Microsoft IQ Turns Workplace Data Into a Metered Agent Brain

Microsoft IQ is a billed context layer for enterprise agents, live in GitHub Copilot, Microsoft Foundry, and Copilot Studio since June 2, 2026. Work IQ APIs followed 14 days later, metered in Copilot Credits rather than a per-user SKU.

The launch was easy to file as a memory upgrade. The part that stuck is the meter: custom agents now pull mail, meetings, files, and live web passages through Microsoft’s own plane, and the bill follows the pull.

Four Engines, One Tenant Boundary

Kyle Daigle, GitHub’s chief operating officer and Microsoft’s developer marketing lead, called Microsoft IQ “a new context layer that grounds agents in both world knowledge and enterprise knowledge.” It is not a single app. It is four engines that sit under GitHub Copilot, Foundry, and Copilot Studio, and they inherit the tenant you already run.

An earlier read treated the launch as shared memory for enterprise agents. That still describes the pitch. The second effect is simpler. Once an agent reasons from Work IQ, Fabric IQ, Foundry IQ, and Web IQ, leaving Microsoft means leaving the brain, not just the model.

THE FOUR IQ ENGINES

Charles Lamanna, Executive Vice President of Copilot, Agents, and Platform, wrote that software is moving “from applications built for people to agents that can reason, retrieve context, and even act on a user’s behalf.” Work IQ, he said, builds a live model of how the company operates by reading email, calendar, meetings, chats, files, people, and line-of-business systems.

Data, context, and insights stay inside the Microsoft 365 tenant trust boundary. Actions are auditable. That is the security claim, and it is also why the layer is sticky. You do not bolt this onto a rival cloud. You turn it on where the mail already lives.

https://x.com/satyanadella/status/2061896510250668490

Work IQ APIs Bill in Copilot Credits

Lamanna said the Work IQ APIs would be generally available on June 16, 2026, and that they would be “the best way for agents to interact with Microsoft 365 data and apps.” Licensing later made the commercial shape plain. There is no separate Work IQ API subscription, SKU, or per-user license.

You pay when you build your own agent that calls the APIs, or when a third-party agent grounds in your Microsoft 365 data through them. Tool calls are the simple line item: 0.1 Copilot Credits per Work IQ Tool API call. Chat and Context are variable, billed by how heavy the query is. Microsoft 365 Copilot licensed users keep Work IQ inside first-party Copilot experiences. Custom and third-party agents draw the credit balance.

A new cost dashboard in the Microsoft 365 admin center lets IT set prepaid or pay-as-you-go Credits, put spend caps on tenants, groups, and users, and watch credit requests. Work IQ is the first product in that dashboard. Copilot Studio is slated to follow.

By early September, that meter was no longer a licensing footnote. Custom agents that never received a Credits purse stalled with a credits-required error after a grace window that operators dated to September 1, 2026. The context layer shipped. The invoice arrived on the same pipe.

Developers did not have to wait until GA to sample the surface. Microsoft put a public preview on GitHub ahead of June 16. The production design is four domains, not a grab bag of Graph calls.

WORK IQ API DOMAINS

• Chat: Programmatic access to Microsoft 365 Copilot, including the response and citations a user would see, plus agents inside Copilot.
• Context: The same grounding Copilot would use, returned as agent-ready context instead of a finished answer.
• Tools: Ten generic verbs over Microsoft 365 entities, with MCP progressive disclosure, so agents are not taught hundreds of data-specific tools.
• Workspaces: Tenant-bound storage for files, memory, progress, and intermediate output while a long-running agent works.

Lamanna’s speed claim sits in those ten tools. Work IQ collapses operations that used to mean round trips across mail, files, and calendar. Specialized models inside Work IQ also pre-package context so the calling agent spends fewer tokens stitching raw records. Efficiency here is not a slogan. It is how Microsoft keeps high-frequency agent traffic on its own index.

What Hosted Agents Keep Between Sessions

IQ feeds the brain. Foundry Agent Service is the body that can stay awake. Most chat apps still throw the thread away after each reply. Hosted agents are built to keep files and state across turns, resume after idle scale-to-zero, and run in per-session VM-isolated sandboxes with a persistent home directory.

Jeff Hollan runs Microsoft’s AI agent platform. On July 9, 2026, 37 days after Build, he wrote that hosted agents in Microsoft Foundry were generally available, with native compute for any framework, any language, and any model, wired to Microsoft IQ, Teams, and Agent 365.

https://x.com/jeffhollan/status/2075339567503417745

The useful part of that GA is not another chat box. You can bring LangChain, the Microsoft Agent Framework, or custom code, land it on Foundry, and still point the agent at Work IQ and Foundry IQ. Model choice stays open. The session, the filesystem, the Entra agent identity, and the traces sit on Microsoft’s runtime.

Tracing is generally available for prompt and hosted agents, with OpenTelemetry spans into Application Insights. That is the audit trail enterprises ask for when an agent sends mail or edits a file. Agent 365 extends Entra, Defender, and Purview over local and hosted agents, including ones that never ran on Azure.

THE SHIP CALENDAR SINCE BUILD

1. June 2, 2026: Microsoft IQ goes generally available across GitHub Copilot, Microsoft Foundry, and Copilot Studio, and Web IQ is announced.
2. June 16, 2026: Work IQ APIs leave preview and start burning Copilot Credits.
3. July 9, 2026: Hosted agents in Foundry Agent Service reach general availability.
4. July 29, 2026: Satya Nadella reports Foundry, Fabric, and Work IQ already grounding agents across nearly 90% of the Fortune 500.

It fit the Build 2026 agent-first bet that skipped a Windows 12 splash. The conference sold a control plane, not a new desktop. Hosted agents are how that plane runs overnight jobs without a custom Kubernetes diary for every team.

Why Web IQ Shows Up Inside ChatGPT

Web IQ is the piece that does not care which model you picked. Daigle called it an AI-first web search stack that is model-agnostic and MCP-native, returning relevant passages at nearly 2.5x the speed of the next best alternative. Microsoft’s Web IQ site puts the latency number at 164ms p95.

The Bing search team said the same stack is built on two decades of Bing infrastructure, reworked for multi-step agents rather than a results page. It returns ranked, citation-ready context across web, news, images, and video, and it is designed for direct injection into a model’s context window. Grounding with Bing remains a separate Foundry path. Web IQ is the agent-native one, with a waitlist for most Azure customers.

Nadella later said Web IQ is already used by many of the most popular AI assistants, including ChatGPT. That is the tell. Microsoft is willing to sell web passages to a rival assistant because the scarce asset is not the snippet. The scarce asset is Work IQ sitting on your tenant, Fabric IQ sitting on OneLake, and Foundry IQ sitting on the knowledge bases you already indexed.

Web IQ can look like an olive branch. It still trains operators to treat Microsoft as the grounding vendor even when the reasoner is someone else’s.

Nadella’s July Tally Puts IQ Under the Fortune 500

On the July 29, 2026 fiscal 2026 fourth-quarter call, Nadella stopped talking about IQ as a Build slide and started talking about it as installed base. He put the context layer on top of Fabric and Foundry, then named who was already using it.

Tens of thousands of customers, including nearly 90% of the Fortune 500, are already grounding their agents in enterprise context with Foundry, Fabric, and Work IQ.

Satya Nadella, Chairman and CEO, FY26 Q4 earnings call

THE JULY TALLY

• Foundry customers: 100,000, with revenue more than doubled year over year.
• Paid Fabric customers: 40,000, up more than 60% year over year.
• Agent 365: Nearly 40 million agents registered across tens of thousands of companies.

Levi Strauss & Co. is running models from OpenAI and Anthropic on Foundry as it folds more than 1,000 domain-specific agents into one platform. Telefónica took Foundry as the base of its corporate agent platform, starting with network operations. Those are not Copilot sidebar demos. They are production fleets that already assume Microsoft holds the context.

Nadella also said the model catalog in the cloud had passed 11,000 entries, and that customers building with more than one provider had risen 5 times since the start of the year. IQ is how Microsoft can afford that openness. If the harness, context, memory, and action space sit apart from any one model family, swapping Sonnet for a MAI model does not mean rebuilding retrieval.

The Squeeze on Independent Memory Layers

Dedicated memory vendors still have a job. They sell session memory, user memory, and graphs that follow an agent across clouds Microsoft does not own. Amazon already ships managed short-term and long-term memory inside Bedrock AgentCore. Google sells Memory Bank next to its agent runtime. Those products exist because not every shop will pour mail and warehouse data into one tenant.

Microsoft’s counter is rude in its simplicity. Work IQ does not ask you to build a retrieval pipeline over Outlook and SharePoint. It is the pipeline, permission-aware, already warm, and billed in the same Credits wallet as Copilot Studio. Foundry IQ knowledge bases are generally available with an MCP server that Claude, ChatGPT, LangChain, and the Microsoft Agent Framework can call. The open protocol is real. The corpus still lives in Azure.

Once Work IQ has packed the mail, meetings, and files, swapping in a frontier model often looks like wasted spend. Operators working in Copilot day to day keep making that point in plainer language than the keynote. The context was the missing piece. The model was the expensive one.

Shops that care about grounding hygiene have started splitting the question before they prompt: is this a workplace-data job or a web job? That split is IQ’s product map leaking into habit. Work IQ on, you are in the tenant. Work IQ off, you are on the Bing index. Mixing those without noticing is how an agent quotes a customer thread as if it were public.

Private Preview for Microsoft’s First Reasoner

The model Microsoft wants under this stack is its own. Daigle said MAI-Thinking-1 is a mid-sized reasoner with 35 billion active parameters and a 256K context window, trained from scratch with zero distillation on commercially licensed data. On a blind test, independent raters preferred to Sonnet 4.6, and it matched Opus 4.6 on SWE Bench Pro. It is in private preview on Foundry.

MAI-THINKING-1 AT A GLANCE

• Size: 35 billion active parameters, 256K context, tuned for multi-step instructions, long context, and code.
• Training: Built from scratch, no distillation from third-party models, on clean licensed data.
• Family: One of seven new in-house MAI models, with image, voice, transcribe, and code variants also landing in products.
• Access: Private preview on Foundry, with MAI weights also headed to Fireworks AI, Baseten, and OpenRouter.

Nadella said MAI models on Maia 200 silicon show 40% better performance per watt, and that millions of developers have used MAI-Code-1-Flash on GitHub Copilot with higher acceptance and 10% lower median token use. The company is trying to own the cheap, good-enough reasoner while still hosting everyone else’s frontier weights. IQ is what makes that mix safe to sell. Context stays put when the model changes.

Web IQ remains waitlisted for most Azure customers. MAI-Thinking-1 remains in private preview. The Work IQ APIs are the part already on the bill, and hosted agents are the part already allowed to keep state overnight. For a Microsoft 365 tenant, the context gap did not close with a feature flag. It closed with a credit meter.