Use an AI data schema to narrow which model fields Copilot should prioritize, a verified answer to route a known question toward an author-validated report visual, and AI instructions to supply business context or interpretation rules that the model does not express clearly. They solve different failures. If “sales” maps to five fields, fix the schema. If “busy season” has a company-specific meaning, use an instruction. If an approved visual already answers a high-value question, consider a verified answer. None of them repairs an incorrect measure or unsafe relationship.
Microsoft currently presents these controls through Power BI’s Prep data for AI experience and recommends an implementation order of AI data schema, verified answers, AI instructions, then descriptions. Microsoft labels the broader preparation workflow as preview and notes that outputs remain nondeterministic. Prep data for AI FAQ
Control-selection matrix
| Decision | AI data schema | Verified answer | AI instructions |
|---|---|---|---|
| Primary job | Focus the model objects Copilot prioritizes for data questions | Return an author-selected visual for matching questions | Add business context, terminology mappings, or analysis rules |
| Best signal | Copilot chooses the wrong field or sees too many similar objects | A common, consequential question already has a maintained visual | The meaning depends on organization-specific language or policy |
| Configuration level | Semantic model | Saved to semantic model; created from a report visual | Semantic model |
| User visibility | Consumers cannot see the configured schema | User can receive the selected visual as the answer | Consumers cannot see the instruction text |
| Good example | Expose [Net Revenue], not raw revenue columns and deprecated measures | “Sales by area” returns the approved sales-by-region visual | “Busy season means June through August” |
| Poor use | Trying to solve a wrong DAX measure by hiding evidence of it | Treating any displayed visual as verified without an owner or benchmark | Writing “always be accurate” or pasting a long glossary with conflicts |
| Main limitation | It does not affect every Copilot capability | Coverage depends on a maintained visual and suitable trigger phrases | Instructions are not persona-specific and might not be respected in every experience |
The product-behavior cells summarize Microsoft’s AI data schema, AI instructions, and Prep data for AI documentation. The “good” and “poor” uses are Refinity recommendations.
AI data schemas: reduce the choice set
An AI data schema selects a focused subset of semantic-model fields for Copilot to prioritize when answering data questions. This is the right control when the full authoring model contains objects that are technically valid but irrelevant, deprecated, overly technical, or confusing for the target users.
Use an AI data schema when
- several measures have overlapping names;
- raw numeric columns compete with governed measures;
- the model includes technical tables and keys users never ask about;
- a broad enterprise model serves a narrow user group;
- testing shows Copilot selects a plausible but unintended field.
Do not overstate its reach
Microsoft documents that an AI data schema applies only to capabilities that reference the model schema for data questions. It does not affect report summaries or questions answered directly from report-page visuals. Microsoft also says report-page creation, search, and DAX-query experiences can require the entire model instead of the AI schema. Relationships can still affect answers even when only one related field is selected. AI data schema considerations
That means “not in the AI schema” is not a universal security boundary. Use Power BI permissions and model security for access control.
AI-schema decision record
| Object | Decision | Question supported | Ambiguity removed | Owner |
|---|---|---|---|---|
[Net Revenue] | Include | Revenue trend and variance | Establishes preferred revenue metric | Finance analytics |
Sales[Revenue] | Exclude | None | Prevents unsupported implicit sum | Model owner |
Date[Fiscal Quarter] | Include | Fiscal-period questions | Provides approved period label | Finance data |
Date[Calendar Quarter] | Exclude for this use case | None | Avoids an unresolved calendar choice | Model owner |
Review the register against the representative question set. A tiny schema that cannot answer the intended questions is not “clean”; it is incomplete.
Verified answers: control a known response path
A verified answer associates likely user phrasing with an author-selected visual. Microsoft recommends verified answers for common or nuanced questions when a report visual can derive the answer. Prep data for AI FAQ
Use a verified answer when
- the question is common and consequential;
- an existing visual contains the approved metric, grain, calendar, and filters;
- the visual has a named owner and benchmark;
- users ask the question using a known range of phrases;
- returning a consistent visual is more useful than generating a new analysis.
A verification checklist for the “verified” artifact
Before creating the answer, confirm:
- The visual’s measures have current definitions and owners.
- The visual reconciles to an independent benchmark.
- Its period and filter context are visible or documented.
- It behaves correctly for each target security role.
- Its trigger phrases come from user language, not only field names.
- A model/report change will trigger retesting.
Microsoft provides the feature; Refinity recommends this evidence standard. “Verified answer” should not become shorthand for “somebody selected a visual once.”
AI instructions: express missing business context
AI instructions tell Copilot how to interpret organization-specific terms or approach certain analyses. Microsoft examples include defining a business season and giving rules for how data should be analyzed. AI instruction use cases
Use an instruction when
- a term has a defined local meaning that does not exist as a model object;
- a user phrase must map to a specific business concept;
- an analysis rule is valid for the entire semantic model;
- the rule can be tied to representative questions and tested.
Keep instructions atomic
One instruction should express one rule:
When this user group says “run rate,” use the approved
[Quarter Exit Run Rate]measure. Do not interpret it as year-to-date revenue divided by elapsed months.
This is testable. A paragraph mixing calendar rules, presentation style, data definitions, caveats, and persona guidance is harder to debug and more likely to conflict.
Microsoft says AI instructions are model-level, not report-level; consumers cannot see or disable them; they are not intended to be persona-specific; and they have a 10,000-character limit. Microsoft also identifies circumstances where instructions may not be respected in Power BI Desktop. AI instruction limitations
Worked example: three controls, one question
Question: “How did sales perform in busy season by area?”
The model contains [Gross Sales], [Net Revenue], Product[Area], Customer[Region], and a standard calendar. The organization defines busy season as June through August, while commercial leaders use “area” to mean customer region. A maintained report visual already shows Net Revenue by Region for that period.
| Layer | Configuration | Why |
|---|---|---|
| Semantic foundation | Confirm [Net Revenue] is the approved measure and the visual uses the correct relationship/filter context | The controls below must not sit on a wrong calculation |
| AI data schema | Include [Net Revenue] and Customer[Region]; exclude competing fields not needed for this use case | Reduces field-selection ambiguity |
| Verified answer | Associate representative trigger phrases with the maintained visual | Provides an approved path for the known question family |
| AI instruction | Define busy season as June through August and “area” as customer region for this model | Supplies local language not fully expressed by field names |
| Evaluation | Test phrasing variants, repeated runs, visual values, and security roles | Confirms the combination works under actual conditions |
The controls are complementary because each addresses a different decision. If the source visual double counts revenue, however, all three controls will faithfully preserve a bad answer.
Decision workflow
For each failed question, ask in order:
- Is the approved business answer itself defined and reproducible? If no, resolve ownership, metric, grain, calendar, filters, or security first.
- Did Copilot select the wrong model object? If yes, review naming, measures, descriptions, and the AI data schema.
- Does an approved visual already answer this known question? If yes, evaluate a verified answer.
- Is organization-specific context still missing? If yes, add one focused AI instruction.
- Did the result pass across prompt variants and user roles? If no, localize the new failure before stacking more configuration.
Microsoft product facts vs. Refinity recommendations
Microsoft product facts
- AI data schemas focus the fields Copilot prioritizes for certain data questions.
- Verified answers can return an author-selected report visual for matching questions.
- AI instructions add model-level business context and analysis rules.
- Each feature has capability-specific limitations.
- Copilot output remains nondeterministic.
Refinity recommendations
- Repair the semantic model before configuring AI controls.
- Give each control one named failure to solve.
- Record include/exclude decisions and instruction rationales.
- Apply an evidence standard to the source behind a verified answer.
- Regression-test the combined configuration after every relevant change.
Primary CTA: Check whether your model has the foundations and lifecycle controls these features depend on.
Related reading: