Microsoft’s July 21, 2026 announcement about the future of conversational analytics in Fabric is not a product update. It is the platform direction that redefines what your Power BI semantic model is for. This is the deep read for architects, CTOs, CAIOs, and the analytics leaders who will have to build the operating model underneath it.
What Microsoft Actually Announced on July 21, 2026
On July 21, Microsoft published a post on the Fabric Updates Blog titled “The future of conversational analytics in Fabric.” It described, in Microsoft’s own words, a broader vision: trusted business context in Fabric IQ, consumable consistently across Microsoft 365, Fabric, developer experiences, and future AI-powered applications.
Strip the launch language away and the announcement says four things.
First, Power BI semantic models — the business logic your organization has spent years building — are now the grounding layer for AI answers across the entire Microsoft ecosystem, not just inside Power BI.
Second, two experiences are live today: Microsoft 365 Copilot Chat grounded in your semantic models, and Microsoft 365 Copilot Cowork, where a grounded insight becomes a drafted email, a composed document, and a scheduled meeting inside one conversation.
Third, Fabric Data Agents are the specialization tier — configurable virtual analysts connected natively to lakehouses, semantic models, mirrored databases, ontologies, and OneLake shortcuts, with evaluation support built in, publishable directly into Microsoft 365 Copilot.
Fourth — the one that generated the most questions — the full-screen Copilot experience inside Power BI, and Report Copilot with it, will evolve into the unified Microsoft 365 Copilot experience over the coming months. One Copilot. Every surface. No context switching.
Every one of those four statements has architectural consequences that the announcement did not spell out. This page spells them out.
Fabric IQ Is Not a Product. It Is a Contract.
The most useful way to understand Fabric IQ is not as a feature you turn on but as a contract Microsoft has now written between your data estate and every AI agent that will ever query it.
The contract says: if your organization encodes its business logic in Power BI semantic models — definitions, measures, relationships, security, sensitivity — then every AI surface Microsoft ships will honor that encoding when it answers questions.
That contract is enormously valuable, and it is entirely one-sided in an important way: Microsoft’s side of the contract is automatic. Your side is not. Copilot will faithfully ground on whatever your semantic model says. If the model says the wrong thing — an implicit measure, an undocumented column, a stale definition of revenue, an untested security role — Copilot will faithfully deliver the wrong thing, fluently, to an executive in Teams who has no reason to doubt it.
The correctness problem was never the conversation. It was the trust surface underneath it. Microsoft has now named the trust surface, and it is yours to govern.
For the CFO-level version of this argument, our analysis of the conversation nearly half of finance leaders have never had covers what happens when AI investment outruns governance. This page covers the technical floor underneath that conversation.
The Three Surfaces at Implementation Depth
Microsoft 365 Copilot Chat with Fabric IQ. The user asks a question in Teams or Outlook. Copilot resolves it against the certified semantic models that user can access. Row-level security honors their identity. Object-level security constrains readable columns. Sensitivity labels travel with the answer. The design detail that matters to architects: the answer is grounded in the model, not the report. Fabric IQ skips the visualization layer and queries the model directly — which means the model must be self-sufficient, with measures that produce defensible numbers outside the filter context of any specific report page. Most production models today are not self-sufficient. That is the readiness gap.
Microsoft 365 Copilot Cowork with Fabric IQ. Cowork is not a chat interface; it is an orchestration surface. The grounded answer flows immediately into action — the drafted email, the composed deck, the scheduled meeting, the filed follow-up. The governance implication: the velocity of wrongness increases by an order of magnitude. A wrong dashboard number waits for an analyst to read it. A wrong Cowork answer becomes an email that was already sent. Grounding validation has to be built into the loop before the loop is turned on. Our Microsoft Copilot practice exists precisely for this readiness sequencing.
Fabric Data Agents. When general-purpose grounding is not enough, you configure a Data Agent: a virtual analyst with a defined scope, curated context, and native connections across every layer of Fabric — lakehouses, semantic models, mirrored databases, OneLake shortcuts, graph data, and ontologies. Three elements separate a Data Agent from a chatbot. It supports evaluations, so answer quality is a measured number, not a vibe. It inherits Fabric’s identity, security, and sensitivity posture natively. And it publishes into Microsoft 365 Copilot, so the specialization surfaces where users already work. Our Microsoft Fabric consulting practice designs, builds, and evaluates these agents against a scored rubric.
What Happens to Copilot in Power BI Between Now and 2027
The question every Power BI team is asking: is Copilot in Power BI going away?
No — but its chrome is changing. The full-screen Copilot pane inside Power BI evolves into an embedded surfacing of the Microsoft 365 Copilot experience. Report Copilot, generally available today, evolves in the same direction. Under the hood, the same M365 Copilot that answers questions in Teams answers them inside Power BI. Microsoft’s stated goal is the removal of context switching: one Copilot experience for insights, wherever users work.
What does not change is where your control lives. Prep for AI instructions — the annotations you author on the semantic model to shape how Copilot reasons over it — remain the control surface. They are authored on the model, version-controlled with the model, and they now influence Copilot everywhere it grounds against that model. If your team wrote Prep for AI instructions for Report Copilot, that investment just got promoted to tenant-wide scope. If your team never wrote them, that is the single highest-leverage authoring task available to you this quarter.
The Licensing Math Nobody Is Doing
Two meters run side by side, and most deployment plans only budget for one.
The Microsoft 365 side: Copilot Chat and Cowork experiences that surface Fabric IQ ship through the Frontier tier of Microsoft 365 Copilot licensing — per-user pricing, and notably, the Fabric IQ grounding invocation itself does not consume additional Fabric capacity from the M365 side.
The Fabric side: everything Fabric IQ grounds on still requires Fabric capacity. Semantic models in Fabric workspaces, lakehouses, mirrored databases, OneLake, and above all Fabric Data Agents — all metered against F-SKU capacity. Undersized capacity produces silent throttling; silent throttling produces intermittent AI failures; intermittent failures produce the one thing you cannot rebuild cheaply, which is executive trust.
Before any pilot, model both meters. Our current Power BI cost and licensing guide carries the current Pro, PPU, and F-SKU pricing, and our team sizes capacity against measured workload rather than vendor rules of thumb.
The Ten Things Every Fortune 500 Semantic Layer Is Missing
In twenty-nine years of auditing enterprise Microsoft environments, EPC Group has never once walked into a Fortune 500 tenant whose semantic layer was ready for what Fabric IQ asks of it. These are the ten gaps we find, in the order we find them.
- Implicit measures everywhere. Default aggregations with no authored DAX and no defensible definition. Copilot grounding on implicit measures produces plausible answers that are silently wrong. Every trusted number needs an explicit, documented, tested measure.
- No model-layer metadata. Tables, columns, measures, and relationships without human-readable descriptions in TMDL, version-controlled in PBIP. Copilot reads your metadata. Blank metadata means the model is guessing about itself.
- Orphaned workspace ownership. Forty to two hundred workspaces per tenant with no living, named owner. An AI grounding on an unowned workspace is answering on behalf of an artifact nobody stewards.
- Missing lineage. Datasets with no documented path back to source. Fine for consumption; fatal for audit.
- RLS never tested against AI personas. Row-level security designed for dashboard audiences, never validated against the identities that will now ask questions through Copilot Chat.
- OLS applied at the visual layer instead of the model layer. The AI is not bounded by your report design. Columns hidden in a visual are not hidden from an agent. Object-level security must live in the model.
- Sensitivity labels applied inconsistently. Labels on some artifacts, absent on others, unaligned with Purview policy. The label is the only mechanism that travels with the answer. Our Microsoft Purview practice closes this gap.
- No Prep for AI instructions. The single control surface for shaping AI reasoning over your model — unauthored in the overwhelming majority of tenants.
- No evaluation baseline. Data Agents ship with evaluation harnesses. Almost nobody defines the scored rubric that makes the evaluation meaningful.
- No certified inventory. No single artifact that says: this model is approved for grounding, by this agent, for this audience, under this retention policy, sponsored by this human, recertified on this date.
Organizations at zero on this list ship Fabric IQ into production with defensible answers. Organizations that skip it ship a very confident tour guide through an ungoverned estate. Closing all ten is the core of our semantic model certification work inside the Power BI consulting practice.
The Governance Operating Rhythm
Ten certified models today become fifty in six months. The certification decays without an operating rhythm, and the rhythm needs a named executive.
Monthly: certified model inventory review — what was certified, what is due for recertification, which Prep for AI instructions changed, which RLS matrices were retested.
Quarterly: the audit committee brief — the board-visible artifact documenting what your AI is grounded on, who authorized each grounding, the evaluation-scored answer-quality trend, and what changed.
Annually: the strategic review — which models retire, which Data Agents evolve, which policies renew.
This is an executive cadence, not a technology cadence. For enterprises not ready to hire a full-time Chief AI Officer, our Virtual Chief AI Officer practice provides the named, accountable, fractional executive who owns this rhythm — and our AI Governance practice stands up the Center of Excellence that keeps it running after the consultants leave.
Why Your Cyber Renewal and Fabric IQ Are the Same Conversation
Underwriters writing AI, cyber, and E&O coverage are converging on a short list of questions: Can you demonstrate every AI answer is grounded in a defensible source? Can you produce the approval matrix? Can you show measured answer quality? Is there a named accountable executive?
Every one of those questions is answered natively by the architecture described on this page. The certified semantic model is the defensible source. The certified inventory is the matrix. The Data Agent evaluation harness is the quality evidence. The vCAIO is the named executive. Organizations that build the Fabric IQ operating model correctly produce their AI insurance readiness evidence as a byproduct — not as a separate project.
The 30-Day Fabric IQ Readiness Plan
Days 1–7: Inventory. Every workspace, every semantic model, every measure the business trusts, every owner, every label, every gap against the ten-item list above.
Days 8–14: Prioritize. Not every model deserves grounding. Select the top ten candidates by decision value — the models whose answers executives will actually act on in Teams.
Days 15–24: Certify the first three. Explicit measures, TMDL metadata, RLS/OLS testing against AI personas, sensitivity alignment, Prep for AI instructions authored and version-controlled.
Days 25–30: Pilot and measure. Enable M365 Copilot Chat against the certified three. Log every question and answer. Draft the grounding matrix. Write the board memo.
EPC Group delivers this as a fixed-fee, senior-led, thirty-day engagement — the same named principals at day thirty that you met at day one. It is the natural first step before any Copilot Cowork rollout or Data Agent build.
Frequently Asked Questions
Is Copilot in Power BI going away?
No. It is evolving into the unified Microsoft 365 Copilot experience surfaced inside Power BI over the coming months. Report Copilot evolves in the same direction. Your semantic-model-side investments — including Prep for AI instructions — carry forward.
Do I need additional Fabric capacity to use Fabric IQ in M365 Copilot?
The M365 Copilot Chat and Cowork experiences ground through Fabric IQ without additional Fabric capacity consumption on the M365 side. The underlying Fabric estate — semantic models in Fabric workspaces, lakehouses, and especially Fabric Data Agents — still requires appropriately sized F-SKU capacity.
What is the difference between Fabric IQ and Work IQ?
Fabric IQ carries your trusted data context — semantic models, definitions, measures. Work IQ carries your work context — emails, chats, meetings, documents. In Microsoft 365, answers blend both: the certified number and the surrounding organizational context, together.
What are Fabric Data Agents and when do I need one?
Configurable virtual analysts connected natively across Fabric — lakehouses, semantic models, mirrored databases, OneLake shortcuts, ontologies — with evaluation support and publishing into M365 Copilot. Build one when a business domain needs deeper, specialized analysis than general-purpose grounding provides.
What does "approved for grounding" mean?
A semantic model that has passed certification: explicit measures, complete metadata, named ownership, documented lineage, tested security, aligned sensitivity labels, authored Prep for AI instructions, an evaluation baseline, and an entry in the certified inventory matrix with a human sponsor and a recertification date.
How do Prep for AI instructions work?
They are lightweight instructions authored directly on the semantic model that shape how Copilot reasons over it — terminology, preferred measures, business context. They are the model-side control surface, and they now influence the unified M365 Copilot experience everywhere it grounds against your model.
What Microsoft 365 licensing tier includes these experiences?
The Fabric IQ experiences in M365 Copilot Chat and Cowork ship through the Frontier tier of Microsoft 365 Copilot licensing. Confirm current tier details against your enterprise agreement — and model both the per-user and the F-SKU meters before piloting.
How fast can we be ready?
Thirty days to a certified first wave and a defensible grounding posture, following the plan above. Full-estate certification depends on model count and current hygiene — typically 90 to 180 days for a Fortune 500 tenant.
Does EPC Group work alongside our existing integrator?
Yes. Our AI Deployment Second Opinion model reviews and validates work in progress without requiring you to replace anyone. Many engagements start exactly there.
Who leads the engagement?
Named senior principals — the same people at week one and week thirty. EPC Group is a Microsoft Solutions Partner holding all six designations, G2 Leader in BI consulting for six consecutive quarters, founded and led by Errin O’Connor, four-time Microsoft Press bestselling author and original member of both the Project Tahoe SharePoint beta and Project Crescent Power BI beta programs.
Related EPC Group Practices
- Power BI Consulting — semantic model certification for AI grounding
- Microsoft Fabric Consulting — OneLake, capacity, Data Agents
- Microsoft Copilot Readiness — rollout sequencing and approval matrices
- AI Governance & Center of Excellence
- Virtual Chief AI Officer
- AI Insurance Readiness
- Microsoft Purview
- Microsoft Frontier Company — this page’s parent topic
- Power BI Cost & Licensing Guide
Primary sources: Microsoft’s announcement · The future of Copilot in Power BI · Fabric IQ overview on Microsoft Learn · Fabric Data Agents
Start a Fabric IQ Readiness Conversation
Thirty days to a certified first wave and a defensible grounding posture. Tell us about your estate and we will scope the inventory.
Contact EPC Group: contact@epcgroup.net · (888) 381-9725