Copilot in Power BI lets people ask questions of their data in plain language, draft report pages and summarize what changed. Switching it on takes minutes. Whether people keep using it depends on the first answers they get, and those answers depend on the semantic model underneath: the tables, relationships and calculations that every report reads.
A model built up over years for a few report authors is rarely ready for that. It is full of technical columns, cryptic names and near-duplicate calculations that its authors know to ignore. Copilot does not know. The preparation below is what we put in place before switching Copilot on for a leadership team. None of it is exotic, and all of it also makes the model easier for people to use.
Start with one model people already trust
Do not try to prepare everything at once. Pick one model that sits behind a report leaders already use, with a business owner willing to sign off its definitions; sales performance or the monthly financial results are typical candidates. A trusted model gives Copilot a fair start and gives you a reference point: if an answer disagrees with the report everyone knows, you will hear about it quickly, and you will know which side is right.
Resist the temptation to start with the largest model you have. A model that tries to serve every department carries every department’s ambiguities, and Copilot will find them all in the first week.
Name and describe everything that stays visible
Copilot works from what the model shows it. Every visible table, column and measure needs a name a business person would recognize and a short description of what it means, what it includes and what it leaves out. Write for someone who has never seen the model, because that is exactly Copilot’s position. The analytics team can draft the descriptions, but the business owner should approve those for the measures leaders will ask about.
Then shrink what is visible. Hide anything nobody should ask about:
- Keys, codes and identifiers used only to join tables.
- Columns that exist only to sort, format or filter other fields.
- Intermediate calculations that feed other measures but mean nothing on their own.
- Old or duplicate fields kept alive for a report that no longer exists.
Every hidden field is one less wrong turn. Teams are often surprised how small the visible model becomes, and how much easier it is for people to find their way around it as well.
Teach it the words people use
Leaders do not ask for the sum of a net amount column by territory description. They ask how sales did in the West. Add synonyms for the terms people actually use, including abbreviations, product family nicknames and region names as people say them. Make the calendar explicit: tell the model which table is the calendar, name fiscal periods the way the business does, and state which date each measure follows. Check the list with a few future users; they will name terms the analytics team never uses.
Set formats and categories while you are at it: currencies, percentages, geographic fields. They decide whether an answer arrives with its currency and unit or as a bare number that nobody can interpret.
Certify a small set of measures
Most leadership questions come back to a short list: revenue against plan, margin, volume, cash and service. Each of those deserves exactly one measure, defined with the function that owns it, described in plain words and visible under a name nobody can misread. Give each one the same name in the model, in reports and in the monthly pack. Retire or hide the near-duplicates. If several measures could answer “what was margin last quarter?”, Copilot will eventually pick the wrong one, and it will do so with confidence.
Once the definitions are agreed, endorse the model as certified so that people can see it is the reference version. Certification is a promise; keep it by putting every change to a certified measure through review.
Write the test before you switch it on
Ask the business owner and a few future users for the questions they really ask, in their own words, and record the correct answer to each from the certified report. Include the awkward ones: questions with ambiguous dates, comparisons with plan, and questions the model cannot answer, where the right response is to say so. Keep the set small enough to run often and broad enough to cover every certified measure and the main ways leaders slice it.
Run the full set before launch and after every change to the model. Agree a pass mark with the business owner, and do not open Copilot to leaders until the model clears it. When an answer is wrong, fix the cause in the model, whether a name, a description, a synonym or a duplicate measure, rather than coaching users to phrase their questions differently.
Roll out to one team first
Launch with one team that uses the model every week and is willing to report problems. Give them a short briefing: what Copilot does well, how to check where an answer came from, and how to flag one that looks wrong. Review flagged questions every week and feed the fixes back into the model and the test set. Agree in advance what success looks like for the pilot: how often the team uses Copilot, how many flagged answers turn out to be wrong, and how quickly fixes reach the model.
Watch what people actually ask, not only whether the answers are right. Questions the model cannot answer point to the next measure worth adding; fields that nobody ever asks about are candidates to hide.
Widen access only when the first team’s questions are answered reliably and the fixes have slowed down. Each new team brings its own vocabulary, so expect a short round of synonyms and descriptions with every expansion, and add the new team’s questions to the test set before they start.
The bottom line
Copilot is only as good as the model it reads. Focused work on names, descriptions, hidden fields, synonyms and certified measures, checked against real questions and proven with one team, is the difference between an assistant leaders rely on and one they try once and abandon. The same work makes the model better for every report built on it, so none of it is wasted, whatever the pace of the rollout.