Suppose a sales manager asks Copilot, “How are sales performing this quarter?” The model has one measure for gross revenue and another for net sales. It also tracks both Order Date and Shipment Date. Copilot could return a convincing answer using the wrong measure or period without showing an error.
An AI-ready semantic model makes those choices clear before anyone asks a question. Microsoft provides features that give Copilot more guidance, but the underlying calculations must still be correct. This guide follows the work from the first model review through approval and ongoing checks.
Why Copilot Reliability Depends on the Semantic Model
The Problem With Vague Questions
“How are sales performing?” could mean revenue to the finance team and order volume to an operations manager. Without an agreed measure, Copilot has to interpret the question. Two reasonable interpretations can lead to very different answers, even when both use valid data.
Microsoft's model guidance makes the order clear: fix modeling problems before writing AI instructions. A wrong relationship can inflate sales totals. Summing daily inventory balances can also give a figure that makes no business sense. Neither error goes away because Copilot receives a better description.
Why This Matters Now
Microsoft says Power BI Q&A experiences will retire in February 2027 and recommends Copilot for natural-language questions. Teams moving from Q&A will need to retest familiar questions, rather than assume the same wording will always produce the same answer. Microsoft's Q&A guidance confirms that date.
Poor data preparation is a wider AI problem. In 2025, Gartner predicted that organizations would abandon 60% of AI projects lacking AI-ready data through 2026. That is a general forecast, not a Power BI Copilot failure rate. For BI teams, the immediate task is to agree on business definitions before people depend on AI answers.
What an AI-Ready Semantic Model Needs
Copilot does not need a separate version of your business data. It needs a semantic model built around measures that the business already trusts. Microsoft recommends reviewing the model's structure and documenting what its fields mean before wider use.
A model might include both [Order Count] and [Order Line Count]. They answer different questions, since one order may contain several lines. Copilot can create calculations during a session, but Microsoft recommends keeping important, frequently used calculations as reviewed measures in the model.
How to Build an AI-Ready Semantic Model in Five Steps
The first work happens in the model, not the Copilot settings. Once the calculations are correct, the team can describe the intended meaning and test how Copilot responds. Doing it in the opposite order risks covering a data problem with more instructions.
Step 1: Fix the Structure First
Start with table relationships and the level of detail in each fact table. If one row represents an order line, a measure must account for that when counting orders. Check the active relationships too, particularly when a sales table has more than one date column.
- Review relationship cardinality and mark alternate relationships inactive where appropriate.
- Name fact tables clearly and record what one row represents.
- Place descriptive fields in suitable dimension tables.
- Add useful hierarchies for dates or locations.
- Fix data types and inconsistent values, such as Open and open in a status field.
The difference between Order Date and Shipment Date matters for the same reason. A sales question may use the first, while a delivery question needs the second. The model should make both available without obscuring their purpose.
Step 2: Define and Review Measures Once
Consider [Net Sales]. The business needs to agree on how that measure handles returns, discounts, and cancelled orders. Once the definition is approved, reports and Copilot can use the same calculation instead of maintaining competing versions.
Other measures need similar care. Gross margin percentage should use approved totals rather than averaging row percentages. Inventory reporting may need an ending balance instead of a sum across daily snapshots. Keep [Order Count] distinct from [Order Line Count], because the counts serve different purposes.
After reviewing the model, the team can request certification. Microsoft's endorsement documentation says certification is available only when an administrator enables it and names the people allowed to approve content. Certification is not the same as marking a model Approved for Copilot.
Step 3: Give Fields Names People Understand
A field called CustNo makes sense to its author but may mean little to someone asking Copilot a question. Customer Number is clearer, just as Average Customer Rating is clearer than AvgRating. Names can remove doubt before the question reaches a calculation.
Descriptions handle details the name cannot. Microsoft's semantic model guidance says Copilot uses only the first 200 characters of a table, column, or measure description in the relevant DAX experience. Put the business meaning first and include exclusions that could change the result.
For [Net Sales], the description could say:
Power BI Q&A also supports synonyms that connect everyday words to model fields. Those may be useful in supported Copilot experiences, but a synonym that worked in Q&A still needs checking with Copilot. Don't treat the two experiences as identical.
Step 4: Record the Rules Behind Common Questions
A user might ask for sales “this year” without naming a calendar or fiscal year. That choice belongs in the model's business guidance, along with the default date for sales questions. If no single meaning is appropriate, Copilot can be told to ask for more detail.
Power BI's Prep data for AI provides three features for this work:
For a company whose fiscal year starts in April, an instruction can define that calendar for sales questions. Another can name Order Date as the usual basis for those numbers. Verified answers can cover frequently requested visuals where the business has already checked the result.
These features do not all affect the same Copilot tasks. Microsoft's documentation says AI data schemas do not apply to report summaries or certain DAX work. They also do not replace security permissions on the model.
Step 5: Test Before Marking the Model Approved
The model may look ready, but the real check is how Copilot handles business questions. Build a set with the expected measure and a validated answer for each question. Include vague requests like “How are sales performing?” to find cases where Copilot chooses the wrong meaning.
Power BI Desktop's Copilot pane lets authors test the prepared model. Its skill picker can simulate different Copilot experiences, while How Copilot arrived at this shows details behind supported answers. If a total is wrong, review the measure and relationship before changing an instruction. For date errors, compare the chosen period with the business rule.
Once those checks pass, the model owner can mark the model Approved for Copilot in Power BI service settings. This changes how the model and its reports are treated in standalone Copilot. It does not certify the model or guarantee the same answer every time. Microsoft says approval changes can take up to 24 hours to appear across models with many reports attached.
Prep Data for AI and Approved for Copilot: What Each Does
These controls sound similar, but their jobs differ. AI instructions tell Copilot how to interpret business language. An AI data schema limits the fields considered for supported model questions, while a verified answer points to a chosen visual. Approval is a separate signal that the model has been prepared for standalone Copilot.
Microsoft still describes Prep data for AI as a preview feature. In Power BI Desktop, authoring works with Import, DirectQuery, and local Composite connections. The Power BI service supports additional model types, and Q&A must be enabled on the semantic model for these features. Trial Fabric capacities do not support Copilot.
There are other limits worth testing. Verified answers have restrictions around RLS and OLS during preview, so you should not treat them as security controls. Changes made through Git or deployment pipelines may need a service refresh before Copilot uses them. Microsoft also notes that AI answers can vary, even when the question stays the same.
Governing the Model as It Changes
An AI-ready model will not stay unchanged after approval. Finance may revise a measure, or an analyst may rename a field that Copilot uses. Each change deserves a review against the saved business questions. A named owner can make that part of normal model maintenance instead of waiting for users to report wrong answers.
Keep descriptions and AI instructions with the model's tracked changes. Review the question set when key measures change, not only when the team updates Copilot settings. That gives reviewers a record of the rule behind an answer and when it changed.
Governed AI intelligence for external reporting
The same preparation matters when Power BI reports go to customers outside the organization. Native Copilot in Power BI is not the same as an AI feature in an embedded customer portal. Some teams use a separate AI layer in their Azure environment to query the existing semantic model. That layer still depends on correct measures and the permissions of the user asking the question.
Query audit logs can show administrators how such an AI system reached an answer. They cannot repair an inaccurate measure or replace a model security rule. Both internal Copilot experiences and external AI reporting depend on the business meaning already defined in the model.
Frequently Asked Questions
What Is an AI-Ready Semantic Model in Power BI?
It is a model with checked relationships and reviewed measures. The fields have clear names, while important business rules are documented. Copilot can then work from definitions that the BI team already trusts.
Can AI Instructions Fix a Badly Built Model?
No. They can guide Copilot toward the right measure or interpretation, but they cannot correct a bad relationship. Microsoft recommends repairing the model first.
Does Certification Make a Model Approved for Copilot?
No. Certification is an endorsement granted by authorized reviewers. Approved for Copilot is a separate setting, so a certified model still needs Copilot testing.
How Long Should Measure Descriptions Be?
Put the essential meaning within the first 200 characters. In the relevant DAX experience, that is the portion Copilot uses. Include important exclusions near the beginning.
Will Copilot Give Identical Answers Every Time?
Not necessarily. Microsoft says AI responses can vary between runs. A reviewed question set and how Copilot arrived at this view can help the team spot results that need another look.
References
- Microsoft Learn. What Is Microsoft Fabric? learn.microsoft.com, 2024.
- Microsoft Learn. Prep Data for AI in Power BI. learn.microsoft.com, 2026.
- Microsoft Learn. Copilot Semantic Model Guidance. learn.microsoft.com, 2026.
- Microsoft Learn. Deprecating Power BI Q&A. learn.microsoft.com/ community.fabric.microsoft.com, September 2026.
- Microsoft Learn. Verified Answers for Copilot. learn.microsoft.com, 2026.
- Microsoft Learn. Endorsement: Promote and Certify Content. learn.microsoft.com, 2025.
- Microsoft Learn. Approved for Copilot Setting. learn.microsoft.com, 2026.
- Microsoft Learn. Copilot Pane in Power BI Desktop. learn.microsoft.com, 2026.
- Gartner. Lack of AI-Ready Data Puts AI Projects at Risk. gartner.com, February 2025.