// loading…
Accreditations
Article · Data & Analytics
Copilot is only as good as the model behind it — a seven-step guide to preparing Power BI semantic models for Copilot and Fabric Data Agents, with testing.
Category
Data & Analytics
Written by
Xpertnest Editorial Team
Published
12 Aug 2026
Read
6 min read
Microsoft Copilot and Fabric Data Agents let business users ask questions in plain English and get an answer in seconds. It is a genuinely compelling demo — and it is also where most organisations get caught out.
Because the answer Copilot returns is only ever as good as two things: the quality of the underlying data, and the design of the Power BI semantic model sitting behind it. Get those right and AI becomes a trusted analyst. Get them wrong and you have built a very fast way to distribute incorrect numbers to your business.
This guide covers the seven steps that separate the two outcomes.
Understanding the pipeline explains why each preparation step matters:
User question → Measure and field selection → Date, filter and security checks → Query execution → AI-generated answer
Every stage depends on the one before it. Poor-quality data, unclear measures or unnecessary technical fields at any point can cause the AI to return incorrect or misleading answers — delivered with complete confidence.
Before you build the semantic model, clean the data. This work can be completed in the source system, Power Query, Fabric Dataflows or a data warehouse — wherever it fits your architecture.
Copilot cannot reliably produce accurate answers from poor-quality data. No amount of downstream configuration fixes this.
Once the data is clean, create a clear star-schema design using well-defined fact and dimension tables. Keep relationships simple, use clear date tables and avoid ambiguous filter paths.
This is what allows Copilot and Fabric Data Agents to understand how your business data is actually connected. Ambiguity in the model becomes ambiguity in the answer.
Important KPIs should use explicit DAX measures with meaningful names. Naming is not cosmetic here — it is how the AI identifies what a user is asking for. Rename vague fields to clear business names: Rev2 becomes Net Revenue, SalesAmt becomes Gross Margin %, and Date1 becomes Total Sales.
Every measure description should explain:
A measure named Rev2 with no description forces the AI to guess. A measure named Net Revenue with a clear definition removes the guesswork entirely.
There is a strong instinct to expose everything, on the basis that more data means better answers. The opposite is true.
Show Copilot the approved measures; the customer, product and region dimensions; the correct date fields; and the fiscal calendar. Avoid exposing technical IDs, audit columns, raw calculations and sensitive information.
A focused AI data schema reduces confusion and helps Copilot select the correct business fields. Note that the AI data schema supports relevant data-question experiences, but it may not control every Copilot feature.
AI instructions tell Copilot how business questions should be interpreted in your organisation. For example:
Instructions encode the business logic that lives in people's heads. But be clear about the limit — instructions guide Copilot, they cannot fix incorrect DAX, broken relationships or bad source data.
Verified answers are worth the effort for important or frequently asked questions, such as:
An approved measure, a named business owner, a last-tested date and a review date.
These are the questions where an incorrect answer causes real damage. Verified answers convert them from a risk into a controlled, owned asset.
The most common mistake is approving a model after a handful of successful questions. That proves very little.
Test across:
Measure + Date + Filters + Number + Explanation + Security.
A numerically correct answer can still be wrong if it uses the wrong measure, the wrong date or the wrong business definition. Testing only the number misses most of the risk.
Best suited to exploring Power BI reports and semantic models.
Best suited to conversational analytics across governed Fabric sources — semantic models, warehouses and lakehouses.
When querying a Power BI semantic model, Fabric Data Agents can use its metadata and Prep data for AI configuration to generate DAX queries. The preparation work described above therefore benefits both.
Preparing a Power BI semantic model for AI follows a clear sequence: clean data → star schema → clear measures → focused AI data → AI instructions → verified answers → security testing.
One caveat worth stating plainly: Prep data for AI is currently a preview capability. It can meaningfully improve answer quality, but it cannot guarantee that every AI-generated answer will be correct.
The most reliable results come from clean data and a clear, secure and regularly tested semantic model. There is no shortcut around that.
Usually because of poor-quality source data, ambiguous relationships, unclear measure names or too many technical fields exposed to the AI. Each of the seven steps above addresses one of these causes.
No. Exposing technical IDs, audit columns, raw calculations and sensitive information increases confusion and reduces accuracy. A focused AI data schema produces better answers.
No. Instructions guide how Copilot interprets questions, but they cannot correct incorrect DAX, broken relationships or poor source data. Fix the model first.
More than a few successful questions. Cover KPI and trend questions, fiscal periods, synonyms and misspellings, ambiguous questions, unsupported requests and unauthorised questions — then verify measure, date, filters, number, explanation and security on each answer.
Written by
Xpertnest Editorial Team
Insights · Xpertnest
The Xpertnest team writes about applied AI, hyper-automation and managed services, drawing on delivery experience across application support, ITSM and enterprise operations.
Keep exploring
Tell us what you're trying to build — or browse more of our thinking.