The positioning holds up. The availability details are messier than most headlines suggest, though, and admins should get them right before telling anyone that Claude can now read the CEO's inbox.
What Work IQ is and why Microsoft keeps pushing it
Work IQ is Microsoft's name for the context engine under Microsoft 365 Copilot. It is how Copilot knows who your manager is, which thread is about the Contoso renewal, and what was agreed in yesterday's standup. In its Ignite post, Microsoft calls it "the intelligence layer for work" and says you already experience Work IQ every day in Microsoft Copilot. Developers can use it too, because custom agents and apps can use this same intelligence via the Work IQ APIs.
Those APIs are not new this week. In June, Microsoft announced that the Work IQ APIs will be generally available on June 16, 2026. It pitched them as cheaper on tokens because specialized large language models (LLMs) and agents within Work IQ package relevant context and data in a structure that is easier for agents to consume. Ignite builds on that foundation rather than replacing it.
The strategy is easy to see. Constellation Research observed that Microsoft repeatedly has argued that Work IQ is its differentiation. Anyone can license a frontier model. Only Microsoft has your tenant's mail, chats, meetings and files already indexed and permission-trimmed.
Section summary: Work IQ is the context layer Copilot already runs on. It was exposed to developers through APIs that reached general availability in June. Ignite extends it from reading context to acting on it.
The Ignite announcements and their status
Microsoft published two descriptions of the same release: an Azure blog post from Yitzhak Kesselman, CVP of Microsoft IQ, and a Microsoft 365 blog post from Amanda Silver. They do not label every feature the same way, so this table records what each says.
| Capability | What it does | Status per Microsoft |
|---|---|---|
| Business Applications grounding | Adds CRM (Dynamics 365) and Power Platform data, plus ERP, to agent context | Azure post: GA for CRM and ERP. Silver's post: ERP is preview |
| Retrieve | Returns permission-trimmed Microsoft 365 content and federated connector data in one call | GA |
| Real-time meeting context | Agents use Teams transcripts while a meeting is still running | GA |
| Long-running tasks and agent-to-agent (A2A) | Multi-step work, with agents asking clarifying questions as they go | GA |
| Granular MCP policies | Scope access by path, audience and trusted app instead of one tenant-wide switch | GA |
| ChatGPT plugin | Work IQ inside ChatGPT Business and Enterprise | Available |
| Anthropic/Claude plugin | Work IQ inside Claude | Azure post: GA. Silver's post: "coming soon" |
| Memory | Carries context across sessions and products | Public preview |
| Word, Excel and PowerPoint creation and editing | Agents produce real Office files | Public preview |
| Listen event subscriptions | Agents react to events such as a new message instead of polling | Public preview |
| Agent policy governance with human approval | Agents pause for sign-off before consequential actions | Public preview |
| Engage | People and organizational insights | Preview |
The two disputed rows are the ones that matter most for planning. UC Today reported that CRM and ERP grounding, real-time meeting context, long-running tasks, agent-to-agent support and ChatGPT and Claude plugins are now generally available. That matches the Azure post. Silver's Microsoft 365 post is more cautious: it says Business Applications in Work IQ brings ERP (preview) and CRM from Dynamics 365, and Power Platform data together with Microsoft 365 context, and that the Work IQ plugin for Anthropic is coming soon.
What does that mean in practice? If your organization runs on Claude, don't promise users a Work IQ connector until it appears in your tenant. The same goes for ERP grounding: treat it as preview until Microsoft clarifies. The ChatGPT plugin is the one third-party integration that both posts clearly describe as available today, and it is limited to ChatGPT Business and Enterprise customers.
Microsoft also says Work IQ is generally available in Foundry and in Copilot Studio, with access over MCP, REST and A2A. Silver's post adds that new SharePoint tools in the Work IQ API give agents more than 60 native SharePoint operations, including lists, libraries, approvals, Quick Steps and versioning. Her example is that asking an agent to "build me a tracker" should produce a working tracker, not instructions for making one.
Section summary: The CRM grounding, Retrieve, real-time meeting context, A2A and MCP policies are clearly GA. ERP grounding and the Claude plugin get different labels in Microsoft's own posts. Memory, Office file creation, Listen and human-approval governance are all preview.
Microsoft IQ: the new umbrella
Work IQ is now one of four "IQs." Microsoft's Azure Ignite post says Microsoft IQ brings together Work IQ, Fabric IQ, Foundry IQ, Web IQ, and more than 350 connectors, so the knowledge your agents need is captured once, governed in one place, and reused by every agent and app. In the Kesselman post the split is:
- Work IQ: how people work, including documents, conversations and past decisions
- Fabric IQ: governed business data, entities, relationships and rules
- Foundry IQ: curated institutional knowledge such as policies
- Web IQ: timely web context, including news, images and video
The packaging mechanism is "business plugins." Microsoft says the same context and business plugins reach your agents wherever they run, in Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and assistants such as Claude and ChatGPT, through model context protocol (MCP) and APIs. Note the wording: business plugins in the Claude marketplace are not the same thing as a dedicated Work IQ plugin for Claude. That difference is probably the source of the conflicting headlines.
The Dynamics team is leaning in as well. Its Ignite post says because agents built in Copilot Studio can be natively integrated with Work IQ, they bring the same knowledge, context, and intelligence as Copilot, and adds that with Agent Accounts entering public preview in December 2026, plus memory and triggers, agents become persistent and proactive parts of shared business processes.
Section summary: Microsoft IQ packages Work IQ with Fabric, Foundry and Web context and distributes it through business plugins, including to non-Microsoft assistants.
Governance: the part admins should read twice
Agents that only read can leak data. Agents that act can leak data and also break things. Microsoft seems to know this, and the governance section of Silver's post is its most concrete:
- Tenant-wide action ceiling. Admins set an upper limit on agent actions in the admin center, for example allowing read, create and update but not delete. Each agent's policy can narrow that ceiling but never exceed it.
- Approval workflow for Agent 365 agents. The maker proposes a policy and the admin approves or denies it. Work IQ checks the policy on every call using the agent's own identity.
- Split permissions. New
WorkIQ.ReadandWorkIQ.Actpermissions separate reading from acting across Microsoft 365 data, business applications and connected sources. - Allow, block or ask. Policies can permit an action, deny it, or require a human to approve it first. Silver's example is letting an agent update a CRM record without handing it "a blank check."
- Prompt-injection screening. Before an agent sees content, an isolated model with no tool access transforms it into sanitized output. Microsoft also says it is working on provenance signals to judge where data came from and how trustworthy it is.
Microsoft's stated principle is that access control lives outside the LLM, so a model that gets talked into misbehaving still runs into a policy wall. That is the right architecture. Keep in mind that these are Microsoft's descriptions of its own controls, not results from independent testing. Prompt-injection defenses in particular have a poor track record across the industry, and a sanitizing model is a mitigation, not a guarantee. The provenance work is explicitly unfinished.
One caveat matters most. The human-approval governance is listed as preview, so the most important brake on consequential actions is not GA yet, even though several of the action-taking features are.
Section summary: The control model is layered and well designed on paper. Human approval is still preview, though, so pilots should start with read-heavy scenarios.
Practical steps for IT teams this week
Based on what Microsoft has published, a sensible first pass looks like this:
- Inventory third-party AI use. If staff already use ChatGPT Business or Enterprise, decide whether the Work IQ plugin is sanctioned before someone connects it. Microsoft says the plugin respects Microsoft 365 controls and only retrieves data the signed-in user can access. That is reassuring, but it also means oversharing problems in your tenant come along with it.
- Clean up permissions first. Permission-trimmed retrieval is only as good as your SharePoint and Teams permissions. Every agent that can see a sprawling "Everyone" site will see it too.
- Set a conservative tenant ceiling. Read, create and update without delete is Microsoft's own example and a reasonable default.
- Plan for cost. Silver says leaders get agents that already know what changed and can handle the follow-up along with a cost estimator for spend. Published per-layer pricing for the IQs was not detailed in the announcement material, so model usage before scaling up.
- Watch the sessions. Microsoft points to BRK304 ("What's new in Work IQ") and BRK296 (Work IQ governance and trust) for deeper detail.
The bigger picture
Every major vendor now says it is the context layer for AI. Microsoft's argument is that it already holds the context. Microsoft's own figures put Microsoft 365 at 19.5 exabytes of data, with roughly 90 billion emails a week and 15 billion Teams chat and channel messages a week. Treat those as company figures, not audited ones, but the size of the advantage is real.
Opening Work IQ to ChatGPT, and eventually Claude, is an interesting choice. Microsoft is betting that being the trusted source of work context is worth more than keeping users inside Copilot. That's sensible if the governance holds up, and risky if it doesn't. Either way, the ceiling-and-approval model gives admins a better starting point than the "connect your Gmail to this browser extension" approach that has dominated the agent world so far.
Two things are worth watching: when human approval reaches GA, and when the Claude plugin actually shows up. Until both happen, Work IQ can take actions, but admins should keep its permissions narrow.
References
- Microsoft Ignite 2026: What's new in Azure, the platform purpose-built for AI azure.microsoft.com
- Microsoft Work IQ Moves From Understanding Work to Doing It - UC Today UC Today · 2026-11-17T19:15:18.216000+00:00
- Your Privacy Choices Opt-Out Icon azure.microsoft.com