Before AI agents act on your Microsoft Fabric data, six questions decide whether the context they reason on can be trusted: release status, refresh model, staleness tolerance, reach, agent permissions, and ownership. The same six apply whether that context lives in Fabric IQ’s ontology, in Atlan’s cross-platform context layer, or in a semantic model your own team maintains. According to Gartner’s May 2026 projection (Gartner, 2026), 40% of enterprises will demote or decommission autonomous AI agents by 2027 because governance gaps surface only after production incidents.
Most of AI agent governance on Fabric reduces to governing that context, and Microsoft documents more of it than buyers tend to expect.
- Answered on Microsoft Learn: release status, refresh mechanics, data reach, and agent identity.
- Answered only by your team: acceptable staleness, refresh ownership, and each agent’s autonomy.
- Worth asking of any layer: whether outside definitions are reconciled, and whether you can reconstruct what an agent saw.
| Field | What Microsoft’s documentation says (September 28, 2026) |
|---|---|
| What it is | Fabric IQ workload with two core items: ontology (preview) and Power BI semantic model (Microsoft Learn) |
| Microsoft’s framing | A shared context layer Fabric agents can consume (Microsoft Learn) |
| Release status | Ontology item: preview (Microsoft Learn); Fabric IQ GA announced at Build 2026 (Microsoft Fabric blog) |
| Freshness | Upstream changes appear after a manual or scheduled full refresh (Microsoft Learn) |
| Reach | OneLake, including external data referenced in place via shortcuts (Microsoft Learn) |
| Agent permissions | Own Entra Agent ID, creator’s delegated permissions (Microsoft Learn) |
What is Fabric IQ’s ontology, and how does it give AI agents context?
Fabric IQ’s ontology is the Microsoft Fabric item that turns business concepts such as Customer or Shipment into a shared vocabulary people and AI agents can reason over. According to Microsoft’s Fabric IQ overview (Microsoft Learn, 2026), the IQ workload delivers that context through two core items, the ontology (preview) and the Power BI semantic model, over business data in OneLake.
Microsoft’s ontology documentation (Microsoft Learn, 2026) calls the item “a shared context layer that can be consumed by Fabric agents,” binding entity types, properties, and relationships to OneLake data and answering natural-language questions through NL2Ontology.
The bootstrap path is a real advantage for Microsoft-first teams: ontologies can be generated from Power BI semantic models already in production, so certified logic starts the agent context layer instead of a rewrite. “Fabric IQ End-to-End: The Complete Enterprise AI Platform” is among Microsoft’s FabCon Europe 2026 session picks (Microsoft Fabric blog, 2026), sorted alongside the OneLake security talks in the FabCon Europe 2026 sessions picked by role.
With the platform describing its ontology as a context layer for AI agents, the useful questions move from whether agents need one to how that layer is governed.
Is Fabric IQ’s ontology production-ready, or still in preview?
As of September 28, 2026, Fabric IQ’s ontology is in preview. The ontology overview (Microsoft Learn, 2026) carries a preview notice, and Fabric’s What’s New page (Microsoft Learn, 2026) lists it among preview features. The workload is further along: Microsoft’s FabCon session-picks post (Microsoft Fabric blog, 2026) says Fabric IQ’s general availability was announced at Microsoft Build 2026, though Learn still labels the IQ workload preview. Graph and operations agents, two items the ontology works with, reached general availability in June 2026 per Microsoft’s What’s New archive (Microsoft Learn, 2026).
| Workload item | Status in Microsoft’s documentation (September 28, 2026) | Source |
|---|---|---|
| Fabric IQ workload | GA announced at Build 2026; still labeled preview on Microsoft Learn | Microsoft Fabric blog, What’s New |
| Ontology | Preview | Ontology overview |
| Power BI semantic model | No preview label | Fabric IQ overview |
| Graph | Generally available (June 2026); graph-powered AI reasoning is in preview | What’s New archive, Graph overview |
| Operations agent | Generally available (June 2026); newer features such as Investigator insights in preview | What’s New archive, What’s New |
A generally available item can still carry preview features, so check release notes for the specific capability each agent will use. Status moves each release, so date-stamp any readiness decision, the first line of the AI-ready data audit an AI readiness assessment formalizes.
How does the ontology stay current, and what does that mean for an agent’s context?
Upstream changes reach Fabric IQ’s ontology on triggers, not continuously. Microsoft’s ontology overview (Microsoft Learn, 2026) makes the point twice, once for data bindings and once for the graph: new rows and other upstream updates must be refreshed before they appear in the ontology item.
Microsoft’s refresh guidance (Microsoft Learn, 2026) adds that schema edits refresh downstream experiences automatically, while upstream record changes can leave the graph stale until an ingestion runs, triggered by a Refresh now action or a recurring schedule. Because each refresh is a full re-ingest with cost implications, Microsoft recommends batching updates; its graph refresh documentation (Microsoft Learn, 2026) lists schedules from every minute to monthly.
That makes freshness a design decision with a budget: daily lets an agent reason on data up to a day old, per-minute narrows that gap and multiplies cost, and no schedule leaves context as current as the last manual refresh.
So ask what staleness each agent can tolerate, who owns the schedule, and whether the agent can tell how old its context is. Mature answers are still rare: per Deloitte’s State of AI in the Enterprise (Deloitte, 2026), only one in five companies has a mature governance model for autonomous AI agents.
Context freshness and context drift name the failure: an agent answering confidently from context that used to be true. Every context layer should state its refresh model in writing.
What’s the most common mistake enterprises make in AI agent governance?
The most common mistake is treating agent governance as one switch, fully locked down or fully trusted, instead of scoping it to what each agent may see and do. Gartner senior director analyst Shiva Varma names that binary approach as the root cause behind the firm’s 40% projection (Gartner, 2026; reported by Enterprise DNA, 2026).
Scoping starts with a concrete question: whose permissions does each agent act with? Per Microsoft’s operations agent guide (Microsoft Learn, 2026), each Fabric operations agent gets its own Microsoft Entra Agent ID, runs in delegated mode with its creator’s permissions, logs its evaluations, and executes approved recommendations with the creator’s access, the core of AI agent identity and a start on scoping past Gartner’s all-or-nothing pattern. What remains is matching scope to task: does your AI agent access control model let you narrow that access where the creator’s own doesn’t fit?
Data governance vs AI governance marks where scoping belongs: one decides who sees a table, the other what an agent does with it, and that holds only if the context carries the policy.
What’s the difference between a semantic layer and a context layer?
A semantic layer defines how metrics are calculated for queries and reports. A context layer carries meaning, lineage, ownership, quality, and policy across the estate so an agent can decide whether to trust what it retrieves. Fabric IQ pairs the two: the Power BI semantic model supplies measures, and the ontology adds entities and relationships (Microsoft Learn, 2026).
| Aspect | Semantic layer | Context layer |
|---|---|---|
| Primary job | Calculates metrics consistently | Tells an agent what data means, where it came from, who owns it, and whether to trust it |
| In Fabric | Power BI semantic model | Fabric IQ’s ontology, scoped to OneLake |
| Typical scope | One BI platform or modeling tool | Every system an agent reads, when built cross-platform |
| Question it answers | How is net revenue calculated? | Which definition applies, is it current, and who certified it? |
Two independent analysts raise cautions in the same InfoWorld piece (InfoWorld, 2025): Moor Insights’ Robert Kramer says the more logic built into Fabric IQ’s semantic layer, the harder it becomes to move; HFS Research’s Suhas AR flags hurdles for organizations not yet invested in Fabric.
Neither caution is unique to Microsoft: any semantic layer, a dbt semantic layer included, concentrates logic in one place. Ontology vs semantic layer and context graph vs ontology cover why agents need the relationships around a definition. The semantic layer answers what a number means; the context layer answers whether an agent should act on it.
Can an ontology built on OneLake reach data and agents outside Microsoft’s stack?
Fabric IQ reaches external data through OneLake, which its overview (Microsoft Learn, 2026) describes as a multicloud data lake unified through shortcuts and mirroring; every IQ item relies on OneLake data tables, and shortcuts reference external data in place without copying it. The same page names Fabric workloads, Foundry, and Copilot Studio as destinations for that context, with data agents publishable to Microsoft 365 and custom apps too. What’s New (Microsoft Learn, 2026) lists a generally available integration adding a Fabric data agent to Copilot Studio as an MCP tool while retaining source permissions, a coherent design for a Microsoft-first estate.
For a mixed estate, the question shifts from data reach to definition reach: when a table’s metric logic and ownership live in dbt or the source platform, are they reconciled into the layer automatically or re-authored by hand, and which agents on which platforms can query the result? Teams with agents on several platforms usually answer both with a multi-cloud context layer, the pattern in multicloud AI agent governance: a context layer is governed only as far as it reaches.
What questions should you ask before trusting any AI agent’s context layer in production?
Six questions separate a context layer you can govern from one that surprises you in production, and they apply to Fabric IQ, to Atlan, and to anything your team builds. The right-hand column shows what Microsoft’s public documentation answers for Fabric IQ as of September 28, 2026.
| Question to ask | Why it matters | What Microsoft’s docs answer for Fabric IQ |
|---|---|---|
| 1. Preview or GA? | Can change before release | Ontology: preview (What’s New); Fabric IQ GA at Build 2026 |
| 2. How does it refresh, at what cost? | Sets the staleness floor | Manual or scheduled full re-ingest (refresh guidance) |
| 3. What staleness can each agent tolerate? | A forecast tolerates lag an operations agent can’t | Not documented; set per agent |
| 4. What can it reach, which agents read it? | Decides multi-platform fit | OneLake incl. shortcuts; Fabric agents, Foundry, Copilot Studio (overview) |
| 5. Whose permissions, scoped to autonomy? | Gartner names binary governance the root cause of failure | Own Entra Agent ID, creator’s delegated permissions (guide) |
| 6. Who owns refreshes and definitions? | Unowned schedules drift; untraceable context can’t be audited | Internal; operations agents keep an activity log |
Microsoft’s public docs answer questions 1, 2, 4, and 5 for Fabric IQ; questions 3 and 6 are yours, alongside the runtime controls in the enterprise AI agent guardrails checklist.
Ownership usually decides the rest: schedules, definitions, and certification each need a named owner, the split context layer ownership works through, and decision traces for AI agents prove afterward what an agent used.
A layer that answers all six in writing is ready to ground a production agent. One that can’t is still a pilot, whoever built it.
How Atlan approaches context for AI agents
The same six questions apply to Atlan, whose context spans many platforms: a revenue metric might be modeled in dbt, reported in Power BI, and queried in a warehouse, each copy on its own schedule, and any context layer, Atlan’s included, is only as current as its last sync.
Atlan is the Context Layer for AI, built to sit across those systems: connectors spanning warehouses, BI tools including Power BI, dbt, and orchestrators, each syncing on a schedule you set. Context Agents work the gaps, flagging where teams define a metric differently or scoring assets on completeness and freshness. The Context Lakehouse stores that context on Apache Iceberg with time travel, so a team can reconstruct what an agent saw, serving it to any agent over MCP; in Context Engineering Studio, a domain owner approves what becomes canonical context engineering on a context graph spanning the estate.
For Fabric-committed teams the scopes coexist: Fabric IQ grounds agents inside Fabric, while a cross-platform layer carries definitions and lineage from everywhere else, sequenced in how to implement an enterprise context layer for AI.
Governing Fabric AI agents starts with the context they reason on
Microsoft has documented more about Fabric IQ’s ontology than most buyers realize: preview status, refresh mechanics, reach through OneLake, and per-agent identity, turning four of six questions into answers you can plan around today. The rest no vendor can answer: how stale can each agent’s context be, who owns the refresh, and do its permissions match its autonomy? Ask them of Fabric IQ, of Atlan, and of every layer your agents read, then write the answers down before the agent ships, mapped against context layer 101, the core components of a context layer, and, one level down, an AI-ready data checklist covering what makes data AI-ready in the first place.
FAQs about governing Fabric AI agents
1. Is Fabric IQ’s ontology generally available, or still in preview?
Fabric IQ’s ontology is in preview as of September 2026. Status is set per item, and some related items are already generally available, so date-stamp any readiness decision.
2. How often does Fabric IQ’s ontology refresh its data?
As often as you trigger or schedule it. Edits to the ontology schema propagate automatically, but upstream row changes appear only after a graph refresh, run manually or on a schedule from every minute to monthly. Each refresh is a full re-ingest with capacity cost.
3. What’s the difference between a semantic layer and a context layer?
A semantic layer defines how metrics are calculated so reports return consistent numbers. A context layer carries meaning, lineage, ownership, quality, and policy across every system an agent reads, so the agent can decide whether to trust a number. Most enterprises need both.
Sources
- What is Fabric IQ?, Microsoft Learn (Microsoft, 2026)
- What is ontology (preview)?, Microsoft Learn (Microsoft, 2026)
- View entity type details: Refresh the graph model, Microsoft Learn (Microsoft, 2026)
- Manage and refresh data in graph in Microsoft Fabric, Microsoft Learn (Microsoft, 2026)
- What is graph in Microsoft Fabric?, Microsoft Learn (Microsoft, 2026)
- Create and configure operations agents, Microsoft Learn (Microsoft, 2026)
- What’s new in Microsoft Fabric, Microsoft Learn (Microsoft, 2026)
- What’s new archive for Microsoft Fabric, Microsoft Learn (Microsoft, 2026)
- FabCon Europe 2026: The sessions we’re most excited to bring to Barcelona, Microsoft Fabric blog (Microsoft, 2026)
- Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure, Gartner (Gartner, 2026)
- Gartner on uniform AI agent governance failure, Enterprise DNA (Enterprise DNA, 2026)
- State of AI in the Enterprise, Deloitte (Deloitte, 2026)
- AI Risk Management Framework, NIST (NIST)
- Microsoft Fabric IQ adds semantic intelligence layer to Fabric, InfoWorld (InfoWorld, 2025)
