Microsoft's Build 2026 bet: agents need context layers, not just better models

At Build 2026, Microsoft unveiled a stack of 'IQ' APIs designed to give AI agents reliable access to enterprise data, identity, and the web. VP Marco Casalaina explains what's actually shipping — and what enterprises are doing with it.

Written by Lena Armitage · Bureau Tech · Jun 12, 2026

The surprising claim buried in the keynote

Microsoft's biggest Build 2026 announcement isn't a new model. It's a bet that the bottleneck for enterprise AI isn't raw capability — it's context. The company is shipping four "IQ" APIs designed to give AI agents reliable, authenticated access to the data they need to do real work: unstructured knowledge, structured business data, the Microsoft 365 ecosystem, and the live web.

That framing comes directly from Marco Casalaina, Microsoft's VP of Products for Core AI and its self-described AI Futurist — a title he defines concretely as being the first person inside Microsoft to try anything new, focused on "about a year out from now."

What the IQ stack actually is

The four IQs are headless APIs — no user interface, no dashboard for end users to navigate. They are, in Casalaina's words, "agent-facing." Developers connect them; agents consume them.

- **Foundry IQ** handles retrieval across unstructured enterprise knowledge and the live web. - **Fabric IQ** gives agents direct access to data stored in Microsoft Fabric and Power BI, bypassing the need to parse reports. - **Work IQ** (generally available June 16) is the agentic interface to Microsoft 365 apps — Outlook, Teams, Word, SharePoint. - **Web IQ** is a new agent-optimized web search stack, including video search and limited browsing, designed for speed and headless operation.

All four are exposed as MCP servers. MCP (Model Context Protocol) is an emerging standard for self-describing, agent-readable APIs with authentication built in. Casalaina describes it plainly: "It's not that fancy. That's really what it is."

Agent identity is the underrated piece

One announcement that deserves more attention than it's likely to get: agents can now hold their own identity in Microsoft's Entra system — the same identity infrastructure used for human employees. That means an agent can have its own Teams inbox, its own email address, and authenticated access to Work IQ on its own behalf. For enterprises worried about audit trails and access governance, this matters more than another benchmark.

The MAI models: what Microsoft is and isn't claiming

Microsoft introduced seven new MAI models at Build, including MAI-Thinking-1. Casalaina is careful about what he claims for them: they're built for token efficiency, cost optimization, and customer fine-tuning — including continued pre-training, which modifies model weights directly rather than adding a fine-tuning layer on top.

He's also explicit about what they're not: "Our MAI models are not distilled. Some model providers, especially some of the less scrupulous ones, will distill other models into theirs." That's a pointed remark in a market where distillation provenance is increasingly contested. Whether Microsoft's data-provenance claims hold up to external scrutiny remains to be seen — the company hasn't published a model card or training data disclosure for MAI at the time of writing.

Where agents are actually working

Casalaina offers two enterprise deployments worth noting. Bayer built an internal agent system on Foundry that now serves 20,000 employees. AEMO, the Australian Energy Market Operator, built an alert-triage system for grid operators — taking a constant stream of sensor alerts and surfacing the ones that require human action, along with historical context on how similar issues were resolved.

Both cases share a pattern Casalaina calls "human-centered agents": AI handling volume and retrieval, humans making consequential decisions. That's a more defensible framing than full autonomy, and it maps to where enterprise deployments are actually succeeding.

The rubric-based evaluation gap

Microsoft is also shipping rubric-based evaluation in preview — a more granular alternative to standard metrics like groundedness (the degree to which a model's output is anchored to source material, as opposed to hallucinated). Standard evals test whether an agent answers correctly. Rubric evals test whether it follows the right steps: did the restaurant-reservation agent check availability before confirming a table? That's a meaningful distinction for production deployments, and it's one the industry has been slow to address.

Key takeaways

Microsoft's four 'IQ' APIs (Foundry, Fabric, Work, Web) are headless, MCP-exposed context layers for agents — not end-user products. Developers wire them in; users mostly won't see them directly.
The MAI model family is built in-house and explicitly not distilled from other providers' models — a data-provenance claim Microsoft is leaning on as a differentiator.
Work IQ APIs go generally available June 16, giving agents authenticated access to Outlook, Teams, SharePoint, and related Microsoft 365 data.
Bayer has 20,000 employees on a Foundry-based agent system; Australia's energy grid operator (AEMO) uses agents to triage a constant flood of grid alerts — two of the more concrete enterprise deployments cited.
Microsoft 365 Copilot now claims more than 20 million monthly active users, up 6x over the past year, according to Casalaina — though the company has not independently published that figure with methodology.

FAQ

What is Microsoft IQ, and who is it for?

Microsoft IQ is a family of four headless APIs — Foundry IQ, Fabric IQ, Work IQ, and Web IQ — that give AI agents authenticated access to different types of enterprise and web data. They're designed for developers building agents, not for end users to interact with directly. All four are exposed as MCP (Model Context Protocol) servers.

When does Work IQ become generally available?

According to Casalaina, Work IQ APIs are scheduled to go generally available on June 16, 2026.

What are the MAI models, and how do they differ from OpenAI or Anthropic models on Azure?

MAI is Microsoft's in-house model family, announced at Build 2026. Microsoft says they're optimized for token efficiency and customer fine-tuning, including continued pre-training. Casalaina states they are not distilled from other providers' models. OpenAI and Anthropic models remain available through Azure Foundry alongside MAI.

How does agent identity work in Microsoft's system?

Agents can now be assigned their own identity within Microsoft's Entra identity platform — the same system used for human employees. This gives agents their own email inbox, Teams presence, and authenticated access to Work IQ, enabling clearer audit trails and access governance for enterprise deployments.

What is rubric-based evaluation, and why does it matter?

Rubric-based evaluation tests whether an agent follows the correct procedural steps for a specific task, rather than just checking whether its outputs are factually grounded. For example, it can verify whether a booking agent checked availability before confirming a reservation. Microsoft is shipping this capability in preview as part of its Agent Optimizer tooling.

Does Microsoft still offer models from other providers, including Chinese labs?

Yes. Casalaina confirmed that Foundry continues to offer models from OpenAI, Anthropic, Mistral, xAI, DeepSeek, Qwen, and others. Microsoft's position is model-choice agnosticism at the infrastructure layer, with MAI filling a specific niche around efficiency and customizability.

Citations

Microsoft's AI Futurist explains how he uses Copilot — and the real-world problems enterprises are solving with agentsMarco Casalaina interview covering Microsoft Build 2026 announcements including IQ APIs, MAI models, agent identity, and enterprise deployments at Bayer and AEMO
Microsoft Build 2026 — primary source for product announcementsMicrosoft announced Foundry IQ, Fabric IQ, Work IQ, Web IQ, seven MAI models, Scout personal agent, and rubric-based evaluation at Build 2026
VentureBeat — source publication for Casalaina interviewBureau research source: VentureBeat reported on Microsoft Build 2026 and the Casalaina interview