Business Intelligence has come a long way. Not long ago, it felt like an exclusive club where only enterprise BI teams had access. Business users had to submit a request, wait weeks, and eventually receive a static report, often outdated by the time it arrived.
Then came the era of self-service BI. Tools like Power BI changed the game, putting analytics into the hands of business users. You didn’t have to code to explore your data you could drag, drop, and discover insights on your own. That was huge.
But the story doesn’t stop there. We’re now moving into the era of conversational BI. Instead of building everything manually, you can ask your data a question in plain English and Copilot in Power BI does the heavy lifting. Want to see last quarter’s sales trend or get a quick summary of customer churn? Just type it in. Copilot will generate the visuals, the DAX, or even a narrative explanation.
Here’s the catch: Copilot is only as good as the data and AI instructions it receives. If your semantic model is messy, inconsistent, or unclear, the answers Copilot generates might miss the mark. That’s why preparing your data model and more importantly, giving Copilot clear AI Instruction is a game changer.
Think of AI Instructions as the context Copilot needs. They define what terms mean, which metrics matter, the business rules to follow, and how to avoid confusion. Without them, Copilot is guessing. With them, Copilot sounds like it has worked in your company for years.
In this blog, we’ll break down how to craft great AI Instructions in Power BI, explore best practices, and show why a little preparation upfront can unlock much better insights from Copilot.
Preparing Data for AI in Power BI: The Role of AI Instructions
When Microsoft introduced Prep Data for AI in Power BI, the idea was simple: if you want Copilot to give you great answers, you have to set the stage properly. Think of it like hosting a guest you wouldn’t invite someone over without first tidying up the house, right? Same thing here: prep your semantic model so Copilot feels “at home” with your data.
You will find this Prep Data for AI Option in Power BI Desktop under the Home tab.
Figure 1 Prep Data for AI in Power BI Desktop
There are three options to prepare your data. You can simplify your model by unchecking unnecessary fields from your model. Those might be the columns like keys, offset or any columns that are not directly related to business context but necessary to build the data model.
Then there is verified answer, you can set a visual from your report as a verified answer for a particular question. When users ask the question, Copilot answers the visual.
Finally, we have AI instructions, where you can write all your instructions to guide Copilot to answer properly. All the instructions will be grounded in Fabric for the current semantic model.
When to Use AI Instructions in Power BI Copilot
If you look at a typical Power BI workflow, it usually goes like this:
Extract Data → Prepare Model → Create Report.
That flow works fine when you’re only building traditional dashboards. But once Copilot enters the picture, there’s another step to think about: Prep for AI.
Here’s how it plays out:
- Extract Data – You pull the data from your source systems. At this stage, AI isn’t really in the picture yet.
- Prepare Model – You shape tables, create relationships, build measures. This is where most of the heavy lifting happens.
- Create Report – You design visuals for your end users.
Up until now, this has been the “classic BI cycle.”
But for AI to give you reliable answers, you need two extra layers:
- Feedback & Validation – Once the model is built, get business users or subject matter experts to validate it. If the metric definitions are unclear or the schema is too complex, Copilot will only amplify that confusion.
- Prep for AI – This is where AI Instructions live. After your model is validated, you add context, business rules, synonyms, and guardrails. In other words, you teach Copilot how to “talk business” in your organization.
The takeaway: Don’t leave AI Instructions as an afterthought. Think about them after your model is validated but before you expect Copilot to shine. This is the stage where your data becomes truly AI-ready.
Best Practices for Writing AI Instructions in Power BI Copilot
Writing AI Instructions isn’t about telling Copilot what to say word for word; it’s about shaping the way it thinks and responds. Clear instructions help Copilot understand your data, speak the language of your business, and stay consistent in every answer.
In this section, we’ll walk through some essential building blocks of great AI Instructions.
Define Domain Scope for Copilot
Clearly define the scope of your semantic model so Copilot knows what domain it represents. Write a short summary about the data and the business area it support Finance, Operations, Marketing, or others. This helps Copilot understand what types of questions it can reliably answer.
Example: “This model covers financial reporting. Expect answers around revenue, expenses, profit, and budget vs. actuals. It does not cover customer demographics or operational KPIs.”
Know Your Audience: Business Users First
Always specify the intended audience for the AI instructions. If your audience is business users, focus on simplifying definitions, explaining metrics in plain language, and avoiding technical jargon. For example, Instead of writing:
“Gross Margin % = DAX expression: DIVIDE([GrossMargin], [Revenue])”
Write it in plain business terms:
“Gross Margin % tells us how much profit we keep from every dollar of revenue. It’s calculated as Gross Margin divided by Revenue, and it excludes returns and discounts.”
See the difference? The second version is something any business user can understand right away. It feels approachable, clear, and aligned with the way they naturally ask questions, without drowning them in technical syntax.
When the audience is business users, let Copilot explain metrics like a colleague at the watercooler, not like a developer reading DAX.
Keep AI Instructions Aligned with Your Brand Voice
Your AI instructions should align with the overall brand voice you’ve chosen. If your tone of voice is conversational, keep your examples and clarifications friendly and clear. If your brand leans professional and precise, avoid casual phrasing or humor. The goal is consistency so Copilot always “sounds” like your organization, no matter who is asking questions.
Response Style: Start with a Summary, Then Add Detail
When guiding Copilot, don’t just tell it what to answer; tell it how to answer. Business users often want a quick takeaway first, but also need enough context to trust the number.
That’s why a “Brief First, Details Next” style works so well. Copilot should give the topline result immediately, then follow up with clarifications, assumptions, or breakdowns.
Example:
Brief: “Total revenue for Q3 was $12.5M.”
Detail: “This figure includes all sales regions and excludes product returns. North America contributed 45% of the total.”
With this approach, business users get instant answers to keep decisions moving—while analysts and finance teams can dive into the details to double-check accuracy.
Handle Ambiguity by Guiding Copilot to Ask Questions
Copilot should never guess when a question is unclear. Instead, guide it to ask clarifying questions.
Example: If a user asks, “Show me revenue,” Copilot might respond: “Do you want gross revenue or net revenue after returns?”
This avoids incorrect answers and builds trust with users by showing that the AI understands nuance.
Define Business-Specific Terminology and Acronyms
Every business has its own “dictionary” of terms and acronyms. Document them clearly so Copilot knows what they mean and how to use them.
Example:
GM% (Gross Margin Percentage): Profitability metric, calculated as (Revenue – Cost of Goods Sold) ÷ Revenue.
AR (Accounts Receivable): The total money owed to the company by customers for delivered products or services. For example, “AR Aging” refers to tracking outstanding invoices by due date buckets (30, 60, 90 days, etc.).
Ops (Operations): Refers to supply chain, logistics, and production metrics in this model.
Defining these terms ensures Copilot speaks the same language as your people.
AI Instructions in Power BI Are an Ongoing Cycle
The Semantic Model Developer defines the model, while the Decision Maker tests it through real questions.
Feedback flows back from the consumer, highlighting gaps or confusion.
The developer then tunes Copilot with better instructions, making the next answer smarter and closer to business reality.
Conclusion
The key takeaway is simple: Copilot is only as good as the instructions it receives. With clear AI Instructions, you move beyond generic responses and unlock insights tailored to your business. As Power BI evolves, organizations that invest in preparing their data and defining instructions will be the ones getting the most reliable, business-ready answers from Copilot.
