There are two wrong stories about generative AI in analytics. One says the dashboard is dead and executives will simply ask a model. The other says it is a demo feature that changes nothing. Having watched both Copilot in Power BI and Fabric data agents work against real semantic models, I think the truth is more useful than either: AI changes who can ask and how fast they get an answer, and it makes the quality of the underlying model the single determinant of whether the answer is right.
What the tools do today
Copilot in Power BI offers generative capabilities to three audiences: business users who ask questions and get summaries of reports, report authors who get help creating pages and visuals, and semantic-model authors who get help writing DAX and descriptions. It has requirements (a capacity, admin enablement) and prompt limits documented by Microsoft [1].
Fabric data agents let people ask plain-language questions about data in OneLake (lakehouses, warehouses, semantic models and KQL databases) and receive answers; the agent translates the question into queries against the sources you configure, with instructions and example queries you supply to steer it, and it is subject to prerequisites, governance settings and limitations that are worth reading in full before promising anything to a leadership team [2].
Both are, at bottom, natural-language front ends to a semantic model. They do not know what revenue is. They read the measure called Revenue, its description, its relationships and its synonyms, and they write a query. If that measure is wrong, ambiguous or missing, the answer will be fluent and wrong.
The prerequisite nobody wants to hear
Every AI-in-BI feature is downstream of the semantic layer. The things that make a model good for humans (a star schema, business-readable names, one definition per measure, descriptions, a marked date table, row-level security) are exactly the things the model needs to answer correctly [3]. In practice this means:
- Names are prompts.
Net Revenue (USD)with a description “Invoiced amount net of returns and discounts, excluding tax” is answerable.Amt2is not. - Synonyms are vocabulary. If sales says “bookings” and finance says “orders,” the model needs both.
- Security is the boundary. AI answers inherit the user’s row-level security; a model without RLS answers every question for everyone [4].
- Ambiguity is a defect. Two measures that could both be “margin” mean the agent guesses. Retire one or rename both.
Organizations that invested in dimensional modeling and a governed semantic model are ready for this now. Organizations that report from spreadsheets and ad-hoc SQL have a modeling project ahead of them, not an AI project.
What changes in data management
| Area | Before | With generative AI in the loop |
|---|---|---|
| Metadata | Nice to have | Load-bearing: descriptions, synonyms and sample questions are part of the model’s definition of done |
| Data quality | Found in reconciliation meetings | Found by anyone asking any question; tests at the grain become mandatory [5] |
| Governance | Who can see which report | Who can ask which question of which data; agents must be scoped and logged [6] |
| Lineage | Documentation | Explanation: “where did this number come from” is now a routine user question |
| Freshness | Refresh schedule | Stated on every answer, or users will assume “now” |
The DAMA Data Management Body of Knowledge has described these functions for years: metadata, quality, governance, security [7]. AI did not invent them. It removed the option of neglecting them, because the consequence of a bad definition is no longer a wrong chart that an analyst catches, but a confident sentence delivered straight to a decision maker.
What changes for the BI analyst
Less time building the fortieth variant of a bar chart; more time on the three things that AI cannot do for you. Defining: deciding what revenue, churn and on-time mean, and encoding it once. Verifying: building the reconciliation and tests that prove the model, and auditing what the agent said against what the data shows. Deciding what to measure: the balanced scorecards, early-warning signals and targets that come from understanding the business, not the tool. Analysts who move toward modeling and verification become more valuable; analysts who only build visuals compete with a prompt.
Where it genuinely helps now
- Drafting DAX and descriptions for a well-named model, then reviewing them.
- Narrative summaries of a report page for people who will not open the report.
- Long-tail questions that would never justify a report page (“how many Georgetown shipments over 20 CBM were delayed by weather in August”) against a governed model.
- Generating the first draft of tests, documentation and data contracts from a model’s metadata.
Where it does not, yet
- Deciding which of two plausible definitions the business means. That is a governance decision.
- Numbers that must be audited. A regulatory filing comes from a paginated report with a fixed query, not from a conversation (see SSRS is not dead).
- Models with poor metadata or no security. The agent will answer anyway, which is worse than not answering.
A 90-day readiness plan
- Days 1–30: inventory and clean one model. Pick the model executives use most. Rename, describe and add synonyms to every visible table and measure; retire duplicates; confirm RLS; write ten sample questions with known answers.
- Days 31–60: pilot with a scorecard. Enable Copilot or a data agent for a small group against that model. Score every answer against the known-answer set and log the misses; most misses are metadata, not AI.
- Days 61–90: govern and extend. Write the usage policy (what questions are in scope, how answers are verified, what is never asked of an agent), add tests at the grain to the pipeline, and extend the metadata work to the next model.
The organizations that get value from AI in analytics will be the ones that did the unglamorous modeling and governance work first. That has been true of every BI technology for thirty years. This one just makes the penalty for skipping it arrive faster.
References
- Microsoft Learn, “Copilot for Power BI overview,” learn.microsoft.com/en-us/power-bi/create-reports/copilot-introduction (accessed 9 Oct 2026).
- Microsoft Learn, “Fabric data agent concepts,” learn.microsoft.com/en-us/fabric/data-science/concept-data-agent (accessed 9 Oct 2026).
- Microsoft Learn, “Understand star schema and the importance for Power BI,” learn.microsoft.com/en-us/power-bi/guidance/star-schema (accessed 9 Oct 2026).
- Microsoft Learn, “Row-level security (RLS) with Power BI,” learn.microsoft.com/en-us/fabric/security/service-admin-row-level-security (accessed 9 Oct 2026).
- dbt Labs, “Add data tests to your DAG,” docs.getdbt.com/docs/build/data-tests (accessed 9 Oct 2026).
- Microsoft Learn, “Governance and compliance in Microsoft Fabric,” learn.microsoft.com/en-us/fabric/governance/governance-compliance-overview (accessed 9 Oct 2026).
- DAMA International, “DAMA Data Management Body of Knowledge (DAMA-DMBOK),” dama.org/cpages/body-of-knowledge (accessed 9 Oct 2026).
