Power BI Verified Answers in Copilot - What Works and What Falls Short

A finance director types "sales by region last quarter" into Copilot and gets back a clean bar chart. It looks right. The problem is that the measure Copilot picked includes returns, and the board pack your CFO prepared uses net sales. Two people, one question, two different numbers, and a meeting that starts badly.

This is the exact problem verified answers are meant to fix. Australian organisations are rolling Copilot out to people who have never opened a semantic model, and those people will trust whatever chart appears on screen. If the chart is wrong, or just subtly different from the official number, you have an adoption problem that is hard to walk back.

We have been setting up verified answers on a handful of client models over the past few months. Some of it is genuinely good. Some of it is still rough. Here is what we have learned.

What a verified answer actually is

A verified answer is a human-approved visual that Copilot returns when a user's prompt matches a trigger phrase you have set. Each one has one or more trigger phrases, a visual, and optionally some filters that users can adjust in natural language.

The important bit is where it lives. Verified answers are stored on the semantic model, not on the report. That means one verified answer set up once is available in every report built on that model. We find this is the main reason to bother. You stop having five analysts each building a slightly different version of "monthly revenue trend", and the answer in Copilot matches the answer in the official report.

When a user types a prompt, Copilot first checks for an exact or semantically similar match to a trigger phrase. If it finds one, it returns your curated visual instead of generating a fresh answer. Copilot then shows a verified checkmark, the phrase it matched, a short text summary, and an explanation of the filters it applied. That transparency matters. People are far more willing to trust a chart when they can see why it appeared.

Setting one up

The workflow is simple enough. In Power BI Desktop, you select a visual, open the three-dot menu on the visual header, and choose Set up a verified answer. In the service, you need to be in edit mode on a Copilot-enabled workspace where you have authoring permission on the semantic model. The same menu option is there.

Then you add trigger phrases. You can type your own or pick from Copilot's suggestions. Our rule of thumb is five to seven phrases per answer. Write them the way your managers actually talk, not the way the data dictionary reads. "How are we tracking on sales this quarter" is a better trigger than "Quarterly revenue KPI actuals".

Filters are where it gets interesting. Verified answers support up to three filters that users can adjust by prompt. You can filter by basic categorical values, advanced categories like starts-with or contains, and date ranges with fixed start and end points. Relative dates like "YTD" or "last three months" are not supported, which catches out almost everyone the first time. If you need relative dates, you will need to build several verified answers with fixed ranges, or accept that users will type a fixed range themselves.

There is also a trap around existing filters. A filter only appears under "Available to users" if it exists in the report and has no value applied. If your region filter is set to Northwest on the visual, it becomes locked into the verified answer and users cannot override it. Slicers are not carried across at all. We now clear every filter on the source visual before we start, which saves a fair bit of confusion.

The limits that will bite you

The documentation lists plenty of constraints. These are the ones that have actually caused us problems.

Hidden fields break things. If any field used in the visual, including filter fields, is hidden in the model or in schema selection, the verified answer will not reliably return. This is easy to miss because hiding fields is a common tidy-up step for end users. Check your schema selection before you blame Copilot.

Report measures are not supported. Verified answers do not work with visuals that use report-level measures. The measure has to live in the semantic model. If your team has been putting calculations in the report layer to avoid touching the model, you will need to move them before this works.

Several visual types are off the list. Custom visuals, decomposition trees, key influencers, and forecasting visuals built with R or Python are all unsupported. Textbox visuals are also excluded. So the chart you were hoping to surface as a verified answer might simply not be available.

Row-level security is not a security feature here. The documentation is clear that RLS and OLS are not fully supported for verified answers. In most cases the system will not match or return a verified answer that contains protected data, but there are edge cases, and the file format in Git can expose content. We tell clients to treat verified answers as a consistency tool, not a control. If sensitive data sits behind a verified answer, the access control has to be somewhere else.

Edits are clumsy. Editing a verified answer's trigger phrases and filters happens in the Prep data for AI dialog, not by editing the source visual. Changing the visual in the report does nothing to the verified answer. If you need a different chart, you delete the old answer and create a new one. This will improve, but right now it is a process problem, and it needs a named owner.

There is a propagation delay. After publishing from Desktop, new or changed answers can take around 15 minutes to appear. Our experience is that it is sometimes longer. When someone says "it doesn't work", check the timing before you check the configuration.

Where it works well

For a well-scoped model with a clear set of recurring questions, verified answers are excellent. We worked with a mid-sized logistics business where operations managers asked the same handful of questions every Monday: on-time delivery by depot, cost per pallet by lane, and open exceptions by customer. We built eleven verified answers, each with five or six trigger phrases and a region filter. Within a few weeks, the managers had stopped asking the analytics team for the same charts, and the numbers in the chat matched the numbers in the weekly pack.

The consistency benefit is the one we would prioritise. Copilot's general answering is impressive, but it will sometimes choose a different measure or a different date basis than your official report. A verified answer removes that ambiguity for the questions that matter most. The model also gets better at understanding phrasing over time because users keep interacting with the approved answers.

Our honest take

Verified answers are useful but not finished. Three things we would tell any client considering this:

First, the Q&A deprecation matters. The documentation says Q&A experiences are going away in February 2027, and Microsoft recommends Copilot for Power BI instead. Verified answers sit inside the Q&A setup flow in the current preview, so you should not build your entire self-service strategy around them as a long-term fixture. Use them for the highest-value questions now, but plan for the feature set to shift. Check the current Microsoft guidance before you commit a lot of effort.

Second, do not try to verify everything. The limit is 250 verified answers per model, but the real constraint is governance. If you have 200 answers nobody owns, they will drift out of date and users will lose trust. We recommend starting with ten to twenty questions that come up repeatedly, assigning an owner to each, and reviewing them monthly.

Third, test with real prompts from real users. Our own test prompts always work. The prompts from a sales manager who types in full sentences with typos rarely do. Run a short pilot with three or four people from the team who will actually use it, and record the prompts they type. You will learn more in one afternoon than in a week of internal testing.

A practical checklist

Before you publish a verified answer, we run through this list:

• The visual uses only model-level measures, not report measures.
• Every field in the visual and every filter field is visible in both the model and schema selection.
• Filters you want users to adjust are set to All, not to a specific value.
• Date filters use fixed ranges, not relative ones.
• Five to seven trigger phrases, written in the words your users use.
• You have tested the answer in Desktop with the skill picker set to answer data questions, with create-page skills turned off.
• Someone is named as the owner, with a date to review it.

If you are working through this on a live model and want a second pair of eyes, our Power BI consultants do this kind of review regularly. If the wider question is how AI fits into your reporting and decision-making, our AI for business intelligence team can help you map where verified answers add value and where they do not. For teams who need to get their people comfortable with Copilot in the first place, our Copilot training sessions cover the day-to-day habits that make these answers useful.

The short version: verified answers are a good way to make Copilot's answers match your official numbers for the questions that matter most. Start small, assign owners, and test with real people before you call it finished.

Reference: Prepare your data for AI - Verified answers (Microsoft Learn)