Building an agent on your company's data with Microsoft IQ

An illustration of a wooden card catalogue with two drawers pulled open and two still sealed shut
AI-generated illustration.

Microsoft IQ is Microsoft's name for the layer that gives an agent your company's context: who's working on what, what your numbers mean, and which documents count. It's four services under one brand, and there's no IQ SKU to buy. You pay for what sits underneath: Copilot seats, Copilot Credits, Fabric capacity and Azure AI Search.

I went through the documentation, the pricing pages and the analyst coverage, because the launch material makes it hard to tell what you can build on today. The short version is that there's a generally available path for an agent over your Microsoft 365 data, your documents and your Power BI numbers. The piece Microsoft leads with, the Fabric IQ ontology that's meant to teach agents what your business terms mean, is still labelled preview in the docs, and part of the bill is unpublished.

Everything below is as of 5 November 2026. Ignite runs 17 to 20 November, and the status table is the part most likely to move, so I'll update this after it.

What it is

Microsoft's IQ hub lists four services. Each gives an agent a different kind of context.

Service	What it gives an agent	How an agent reaches it
Work IQ	Microsoft 365 work data: mail, calendar, files, people, chat and sites	Copilot Studio, plus an API and an MCP server for your own agents
Fabric IQ	Business data in Fabric and Power BI, plus an ontology that defines your business terms	Fabric data agents, and a plugin in Copilot Cowork
Foundry IQ	Retrieval over documents in Blob storage, OneLake, SharePoint, existing search indexes and the web	A knowledge base in Azure AI Search, exposed to Foundry agents as an MCP tool
Web IQ	Web grounding	Limited access

The Foundry IQ FAQ puts the design plainly: "Each IQ workload is standalone, but they can work together to answer virtually any organization-specific question from an agent." The combining is where it gets uneven. Copilot Studio supports Work IQ, and Microsoft's own training said in June that "support for Foundry IQ and Fabric IQ in Copilot Studio continues to expand" (Learn). Knowledge sources don't cross over either. The same FAQ says "you can't use Foundry IQ knowledge sources in Copilot, and you can't use Copilot knowledge sources in Foundry IQ." So the platform you choose for documents decides where those document connections live.

Work IQ looks the furthest along to me. Microsoft says its MCP server "collapses hundreds of operations into just 10 generic tools", and that "a Rego-based policy engine enforces detailed, context-aware rules on every request" (Work IQ overview). Treat the ten as Microsoft's phrase: the MCP reference page tabulates twelve named tools.

Web IQ is the newest. It was announced at Build in June with sub-200 ms grounding and zero data retention, and it's available only to select Azure customers, so I've left it out of the rest of this piece. One naming note: Azure AI Foundry is now Microsoft Foundry, so tutorials from last year use the old name.

What has shipped

Status as of 5 November 2026, from the docs page for each item, with the page's last-updated date where it shows one.

Piece	Status	Source
Work IQ API	GA since 16 June 2026	Licensing notice
Work IQ in Copilot Studio	GA	Learn, 22 Oct
Work IQ MCP `retrieve` tool	Preview	Learn, 5 Oct
SharePoint tools in Work IQ	Preview since 3 Nov 2026	SharePoint blog
Fabric data agent	GA	Learn, 21 Oct
Fabric data agent in Copilot Studio	Preview	Learn
Fabric IQ plugin in Copilot Cowork	GA	Learn, 8 Oct
Fabric IQ workload and ontology	Preview in the docs; GA in the Build blog	Learn, 29 Sep
Foundry IQ knowledge bases	GA, with some features in preview	Learn
Web IQ	Limited access	Foundry blog
Entra Agent ID for Copilot Studio agents	Preview	Learn

Fabric IQ is the confusing one, and Microsoft's own pages disagree three ways. The Build blog in June called it generally available. A Learn training unit the same month counted it among "three generally available services". The Fabric IQ overview, updated in late September, still describes "the IQ (preview) workload" and labels the ontology preview throughout. My read is that the brand and some of its items are GA and the ontology isn't. Check the label on the specific item you plan to use, on its docs page.

Foundry IQ has a smaller version of the same problem. The knowledge bases were announced GA at Build, but the overview says "some Foundry IQ features are now generally available, while others remain in preview", and the connection walkthrough needs a preview SDK or the 2026-08-01-preview REST API, with query planning and answer synthesis both marked preview.

Building on it

There are four realistic paths. The cheapest involves no build at all, and it's the one I'd start with.

Fabric IQ in Copilot Cowork

If your numbers already live in Power BI, the Fabric IQ plugin lets people ask questions of reports from inside Copilot Cowork. "Queries run against Power BI as you, so existing item permissions and row-level security (RLS) continue to apply." It needs no Fabric capacity beyond what your Power BI content already uses, but it does need Microsoft 365 Copilot licences and Cowork's usage-based billing turned on. Answers don't yet cite the report they came from, and DLP isn't supported in Cowork. What it buys you is a cheap answer to the question that decides everything after it: are your semantic models good enough for an agent to answer from?

A Copilot Studio agent with Work IQ

Work IQ gives a Copilot Studio agent the user's SharePoint, OneDrive, Outlook mail, Teams, calendar and people. Two lines on that page deserve a decision before launch. "Selection is limited to whole workloads," so you can switch on SharePoint but you can't narrow it to one site. And the page says it plainly: "Write operations are on by default." The page also notes that "Work IQ uses the GitHub Copilot harness and Copilot Credits", which puts it on a different harness and billing model from the older per-workload tools it sits beside.

A Copilot Studio agent with a Fabric data agent

This is the path for questions about numbers. A Fabric data agent turns a question into read-only SQL, DAX or KQL over up to five sources (lakehouses, warehouses, KQL databases, Power BI semantic models, ontologies and Microsoft Graph) and runs it as the person asking. You publish it in Fabric and add it to a Copilot Studio agent. The prerequisites are an F2 or larger Fabric capacity (or P1), cross-geo processing and storage for AI switched on in the tenant, both agents in the same tenant, and a Microsoft 365 Copilot licence for whoever builds it. The connection is preview, and the page says it's "only validated for Microsoft Teams."

The data agent's limits are worth reading before you promise anyone anything. It answers from tables only: it "doesn't support unstructured data, such as .pdf, .docx, or .txt files." It works in English. "At present, responses are capped at a maximum of 25 rows and 25 columns." "You can't change the LLM that the Fabric data agent uses." Microsoft says the product team tested it "across a range of public and private datasets", and publishes no accuracy figure from that testing.

A Foundry agent with Foundry IQ

Foundry IQ is the pro-code path, and it's retrieval over documents built on Azure AI Search. You create knowledge sources (Blob storage, OneLake, SharePoint and existing indexes, or SharePoint and the web queried remotely), group them into a knowledge base, and connect a Foundry agent to the knowledge base's MCP endpoint using the project's managed identity (walkthrough). knowledge_base_retrieve is the only tool Agent Service supports for it, hub-based projects aren't supported, and the retrieval effort setting caps your sources: up to ten at minimal, which skips LLM planning, three at low and five at medium (FAQ).

Microsoft's Build post says Foundry IQ now unifies Work IQ, Fabric IQ, Azure SQL, File Search and MCP sources behind one SLA-backed retrieval endpoint. The docs' list of knowledge sources doesn't include those yet, so I'd treat it as direction until the docs catch up.

Which path for which data

Your data	Where I'd start
Mail, meetings, Teams, files people already share	Work IQ in Copilot Studio
Numbers already modelled in Power BI	Fabric IQ in Cowork, then a Fabric data agent
Policies and process documents across several stores	Foundry IQ
Live records in an operational system	Your own API or MCP server, or mirror the database into Fabric first

Security

The default identity is the user's, and that's the best thing about the design. Work IQ says "every request runs in the context of a specific user and only accesses what that user is allowed to see or do." The Fabric data agent uses "the requesting user's identity and permissions", row-level and column-level security still apply, and it respects Purview DLP and access restriction policies on the underlying data. The Purview audit and risk features for agents are still preview.

Foundry IQ is where you can hurt yourself. Per-user filtering works by passing the user's token in the x-ms-query-source-authorization header, and the walkthrough is blunt about the alternative: "Without the token, permission-enabled sources return results unfiltered." Document permissions are only enforced when the source supports them and they've been "explicitly configured for synchronization" (FAQ). Remote SharePoint is the exception: it enforces permissions itself through the Copilot Retrieval API, and every user needs a Microsoft 365 Copilot licence for it.

The governance defaults differ by platform. In Copilot Studio, Power Platform DLP policies apply to every agent, and Microsoft 365 audit applies once you deploy to Teams or Copilot. Foundry needs RBAC, policies and logging configured explicitly, because, in Microsoft's words, "governance doesn't come by default" (Learn).

Then there's residency. Connecting a data agent to Copilot Studio comes with this sentence: "responses returned by Fabric data agents may be sent outside of Fabric's compliance boundary or geographic region, and processed and/or stored according to the Microsoft Copilot Studio(s) applicable terms and data handling policies" (Learn). If you've promised anyone that their data stays in a region, take that sentence to your compliance team before the pilot starts.

What it costs

You can't price IQ as a line item, because there's no IQ SKU. You price the pieces, and as of 5 November 2026 one of them has no published rate. US list prices:

Item	Price	Source
Copilot Business	$21 per user per month list; $18 on annual billing under a promotion that runs to 31 December 2026, or $25.20 on monthly billing	Microsoft, pricing page
Microsoft 365 E7 (E5, Copilot, Agent 365, Entra Suite)	$99 per user per month, on sale since 1 May 2026	Office Watch
Agent 365 on its own	$15 per user per month	Office Watch
Copilot Credits	$200 per pack of 25,000 credits a month	Copilot Studio
Fabric capacity for a data agent in Copilot Studio	F2 or larger; check the Azure calculator for your region	Learn

The useful surprise is what's free. If a user has a Microsoft 365 Copilot licence, agents they use in Copilot Chat, Teams or SharePoint aren't charged for classic answers, generative answers or Graph grounding "when the agent runs with your authenticated identity" (Learn), subject to fair-use limits. My read is that a question-and-answer agent for licensed staff adds little to the bill beyond the seats, as long as it sticks to answers and grounding. Work IQ called from your own app or a third-party agent is different: it's billed in Copilot Credits by usage, and the licensing notice says there's no separate Work IQ subscription or per-user licence.

The unpublished part is Fabric. A Microsoft Message Center notice (MC1469820, via a mirror) said that from 1 October 2026, Copilot in Fabric, Fabric data agents, AI Functions and Fabric IQ ontologies would move to a dynamic model that meters the resources each task uses. The Learn page on data agent consumption still shows the older per-token rates, and I couldn't find published rates or a calculator for the new model. So you can't forecast this cost yet. You can cap it.

Two controls do the capping. Copilot Credit spending policies set limits per user, environment or agent. A Fabric Copilot capacity routes a chosen group's Copilot and data agent usage to one designated capacity, which must be at least F2 or P1 and sit in the tenant's home region. Put a pilot on its own small capacity and the worst case is that capacity's bill.

The case against

The sharpest criticism came early. Barry Briggs at Directions on Microsoft called the IQs "three disconnected services which – frankly – highlight Microsoft's organization chart" when there were three of them. A year on, the conflicting status labels and the knowledge sources that stay on one platform make his point for him. His more important point is about data: "Few enterprises have a well-architected enterprise data model (EDM)", and building an ontology "is a major effort and requires on-going maintenance." He still concluded that "for Microsoft-centric enterprises they're likely the right choice", and recommended you "start at the department level before taking on something as massive as a full EDM."

The evidence of value is thin. The only named customer result I found is Heritage Grocers Group, where a partner's Fabric agents removed more than five hours a week of manual work building sales recaps. That's a partner case study on Microsoft's marketplace blog, with one number. Forrester's 116% three-year ROI for Microsoft 365 Copilot was commissioned by Microsoft, models a composite organisation, and predates IQ. Adoption tells you more. In January, Microsoft claimed 15 million paid Copilot seats, which Directions worked out at about 3.33% of Microsoft 365 commercial customers; by July Copilot had passed 30 million. In a 2025 Gartner survey of 187 IT leaders, only 5% of those who had finished Copilot pilots said they were moving to larger deployment.

There's no accuracy number either. The data agent docs give none. The one quality figure it gives is for Foundry IQ's retrieval, where the FAQ says "agentic retrieval achieves approximately 36% higher response quality than traditional single-shot RAG", on Microsoft's own benchmark.

The rest is churn and lock-in. Names have changed mid-preview, the status labels disagree, and the Foundry IQ walkthrough runs on a preview API version. Fabric IQ's items "rely on OneLake data tables", with outside data brought in through shortcuts (Learn), so the business definitions you write there live in Fabric. Gartner, quoted by Office Watch, called Agent 365 "a work in progress with limited net new functionality to justify its $15 pupm price point."

Where I'd start

If I were starting on this tomorrow, I'd build on the GA half and leave the ontology alone until it's out of preview and has a price.

First, point Fabric IQ in Cowork at one existing semantic model and ask it the questions people actually ask. If the answers are poor, the problem is the model, and an ontology layered on top won't fix it. That's the cheapest way to test Briggs's point about data models on your own data.

Second, if the answers hold up, build one Fabric data agent over that model, running as the user, published to Teams through Copilot Studio, on its own Fabric Copilot capacity so the spend has a ceiling. Clear the residency sentence with compliance first, and decide whether Work IQ's default write access belongs anywhere near it.

Third, measure from the first day: questions answered, and minutes saved per answer against how people get that answer today. I couldn't find that number published for this stack, so a pilot has to produce its own.

Then look again after Ignite. The things I'll be checking are whether the ontology reaches GA and what it costs, whether Fabric publishes rates for the new AI metering, whether the data agent in Copilot Studio leaves preview, and how far Copilot Studio's support for Fabric IQ and Foundry IQ has got.

Sources

Microsoft:

Analysts, press and others: