report viewers stop needing their own Pro licenses.
How many requests can a small capacity handle?
About 432 a day on an F2, by our math, if Copilot ran Microsoft’s 400 CU-second example request and nothing else. Real capacities also run refreshes and queries, so plan on a fraction of that. Microsoft puts an F64 at over 13,824 such requests a day.
Copilot in Power BI or Microsoft 365 Copilot, which one do people actually use?
Usually both, for different jobs. Analysts and report builders live in the Power BI Copilot pane. Managers who rarely open Power BI ask Microsoft 365 Copilot in Teams or Outlook, which can pull Power BI insights if they hold a Microsoft 365 Copilot license and normal access to the content.
Does Copilot send our data to OpenAI?
Prompts, metadata, and conversation context go to Azure OpenAI running in Microsoft’s datacenters, not to OpenAI’s public service. Microsoft says your data stays in your tenant boundary and isn’t used to train foundation models. Table contents are only sent when a prompt needs them.
Can Copilot write DAX for us?
Yes, in Desktop and in the DAX query view, using your model’s table and column names. Treat it like a fast junior analyst. Review every measure before it goes into a production model.