Analytics Insight framed this as a business intelligence strategy question. Its useful point is narrower and more practical than its "framework" headline suggests. Different teams often use the same business term for different measures, and semantic modeling is where those definitions get reconciled. Microsoft's release notes and documentation, Gartner's 2026 research and Tableau's August release all show vendors moving in that direction. Microsoft's own guidance is also frank about what goes wrong when the groundwork is skipped.
Power BI's June 2026 update puts semantic models at the centre
A semantic model is the layer between raw tables and the people or tools querying them. It holds the business logic. In Power BI, Microsoft describes it as including tables, columns, measures, relationships and calculation groups, plus properties such as descriptions, data types, format strings, data categories and linguistic metadata (synonyms and relationship verbs). A report on "active customers" draws on a measure defined in the model. When Copilot answers a question, it reads the model's schema to work out which fields to use.
Microsoft's update archive lists the June 2026 release as version 2.155.756.0. Four items in it bear on semantic modeling:
| Feature | Status in June 2026 | What Microsoft says it does |
|---|---|---|
| Fabric Apps for semantic models | Preview | An "AI-first" way to build operational data apps on top of semantic models, reusing the model's governance and business logic |
| Copilot in web modeling | Preview | Analyses model structure and applies schema changes (renames, relationships, DAX measures) from natural-language prompts |
| DAX user-defined functions (UDFs) | Generally available | Reusable functions with typed signatures, available in both Desktop and web modeling |
| DAX UDF optional parameters | Shipped with the June release | Parameters can have default expressions, so callers can leave some arguments out |
The same release made shape maps generally available and added a date picker slicer as a preview, along with other visual improvements. Those are useful for report authors, but they are separate from the semantic-model story. The two AI features that matter most here are still previews, so plan for them to change.
Microsoft has kept going in the same direction since June. The July 2026 update (version 2.156.951.0) added TMDL view on the web, a Model Options dialog in the Power BI service, and measure descriptions written as triple-slash comments in DAX query view. That last one matters because descriptions are among the properties Copilot reads.
Why "revenue" needs one definition before Copilot can answer
Analytics Insight lists the usual problem terms: revenue, active customer, churn, margin and forecast. Each can have several legitimate definitions. Finance may count revenue when it is recognised, while sales counts it when it is booked. One team's "active customer" might mean a login in the last 30 days, another's a paid invoice in the last quarter. Two dashboards that disagree are a nuisance. A chat assistant that quietly picks one definition is a bigger risk, because the person asking may never know a choice was made.
Microsoft's documentation shows how this plays out. For the "Ask data questions" experience, Copilot builds its context from report metadata on the current page, the current chat history, the model schema, the full linguistic schema, and properties such as descriptions and data types. Hidden fields, hidden report pages and tables marked private are left out. So the model's structure, names and descriptions directly shape what Copilot can see and how it reads a question.
Microsoft's own examples show the failures. Asked how many units sold in Australia, Copilot filtered on the Country column in the Customer table instead of the one in the Sales Region table. Asked for profit by country in 2024, it said there was no 2024 data. It then filtered on a customer Birthday column rather than the model's marked date table. The fixes Microsoft suggests are all modeling work:
- Hide or rename fields that aren't meant for analysis.
- Remove misleading synonyms using the Q&A setup.
- Stop using the same field names across different tables.
Microsoft states the conclusion without hedging. Its guidance says that without preparing the data, the semantic model and the users, Copilot mainly produces low-quality and inaccurate outputs that may be incorrect or even misleading. Its tutorial on preparing models for AI makes the same point: "Without proper preparation, Copilot can generate less helpful and inaccurate results." The documentation also notes that Copilot outputs are nondeterministic, meaning that it's possible for a user to receive a different output from a Copilot experience, despite using the same prompt and grounding data. A well-built model makes errors less likely. It cannot rule them out.
Gartner treats composite semantic layers as the 2026 direction
Gartner's February 17, 2026 research note, Top Trends in D&A for 2026: Making Composite Semantic Layers Interoperable, starts from the view that a single universal semantic layer is still out of reach. Gartner describes a shift toward composite semantic layers instead. These coordinate different semantic artifacts, such as BI models, data catalogues and metric definitions, to close context gaps, reduce analytical silos and keep business logic consistent across systems. This is Gartner describing a trend. It does not mean any particular cross-platform implementation has become standard.
A May 11, 2026 Gartner press release ties this directly to AI. Gartner argues that AI agents need context, including semantic representations of data, to understand an organisation's relationships and rules. It says a conventional schema alone doesn't capture business meaning, and advises data and analytics leaders to build a "context layer" into their infrastructure. Gartner predicts that by 2027, organisations that prioritise semantics in AI-ready data could increase agentic AI accuracy by up to 80% and cut costs by up to 60%. These are forecasts with "up to" ceilings, not measured results.
For a Power BI shop, the practical reading is modest. A well-governed Power BI semantic model is one semantic artifact. The composite-layer idea is about making several of them work together across platforms. Microsoft's Fabric Apps preview, which reuses a model's governance and logic in operational apps, and its Frontier-programme work bringing Power BI data answering into Microsoft 365 Copilot Chat, both extend how far a single model's definitions reach. Neither is a cross-vendor composite layer.
Copilot in web modeling edits models behind checkpoints and Write permission
Copilot in web modeling is the most operationally significant preview in the June release. It lets Copilot change a model, not just answer questions about it. Microsoft describes it as an AI-powered assistant integrated into the Power BI service semantic model experience. It helps model authors analyze and improve their semantic models using natural language. Microsoft's Power BI updates blog said it was rolling out this month in the Power BI service in June. It can flag inconsistent naming or unclear structure, then make schema changes such as renaming tables and columns, creating relationships, and generating DAX measures.
How a session works
- Open a semantic model in Model view in the Power BI service, switch to Edit Mode, and select Copilot from the ribbon.
- Enter a prompt or pick a suggested one. Before responding to the first prompt in a session, Copilot asks your permission to review the model or suggest changes.
- After you grant permission, Copilot creates a restore checkpoint at the start of the session using semantic model version history. You can use this checkpoint to revert the semantic model back to its original state before Copilot changes were applied in that specific chat session.
- Review each proposed change before accepting it. If a session goes wrong, roll back to the checkpoint.
The permission model follows existing Power BI roles. You can use Copilot to make changes only if you have the Write permission to edit the semantic model. Workspace Admins, Members, and Contributors all have Write access and can use Copilot to edit the active model. Copilot's editing access is limited to the model open in the current Copilot pane. It can read the schema of other models you can access in the same workspace, but those stay read-only. One documented limitation: you can't use web-modeling Copilot to set "prep data for AI."
Two operational details deserve admin attention. First, the checkpoint only covers changes made in that chat session. It doesn't replace your normal change control for models many reports depend on. Second, you can't switch Copilot on for one experience and off for another. Microsoft's guidance says that enabling Copilot in Fabric for a user or security group on a capacity turns on every Copilot workload and experience for those users. If you want to allow Copilot for report consumers without giving it editing reach, workspace roles are your control, not a feature toggle.
Licensing depends on the experience. Microsoft's semantic-model guidance says Copilot in Power BI Desktop requires a paid Fabric capacity of F2 or higher, or Power BI Premium P1 or higher. That capacity requirement is separate from whether the connected model sits in a Pro or Premium Per User workspace. Microsoft's Copilot overview points readers to experience-specific requirements pages, so check the requirement for the exact experience you plan to use. A Pragmatic Works walkthrough of the web-modeling preview reports that you must have access to the semantic model, publish to a Fabric capacity workspace, and ensure Copilot is enabled in your tenant, and that availability depends on tenant, region and admin settings.
DAX user-defined functions standardise logic, but someone still has to approve it
DAX UDFs reached general availability in June. They tackle a long-standing modeling headache: the same calculation copied, with small differences, across many measures. A UDF defines that logic once with a typed signature and can be called from anywhere. Optional parameters with default values mean callers only pass the arguments they need. Microsoft says UDFs now work across both Desktop and web modeling.
For governance, this is a real improvement. If the churn formula lives in one function, fixing it fixes every measure that uses it. That also makes it easier for a reviewer to confirm the whole model uses a single definition.
A function makes a definition consistent. It doesn't make it correct. Whether "churn" means cancelled subscriptions or lapsed usage is still a business decision, and a UDF will apply the wrong answer just as consistently as the right one. The same applies when Copilot writes measures. Microsoft warns that inexperienced developers should validate and understand any DAX Copilot generates before adding it to a model. Copilot doesn't know where a measure will be used, and code that works in one query can give wrong results in a different report filter context.
Tableau Next's Semantic Model Builder shows Salesforce making the same bet
Microsoft isn't alone. Tableau's official August 2026 release notes say the Semantic Model Builder (Smart Canvas) in Tableau Next is now generally available. It offers auto-layout, infinite navigation, inline editing and domain-based grouping. Analytics Insight adds that the release includes AI tools for model creation and calibrating question-and-answer results. Tableau's own page labels Tableau Agent data connection and analytics creation as beta, so those AI pieces are less mature than the modeling canvas.
The comparison shows where vendors are investing. It doesn't mean the two platforms are equivalent. Power BI's June AI features for models are previews built on an established semantic model format. Tableau Next's builder is a generally available authoring canvas with beta agent features around it. For organisations running both, this is where Gartner's composite-layer argument becomes practical: two authoring tools can easily end up with two definitions of "revenue."
What this means for Power BI teams: fix the definitions, then enable the assistant
Finish the definition work before you roll out AI features to business users. Teams whose users already rely on Copilot for questions have the most urgent job. They should test their models the way Microsoft recommends, using the "Ask data questions" experience with realistic prompts, before assuming the answers are trustworthy. Teams still on report-only use can take their time. Microsoft notes that if users won't use Copilot, developers don't need to test for it. Microsoft also suggests that if a complex model can't give consistently reliable answers, the right call may be to advise users not to use Copilot with it.
For model authors, Microsoft's guidance comes down to a short checklist:
- Get the business to sign off on definitions for your highest-impact metrics (revenue, churn, active customer, margin) before encoding them, and put the approved logic in UDFs so each has one source.
- Use human-readable names and add descriptions for tables, columns and measures. Hide helper fields and avoid identical field names across tables, because Copilot reads names and descriptions and skips hidden fields.
- Set up linguistic modeling (synonyms and relationship verbs) through the Q&A tools, and remove synonyms that point Copilot at the wrong column.
- Treat every change from Copilot in web modeling as a change to review. Don't clear chat sessions you may need to roll back until you've confirmed the result, and keep version control on shared models.
- Use workspace roles to decide who can have Copilot edit models. Admins, Members and Contributors all have Write access, and there's no separate switch for the editing experience.
- Check the specific capacity and licence requirement for each Copilot experience you enable. The F2/P1 threshold Microsoft documents applies to Copilot in Power BI Desktop, not to every Copilot surface.
Microsoft's June 2026 release gives Power BI teams faster ways to build semantic models: Copilot that can restructure a model on request, functions that centralise logic, and apps that reuse a model's governance. Microsoft's own documentation still says Copilot mainly produces poor results when the model and its definitions haven't been prepared. Gartner reaches the same conclusion at an industry level. With Copilot in web modeling and Fabric Apps still in preview, and Microsoft shipping model-authoring updates every month, now is a good time to agree on and write down metric definitions, before these features become generally available and more users start relying on them.