Power BI Copilot can give a wrong answer when the semantic model leaves an important choice unresolved: which measure represents the business term, which date controls the period, which grain is safe to aggregate, or which filter a user intended. A dashboard can still look correct because its author already made those choices in a visual. Copilot must infer them from the question and the model context. Start by tracing the answer through terminology, schema, calculation, security, and testing—not by endlessly rewriting the prompt.

Microsoft states that an unprepared model can produce less useful or inaccurate results, and that Copilot output is nondeterministic even when the prompt and grounding data are the same. Microsoft’s current preparation features can reduce ambiguity, but they cannot ensure one specific output every time. Prepare your data for AI

The failure path

A natural-language answer passes through more decisions than a fixed report visual:

  1. Interpret the user’s words.
  2. Select model objects that appear to match.
  3. Generate or select a semantic query.
  4. Apply relationships, filters, time logic, and the user’s permissions.
  5. Choose a visual and summarize its result in language.

Microsoft documents that Copilot grounds on semantic-model information such as tables, columns, measures, relationships, calculation groups, descriptions, data types, and format strings. For data questions it can return a generated visual plus a natural-language summary of the query result. Use Copilot with semantic models

That makes the model—not the prompt—the first place to investigate when an answer is plausible but wrong.

Diagnostic matrix: symptom to likely control

SymptomLikely failure mechanismEvidence to inspectFirst control to test
“Sales” returns a valid but unexpected numberCompeting measures or columns can all plausibly mean salesMeasure list, descriptions, report definitions, owner-approved metric glossaryExpose one approved measure in the AI data schema; hide or rename unsafe alternatives
The value is right for a different periodOrder date, invoice date, fiscal calendar, or rolling-period logic is ambiguousDate relationships, calculation groups, fiscal calendar specification, expected-answer queryMake the calendar contract explicit in measures and terminology; add a focused instruction only after the model is sound
Totals are inflatedFacts at different grains are being combined or summarized through an unsafe relationshipTable grain, relationship cardinality and direction, DAX query, duplicate-key testsCorrect the model or measure; do not try to prompt around a grain defect
Answers vary when the wording changes slightlyTrigger phrases, terminology, or schema context do not cover normal user languagePrompt variants, field-selection traces, verified-answer triggersAdd synonyms/descriptions, then test an AI instruction or verified answer for the specific class of question
Copilot chooses the wrong fieldToo many overlapping fields are availableAI data schema, hidden fields, duplicate business termsNarrow the AI data schema to the fields that serve the target users
The visual is correct but the prose overstates itThe generated summary goes beyond the returned dataQuery result, visual, generated narrativeDefine an answer contract that separates returned facts from interpretation
One user receives an empty or different resultPermission context or row-level security changes the accessible dataRole mapping, “view as” tests, user identity, source permissionsTest each target role; do not validate only as an administrator
Copilot times out or returns a generic errorThe model, DAX, or query is too complex or insufficiently optimizedPerformance Analyzer, DAX query, model size and relationshipsSimplify or optimize the model/query before adding more instructions

The matrix is a Refinity diagnostic method. The underlying product behaviors are described in Microsoft’s semantic-model Copilot guidance and AI data schema documentation.

Worked example: “What was revenue last quarter by region?”

Suppose the executive dashboard shows Net Revenue by fiscal quarter. The model also exposes:

  • a raw SalesAmount column that can be summed implicitly;
  • a Gross Revenue measure used by finance;
  • order date and invoice date relationships;
  • calendar and fiscal-quarter labels;
  • a region on both customer and sales-territory tables.

The dashboard author chose Net Revenue, fiscal quarter, invoice date, and customer region when building the visual. The user’s sentence contains none of those decisions. Several generated queries could be technically valid while failing the business intent.

The durable repair is an answer contract:

Contract fieldApproved interpretation
Business termRevenue means [Net Revenue] for this user group
SourceCertified priority semantic model
GrainOne result per fiscal quarter and customer region
CalendarCorporate fiscal calendar
Date roleInvoice date
FiltersExclude internal transfers; respect the viewer’s security role
Expected evidenceTotal reconciles to the approved finance report for the same period

Then make the model and AI configuration express that contract: expose the approved measure and dimension, remove confusing fields from the AI schema, create a verified answer if a maintained report visual already answers the high-value question, and use a concise instruction for organization-specific language that the model cannot otherwise express.

A five-step investigation that avoids prompt whack-a-mole

1. Freeze the expected answer

Write the expected measure, period, dimensions, filters, security role, and trusted comparison before testing Copilot. “The dashboard says 12.4” is not enough unless you can explain how that dashboard produced 12.4.

2. Reproduce with representative wording

Test the business user’s wording, not only a model author’s ideal prompt. Include a direct question, a common synonym, a follow-up, and an intentionally ambiguous version. Repeat important tests because Microsoft identifies Copilot output as nondeterministic.

3. Localize the first wrong decision

Determine whether Copilot misunderstood the term, selected the wrong field, generated the wrong calculation, applied the wrong time/filter context, or summarized a correct result incorrectly. Fix the earliest failing layer.

4. Change one control

Keep the expected answer and prompt set fixed. Change one of the model, AI data schema, verified answer, or instructions. Microsoft recommends implementing the preparation controls in sequence: AI data schema, verified answers, AI instructions, then descriptions. Prep data for AI FAQ

5. Turn the incident into a regression test

Add the question, variants, expected answer contract, security roles, and tolerance rule to a versioned evaluation set. Microsoft’s content lifecycle guidance recommends testing known benchmarks, comparing Copilot output with trusted reports or analyses, and comparing results before and after model changes. Validate Power BI content

What not to fix with AI instructions

AI instructions are useful for business context and interpretation rules, but Microsoft says they are stored at the semantic-model level, are not persona-specific, and are not guaranteed to be respected in every Copilot experience. AI instructions

Do not use instructions to conceal:

  • an incorrect measure;
  • a many-to-many or mixed-grain double count;
  • an untested row-level security design;
  • an unclear ownership decision;
  • two teams using incompatible definitions under the same name.

Those are semantic-layer and governance problems. Fix them where they originate.

Microsoft product facts vs. Refinity recommendations

Microsoft product facts

  • Copilot uses semantic-model metadata and query results as grounding context.
  • Current Prep data for AI controls include AI data schemas, verified answers, and AI instructions.
  • The controls are configured on the semantic model, not separately for each report.
  • Copilot output remains nondeterministic.
  • AI data schemas apply to certain schema-based data questions, not every Copilot capability.

Refinity recommendations

  • Define an answer contract before changing prompts or AI configuration.
  • Treat wrong answers as traceable failures across terminology, model, query, security, and presentation layers.
  • Test important questions across prompt variants and permission roles.
  • Require evidence for the “correct” comparison result.
  • Promote production use only when failures become versioned regression tests.

When the problem needs a readiness review

A technical readiness review is warranted when the same term can resolve to several measures, leaders disagree about the benchmark, important answers depend on undocumented calendar logic, security has not been tested by role, or the team cannot show a repeatable evaluation set. Those conditions are broader than prompt design.

Primary CTA: Score your model across seven Power BI Copilot readiness dimensions.

Related reading: