Enhancing Natural Language Q&A
When I first heard about Copilot and its ability to interpret natural language queries in Power BI, my first reaction was equal parts excitement and dread. Excitement because users could finally just ask questions to answer some of their more ad hoc questions. Dread because, let's face it, most larger organisations end up with different departments speaking their own private dialects which is often confusing enough for humans, let alone AI.
I remember one subscription-based business I was working with was, naturally, heavily reliant on metrics like ARR. These terms were second nature to the finance and sales teams. Yet, whenever I talked to the marketing department, their reports always referred to the same metrics simply as "Revenue." Cue endless confusion and analysts stuck playing translator between two departments, who were, on the face of it, speaking the same language.
Copilot is remarkably smart, but it's not psychic. It relies on the clarity and consistency of your semantic model. Without clear naming, comprehensive synonyms, and thoughtfully written metadata, Copilot might interpret the same question asked by two different departments as entirely separate queries. Imagine the confusion when "Show me Revenue for last month" from marketing yields completely different results than "What's our ARR for last month?" from finance, simply because the model didn’t recognise these as synonyms.
Image thoughtfully produced on Canva© for Ascendant Analytics
You can see why one essential tool in Power BI is using synonyms. This allows you to explicitly tell Copilot that “Revenue,” "ARR," and "Annual Recurring Revenue" all mean the same thing, in the context of your model. Obviously, Revenue and ARR are different metrics and the business I worked with reported on them using different systems (ERP vs Power BI), so it was OK using a synonym in this way to avoid confusion in my model. Setting up synonyms isn't glamorous work, but it is absolutely crucial.
To make this even easier, rather than guessing which terms users might employ, you can let Copilot provide a first draft. From there, you simply review, refine, and approve these suggestions. It’s like having a thesaurus tailored exactly to your business, no tedious brainstorming required.
Metadata descriptions are equally important, acting as a guide for your model. While I admit this often feels like paperwork, we need to remember that context is key. Without good descriptions, Copilot might misunderstand the intent behind certain columns or measures, causing misleading visualisations and Q&A misfires. It reminds me of Socrates’ comparison of himself to a midwife: he said he did not give birth to ideas himself, instead he helped others deliver their own insights. Metadata works the same way, it doesn’t generate the insights on its own, but it ensures the model can deliver them clearly and correctly.
Initially I wondered how this overlaps with Copilot instructions, but I’ve realised synonyms and metadata set the foundational layer for Copilot to interpret basic terms and phrases. Instructions, on the other hand, expand on this baseline. so use them for explaining complicated business logic or handling tricky scenarios that synonyms alone can't resolve. Think of synonyms and metadata as teaching Copilot basic vocabulary, whereas instructions help it navigate complex company-specific grammar.
It’s also worth acknowledging the reality that language will always have various dialects. Even with the best semantic models and governance, terminology shifts are inevitable. Leadership changes often bring new jargon as eager analysts adopt the new boss's way of speaking overnight. Sometimes you can gently correct this but sometimes you’ll need to equip Copilot to recognise and handle the variations. Good synonym practices can absorb some of these variations, ensuring the business stays aligned without constant semantic policing.
By proactively managing naming, synonyms, and metadata, you’re enhancing Copilot’s performance and reducing confusion across departments. Crucially, you’re helping users trust the data they get back. When finance and marketing can ask Copilot the same question in their respective languages and receive the same clear answer, you know you've got your semantic model right.
In the end, clarity always wins. It's true for people, and it's especially true for AI. Clear, consistent, and well-documented language helps Copilot become an ally rather than another source of confusion. So, do the paperwork, set your synonyms, write those metadata descriptions, and watch as your Q&A experience becomes a whole lot less stressful.