A Copilot Studio agent with the Microsoft 365 Copilot channel configured, running inside M365 Copilot - not an Agent Builder agent. It uses model metadata to generate DAX, query Power BI as you, and explain results.
1. Ask a questionActual prompt from the published M365 test. The tested Power BI semantic model contains agents, creators and usage data.2. Get a model-grounded responseActual output, cropped to the first 8 of 20 rows. People's names are masked; agent names, values and dates are unchanged. View the complete response and test notes.
Observed response, not an independent numerical audit. Experimental starter: configure your own connection, model metadata and business definitions before publishing. This website is documentation, not a live agent.
Published M365 conversations tested. A generic top-20 request returned agents and creators; a top-five follow-up matched the prior first five; advice returned explicitly unexecuted DAX. Complete redacted captures appear below. This is scoped response and consistency evidence, not universal query correctness. The earlier repeated-value concern is deferred and the combined single-turn request remains unresolved.
Built in Copilot StudioPublished through the M365 Copilot channel; not built in Agent Builder.
Existing Power BI modelQuery the semantic layer, not exported spreadsheets.
Your execution identityExplicit end-user connection, no maker fallback.
A custom example, not Microsoft Agent 365. The demonstration model/report happens to be named Agent365. It is not the Microsoft product or an official product schema. Use this pattern with other compatible Power BI models by adapting the model context, queries, configuration, and permissions.
01 / How it works
A small set of capabilities. Many different questions.
Message -> Copilot Studio / topics -> Power BI tool -> answerTopics control the workflow; the embedded Run a query against a dataset connector action executes the DAX. The separate tool box makes its role visible, not a separate deployment. Detailed diagram · Editable detailed source.
A governed metadata snapshot supplies structure. The orchestrator is expected to author a new DAX table expression. Native runtime controls form a bounded execution envelope. Deployment of those pieces is not proof of successful orchestration.
01 / ASKBusiness questionNatural language from an authenticated user.
02 / INTERPRETCopilot StudioRetrieve metadata; generate a new DAX expression.
03 / PREPAREExecution envelopeStructural checks, projection, ordering, and result bounds.
04 / EXECUTEPower BI connectorRun a query against a dataset as the caller.
05 / RETURNSemantic modelMeasures and permitted data -> grounded answer.
Governed metadata, checked as the callerOwner-run definition preparation retains structural context. A zero-row Invoker probe gates runtime disclosure. Refresh and republish after schema changes.
Advice does not execute the business queryCompile proposed DAX and label it unexecuted. Metadata authorization may still probe Power BI. Exact existing-measure implementations are not in the snapshot.
PRESERVE
The top-100 experience
Keep the successful ranking presentation as a regression scenario, including ordering, metric definition, and all requested rows up to the limit.
GENERALIZE
Expressions, not metric templates
Generate new combinations of actual fields, measures, filters, and calculations. The input is a table expression, not every possible full DAX script.
EXPLAIN
DAX with context
Use real model names, explain filter context and time scope, and distinguish a query from a measure. Suggested code is not verified code.
Current contract: one onboarded primary model; generated table expressions, aliases/order, and scalar date bounds; up to 100 rows, 16 columns, and 256 characters per text cell with explicit truncation. No five-metric or owner-field enumeration. The local Python modules prepare metadata and native source; they are not a hosted runtime service.
02 / Example conversations
Ask for an answer. Or ask how.
Public-safe examples
Tested against a Power BI semantic model containing agents. The demonstration model includes agent inventory, creator/ownership and usage data, which is why the test questions ask about agents, creators and interactions. Copilot Studio queries that semantic model with DAX, not the report visuals. Other compatible models require their own metadata and business definitions.
The report behind the agent-focused exampleUser-provided view of the custom report connected to the tested semantic model. This is a Power BI report, not an agent response or a synthetic illustration. Its displayed date window differs from the conversation tests, so the totals are not a like-for-like comparison. The report and semantic model are not included in the solution ZIP.
These examples document the intended experience and evaluation criteria. Synthetic values are clearly marked; supported combinations and test status live in the source documentation.
Question -> ranked result
Keep the useful answer format
Give me the top 100 agents by usage.
Illustrative output: Agents ranked by audited Interactions for the selected period. Return every available requested row up to 100 and state the date scope.
Rank
Agent
Interactions
1
Service Desk Demo
1,284
2
Policy Finder Demo
976
3
Onboarding Demo
811
Synthetic three-row excerpt only. Agent names and values are invented; this table is not test evidence.
Question -> supported aggregation
Reuse a capability, not a canned answer
Give me the top agents and their usage over the last 30 days.
Expected answer contract: retrieve relevant metadata, generate a ranking expression with paired date expressions using the trusted engine UTC anchor, and explain the actual requested scope. Do not substitute an all-history ranking.
Follow with "Add platform and region, apply these two filters, and calculate sessions per user." Those combinations should be authored from metadata rather than require a new metric template. Observe the actual generated expression before claiming success.
A dated usage ranking and creator follow-up are user-observed successful; this example is not a published transcript. With a 12 September 2026 UTC anchor, the requested window is 14 August-12 September inclusive. Production dates are dynamic. Observed event bounds are not refresh timestamps or proof of continuous coverage.
Question -> model-grounded explanation
Teach the query, not just its result
Explain how to rank agents by Interactions in DAX using this model. Explain filter context and do not run it yet.
A good response uses the actual approved schema, explains why the grouping and measure fit the question, and discusses date filters and tie handling.
EVALUATE
SUMMARIZECOLUMNS(
'Date'[YearMonth],
"Sessions", [Interaction Sessions]
)
ORDER BY 'Date'[YearMonth]
Model-grounded illustration, not executed as shown or captured from a live answer. The actual topic accepts a table expression and forms a bounded envelope; it does not accept this entire script as a tool argument.
State whether the code is a DAX query or a measure.
Prefer existing measures with the right business definition.
Label the answer suggested DAX, not executed.
Do not claim to create or save anything in Power BI.
This integration is read-only. Model authoring is not implied by DAX guidance.
Question -> honest boundary
Unsupported should be explicit
Rank agents by customer satisfaction and include their owners' email addresses.
Expected behavior: establish whether an appropriate satisfaction metric exists and what it means. Do not quietly substitute usage for satisfaction. Identity fields follow actual metadata and the requesting user's permissions; the retired blanket demo ban is not a model limitation.
Clarify ambiguity or explain actual access failures rather than inventing fields. Publication requires the owner's approval; personal identities and tenant configuration remain redacted.
Actual published-channel prompts and responses. With the owner's approval, only people's names are covered in the result captures; agent names, usage figures and dates are unchanged. Composites are labeled and no replacement values are invented. Checks establish visible shape and consistency, not independently audited numerical accuracy.
Review the Agents semantic model and provide me the top 20 used agents and their creators.
One generic request, 20 ranked rowsCreators included, displayed usage descending and distinct. The agent declared its own 30-day default. Values were not independently audited; an extra per-row date lacked a label.Context retained across turnsThe five agent, creator and usage triples matched the preceding first five. The agent retained the period but rendered a list rather than the requested table.Model-grounded DAX adviceSame-period platform comparison and filter-context explanation. Prose instead of a code block; neither semantic correctness nor business-query nonexecution was independently instrumented.Current generic capabilitiesMetadata, generated execution, advice and error handling live under Topics. Three older bounded topics remain inactive.Power BI access as the requesting userThe current published-channel consent step, not a historical tool-registration screen. Connection consent does not grant missing model permissions.Rendered model selection observedGPT-5 Reasoning (Preview). Current warnings concern preview-model suitability and no formal Studio evaluation; inference telemetry is not verified.
Yes, these capabilities call Power BI. They live under Topics, with embedded connector actions. There is no dedicated top-100 tool and no new tool per question. Copilot Studio selects reusable topics and fills their technical inputs from the question and verified metadata.
Get model metadata
Trigger: generative selection.
Validates the model alias, probes schema visibility as the requester, then returns the governed catalog or selected tables.
Inputs:modelAlias (String, blank defaults to primary), view (String: catalog/tables), tableNames (String: at most four verified names).
Output: metadata JSON in result, plus the shared status/visibility fields below.
Requires the current-turn/user metadata gate, validates newly authored DAX and forms a bounded query. The embedded Power BI action executes it as the requesting user.
Inputs: the seven query parameters below.
Output: bounded row-envelope JSON in result, the final generatedDax and execution status.
Uses the same expression contract, but compiles the proposed business query without executing it. The answer must label it unexecuted; metadata authorization may still probe Power BI.
Inputs: the same seven query parameters.
Output: proposed generatedDax and advice/status envelope, not business result rows.
Generated query error · Trigger: OnError. Reports the actual error and stops rather than fabricating an answer. No declared task inputs or outputs. Full synthetic error topic.
The actual tool: Power BI > Run a query against a dataset
Connector: Power BI (shared_powerbi). Action: Run a query against a dataset. Operation ID:ExecuteDatasetQuery. These are the names to look for in Studio, not the custom topic names above. The connector executes DAX; Copilot Studio generates it.
Topic
Actual connector action
Get model metadata
Run a query against a dataset: execute a schema-visibility probe, then return the prepared metadata snapshot. This is not a built-in metadata-discovery action.
Run generated DAX
Run a query against a dataset: execute the compiled business query.
Compile DAX advice
No connector action in this topic. Native Power Fx compiles the proposed query; the separate metadata topic may still probe Power BI.
Generated query error
No connector action. Native error handling reports and stops.
Add the tool yourself
Open your agent's Topics page and the query topic. Select + Add node > Add a tool > Connector, search for Power BI, and choose Run a query against a dataset.
Create or select a Power BI connection and retain user credentials / Invoker. The requesting user needs the appropriate Read/Build permissions and connection consent.
Set the fixed workspace/model and bind the query/output as shown below. Do not let the agent invent resource IDs or an impersonated user.
Add the same action in the metadata topic for its generated visibility probe. Keep advice and error handling as native topic logic, not extra Power BI actions.
Action field
Query-topic setting
Workspace (groupid)
Your authorized workspace, fixed configuration.
Dataset (datasetid)
Your semantic model, fixed configuration.
Query text (query)
Power Fx formula Topic.generatedDax (=Topic.generatedDax in native YAML). This is the complete compiled DAX query, not the raw tableExpression input.
Impersonate user (impersonatedUserName)
Leave blank (=Blank() in YAML); execution uses the requesting user's connection.
Output (firstTableRows)
Bind to Topic.RawRows. Preserve Any in the native output schema for arbitrary generated columns.
In the native action source, retain serializerSettings: ={includeNulls:false} and requestTimeoutInMilliseconds: 30000. The metadata action instead uses its prepared probe query, Topic.ProbeRows and the numeric AccessProbe output schema.
The orchestrator fills these arguments; the user asks an ordinary business question. Workspace/model IDs and execution identity are trusted configuration, not model-authored inputs.
Input
Type
Contract
modelAlias
String
Blank defaults to primary; only the onboarded primary alias is supported.
tableExpression
String
New DAX table expression, not a full EVALUATE/DEFINE/ORDER BY script; maximum 12,000 characters.
columns
String
1-16 distinct output aliases, comma-separated.
sortBy
String
Declared aliases and asc/desc; remaining aliases break ties.
limit
Number
Integer 1-100; default 20.
startDateExpression
String
Optional scalar DAX date bound, paired with end.
endDateExpression
String
Optional scalar DAX date bound, paired with start. Dates use UTC_TODAY, QUERY_START and QUERY_END.
Shared outputs
status, stage, probeResultStatus, resolvedModelAlias, error, result and generatedDax are Strings. connectorAttempted, connectorReturned and visibilityVerified are Booleans. Completion or a generated query alone does not prove execution, valid rows or semantic correctness.
Actual Studio configuration
Callable execution topicThe model-facing description distinguishes execution from advice. Ask-before-running is off here; this does not bypass Power BI connection consent.AI-filled DAX, not a fixed queryThe actual tableExpression input is dynamically filled. Studio's display-name warning is retained in this capture; no runtime settings were changed for documentation. The full input description and contract are in the source.
Embedded Power BI action: Run a query against a datasetExecuteDatasetQuery receives Topic.generatedDax, uses Invoker mode and returns dynamic rows as Any. Null omission is explicit; rows are serialized directly. The workspace and semantic-model IDs are redacted, not replaced with fake settings.
Working demo versus downloadable starter: these are genuine captures of the configured demo. The ZIP deliberately has no live metadata or connection binding and stops before queries until customized. Follow the full import guide. Read the full agent instructions.
Scope / Discovery / Scale
Reusable execution is not automatic model understanding.
The primary snapshot contains 21 tables, 244 columns, 166 measure names, and 13 relationships. It is prepared metadata, not live per-turn definition discovery, complete business semantics, or verified conversational reasoning.
Importable solution available. The unmanaged starter ZIP was successfully imported as a separate unpublished agent. It intentionally contains no demo schema or connection binding and stops before queries until configured. Download, import and customize for your report/model.
DEPLOYED / SCOPED SUCCESS
Generic expression capabilities
Metadata, dated ranking and creator follow-up are user-observed working. Other question shapes still need evaluation; structural controls are not a complete DAX parser, semantic proof or cost estimator.
REQUIRES ONBOARDING
Another model's meaning
Prepare current metadata and validate permissions and semantics. No new five-metric template is intended, but changing only a dataset ID is still not a verified onboarding process.
NOT YET DEMONSTRATED
Automatic multi-model discovery
A model catalog, focused metadata retrieval, routing, schema refresh, and cross-model interpretation remain separate work. No model-count or concurrency ceiling has been measured.
Not exposed does not mean nonexistent. The generic contract no longer bans owner/creator fields. Actual schema visibility, permissions, missing metadata, and genuine absence remain distinct. The all-column metadata probe may fail closed for narrower OLS users, who can require a role-appropriate snapshot.
Second-model probe: structure discovery succeeded
A genuinely different model accepted three constant DAX queries. Its definition was retrieved automatically using the documented Fabric API: 78 tables, 694 columns, 17 measures, and 79 relationships, including hidden/generated objects.
This used existing read/write model permissions. No nonempty object descriptions were found; authored-instruction discovery remains unproven. The schema has not been wired into the agent, and this is not evidence of automatic business understanding or read-only-user metadata access.
Observed constant-query request times: 823 / 461 / 538 ms. Three sequential samples, not a benchmark, concurrency test, or complete-chat latency.
DIRECT TESTA constant-only DAX query executed successfully against the model.
DIRECT TESTCounts and a 100-row ranking executed using the audited Interactions metric.
USER REPORTThe user confirmed they liked the top-100 output.
SCREENSHOTActive topics, AI-filled DAX input, Invoker connector configuration and published-channel consent are visible.
OFFLINEThe generic source has 54 Python tests, 5 client-harness tests, and 55 synthetic native Power Fx checks; the native null-serialization caveat is documented.
DIRECT TESTEight varied expression cases passed, including new combinations and the original ranking membership/order.
M365 UITop-20 ranking, consistent top-five follow-up and explicitly unexecuted DAX advice completed. Real redacted captures are included.
Offline/direct checks are authored fixtures. Actual caller traces established generated-query execution; later Studio and M365 responses succeeded. All displayed values are not independently audited; the earlier combined request and deferred repeated-value case remain unresolved. Missing declared fields mean DAX blank/null under the explicit null-omission policy. The rendered picker shows GPT-5 Reasoning (Preview); inference telemetry remains unverified. Current warnings concern preview-model suitability and absence of formal Studio evaluation.
Production questions still matter
Does RLS automatically make this safe?
RLS follows the actual execution identity and workspace role. Verify restricted users, Build permission, and the model's sensitive fields. Agent instructions do not enforce authorization.
Can it answer any question?
No guarantee. The generic contract removes the five-metric mapping but still requires valid metadata, permissions, supported table-expression syntax and correct reasoning. Dated ranking and creator follow-up are user-observed working; the combined request and broader question coverage remain unverified. Exact measure implementations and automatic multi-model routing are not provided.
Is Fabric capacity or MCP required?
A Fabric data agent is not required for the connector/REST route. Power BI licensing still applies. Hosted Power BI MCP is a separate optional integration; its query-generation tool has Copilot entitlement requirements.
Are these real business results?
The M365 result screenshots are real, with owner-approved agent names, usage figures and dates unchanged. People's names are masked. Separately labeled illustrative tables use invented names and numbers; they are not test evidence. Original captures and tenant configuration remain private.
Download the unmanaged ZIP and import it through Power Apps or Copilot Studio Solutions. It creates a separate unpublished agent with no configured channels or bound Power BI connection.
2 / CONFIGURE
Ground it in your model
Bind an end-user connection, prepare authorized model metadata and review your metric definitions, relationships, dates and permissions. Demo report instructions and private model data are not included.
3 / VERIFY AND PUBLISH
Generate, test, then share
Generate and deploy the configured topics using the supplied source. Validate important answers independently, then publish and configure your M365 Copilot channel. The ZIP stops before queries until configured.
Reusable does not mean unlimited. One model alias is configured at a time. Missing business definitions, unsupported expressions and incorrect DAX filter context can still prevent a correct answer. Successful execution alone is not an accuracy guarantee.
People's names are masked in real result screenshots; agent names, usage figures and dates remain visible with owner approval. Configuration captures redact private resource IDs and identities. Illustrations are labeled synthetic. The reviewed ZIP excludes credentials, tenant bindings and the demo model. No open-source license has been selected; public visibility does not grant additional reuse rights.
