The Executive Insight
The marketing story for Copilot in Fabric is simple:
“Ask questions in natural language, get instant insights.”
The operational story is not.
Under the covers, Copilot and Fabric IQ agents are not one magical brain. They are an AI workforce sitting on top of your models, domains, and OneLake:
- They consume capacity every time they generate or refine an answer.
- They inherit every flaw in your semantics and governance.
- They can amplify oversharing and leakage if Purview boundaries are weak.
Issue #15 is about treating Copilot not as “free insight,” but as staff you have to onboard, supervise, and budget for.
What Changed This Year: Copilot Capacity Became a First-Class Bill
Two quiet but important shifts landed over the last year:
- Copilot in Fabric is now broadly GA across workloads (Power BI, Data Engineering, Data Science, Real-Time, etc.).
- Fabric Copilot Capacity (FCC) became available from F2 upwards and can be designated per capacity.
The implications:
- When users invoke Copilot or Data Agents, their tokens and compute don’t just “come out of nowhere.” They’re charged against a Copilot capacity or, if not configured, your existing Fabric capacity.
- A new tenant setting, “Capacities can be designated as Fabric Copilot capacities”, allows capacity admins to flip capacities into FCC without tenant admin involvement.
- If you don’t know who your capacity admins are, you can wake up to AI usage accruing on capacities that were never sized—or budgeted—for it.
In other words: Copilot is now economically real. You don’t just decide if you’ll have AI. You decide where it runs and who pays for it.
The Governance Reality: Anyone Can “Approve for Copilot” (Unless You Fix It)
There’s another less obvious governance gap practitioners have already raised:
- Today, any semantic model owner can flag their model as “approved for Copilot.”
- There is no global review gate for “Copilot‑ready” models; it behaves a bit like the old “promoted content” pattern in Power BI.
That leads to a predictable pattern:
- Enthusiastic model owners mark their models as Copilot‑ready.
- Copilot happily reasons over those models—even if definitions are half‑baked or security is loosely thought through.
- Business users and AI agents start to treat those models as trusted truth because “Copilot can see them.”
The articles calling this out are blunt: admins need more control over who can make a model Copilot‑visible, or you’re back to crowdsourced endorsements in front of an AI audience.
Until Microsoft changes that behavior, the only real mitigation is governance discipline:
- Clear rules about which domains/models are allowed to be Copilot‑enabled.
- Admin review of Copilot usage and approvals.
Copilot doesn’t fix your modeling and governance debt. It makes it visible—and billable.
Fabric IQ Agents: Your First Real AI Colleagues
While Copilot is the chatty front‑end, Fabric IQ and Data Agents are the deeper shift:
- Fabric IQ introduces an ontology and semantic graph that define your business entities, relationships, and rules.
- Data Agents and Operations Agents use that ontology as a mental model of your business: customers, orders, inventory, routes, risks.
- These agents don’t just answer questions. They can monitor data, reason over it, and trigger actions—like an analyst embedded in your data estate.
Think of them as:
A team of junior analysts and ops coordinators who:
- Know your ontology.
- Run queries, synthesize signals, and kick off workflows.
- Never sleep and always bill their time against your AI capacity.
This is powerful:
- You can build domain‑specific agents for finance, operations, supply chain, customer success.
- Agents can combine event data, warehouse facts, and ontology rules to take actions (reroute trucks, flag risky orders, escalate anomalies).
It’s also exactly why AI governance and data governance are now indistinguishable conversations.
AI Safety and Trust: Purview Is the Real Guardrail
Microsoft’s own messaging is consistent:
“AI use is only safe if it’s secure.”
Recent Purview and Fabric announcements add concrete pieces to that:
- Purview Security Posture Management for AI: find overshared Fabric data (reports, dashboards, datasets) that are likely risk points for AI‑driven leakage.
- Data Risk Assessments that scan your top Fabric workspaces for oversharing, mislabeled sensitivity, or AI‑risky access patterns.
- Integration between Purview classifications/labels and Fabric items surfaced to Copilot and agents.
The point is not “AI is dangerous, turn it off.”
The point is:
- AI experiences (Copilot, agents, IQ) should only see what they’re allowed to see.
- Your existing RLS, sensitivity labels, DLP, and domain boundaries must apply consistently, even when access is via a prompt, not a report.
When you align Fabric + Purview, AI becomes a governed access pattern, not a new shadow channel.
Hope for Lean Teams: Start with One “AI-Safe” Domain
This can sound impossibly big if you’re a lean SMB team or a mid‑market partner. It doesn’t have to be.
You don’t need tenant‑wide Copilot on day one. You need:
- One domain where your semantics and governance are strong (e.g., core revenue KPIs, or a well‑modeled ops domain).
- One AI surface (Copilot on a certified semantic model, or a Fabric Data Agent bound to a clean ontology slice).
- One capacity plan that says, “AI for this domain runs here, and we can see its cost and usage.”
Then you:
- Validate that answers are trustworthy.
- Confirm that access stays within policy (Purview + audit).
- Measure whether AI adds decision velocity (time‑to‑answer, time‑to‑action).
Only then expand.
You’re not rolling out “AI for everything.” You’re staffing one AI team member into one well‑run department, and watching how they perform.
Where I Fit In (For Partners and Leaders)
If you’ve been following this series for long enough, then you must be aware of the pattern:
- AI on Fabric is not a separate initiative.
- It is the combined output of your semantics, governance, capacity, domains, and discovery.
Most organizations are being asked to “do something with AI in Fabric” without:
- A clear view of what Copilot can safely see and say.
- A handle on where AI capacity lives and who pays for it.
- A plan for which domain is mature enough to deserve agents.
I work with:
- Partners who need an AI‑on‑Fabric story that goes beyond “look, it writes DAX.”
- CIOs, CDOs, and CTOs who want to say “yes” to Copilot and Fabric IQ—but only where it’s safe, governed, and economically grounded.
- Operators in data‑sensitive industries (healthcare, energy, F&B) where AI mistakes are more than cosmetic.
My focus is simple to describe and hard to fake:
Make sure every AI experience in Fabric is grounded in trusted data, constrained by governance, and visible on the cost ledger.
Isaac Truong | Founder, Allston Yale
Enterprise-grade analytics for $50M–$100M SMBs
Power BI | Fabric | Azure | Data Strategy
? Book a 20-min Fabric diagnostic →
? Subscribe to get Friday Fabric Facts in your inbox (plus early access to templates) ?
LinkedIn: Connect with me for daily Fabric tips
Friday Fabric Facts #15: Originally Posted on LinkedIn, May 8, 2026