An account-planning agent is asked to prepare for a customer renewal. A useful briefing may require recent meetings and messages, accepted revenue and service measures, the customer's contract terms, and current public news about the company. Those are four different kinds of context. A capable language model does not possess them automatically, and one broad search connection does not make them equivalent.

Microsoft IQ is Microsoft's family of context capabilities for agents and copilots. Work IQ covers how people work. Fabric IQ covers the state and meaning of the business. Foundry IQ covers curated institutional knowledge. Web IQ covers current external information. They complement one another, but they are not one database, one endpoint, or one permission boundary.

This article explains what each capability provides, how an agent reaches it, and where permissions and validation still sit with the application. It goes deeper into the context layer introduced in Understanding the Microsoft Agentic AI Stack.

Microsoft IQ at a glance

One request can require four different context questions.

A matrix for a customer renewal. Work IQ answers what was recently discussed, owned by Microsoft 365 content owners and bounded by the signed-in user's Microsoft 365 permissions. Fabric IQ answers revenue and open service issues, owned by data and semantic model owners and bounded by Fabric item and data permissions. Foundry IQ answers what the contract and renewal policy say, owned by knowledge base and source owners, with source permissions where supported. Web IQ answers what changed publicly, from external publishers, through a limited-access service.
Each IQ capability answers a different question, with a different evidence owner and access boundary. The agent or its host must connect each one it needs and assemble the evidence itself.

Microsoft IQ Answers Four Different Context Questions

The four capabilities are easiest to understand by the question each one answers.

  • Work IQ: What has this person or team been working on, with whom, and in which Microsoft 365 content and conversations?
  • Fabric IQ: What is true about the business according to governed data, accepted measures, entities, relationships, and operational signals?
  • Foundry IQ: What do the organization's policies, manuals, contracts, product documentation, and other authoritative sources say?
  • Web IQ: What current public information could change the decision?

The separation matters because the evidence has different owners and controls. A private email thread, an approved margin measure, a contract clause, and a public news article should not be retrieved, interpreted, or cited in the same way.

Microsoft IQ supplies context only. The model still performs inference, the runtime still chooses tools, the application still validates the response, and an authorized system still executes any action.

Work IQ Understands How Work Happens

Work IQ is the workplace intelligence layer for Microsoft 365. It can reason over supported email, meetings and calendars, OneDrive and SharePoint documents, Teams messages, people and organizational context, Planner plans, and enterprise search results.

The objective is not merely to return files. A search endpoint may locate a project document; Work IQ can also help an agent see which meeting discussed it, who is involved, and what follow-up the user probably means.

Microsoft describes Work IQ from two views. Data, Memory and Inference explain how workplace understanding is built; they are a conceptual model, not three APIs. Chat, Context, Tools and Workspaces are what a developer can use: a finished answer, agent-ready evidence, operations on Microsoft 365 resources, and SharePoint Embedded storage for long-running work.

Inside Work IQ

How workplace understanding is built is different from how an agent consumes it.

Work IQ cutaway inside one boundary of signed-in user, Microsoft 365 permissions and tenant policy. Data, Memory and Inference form the conceptual layers. Chat, Context, Tools and Workspaces are the developer surfaces, all exposed through the Work IQ API endpoints. Agent-to-Agent, REST and Model Context Protocol are the three routes into those endpoints.
Data, Memory and Inference describe the intelligence model; Chat, Context, Tools and Workspaces describe the usable surfaces. Three routes reach those surfaces, and availability varies by host integration.

The useful architecture question is what the agent needs: a finished workplace answer, selected evidence, an operation over Microsoft 365, or persistent working state.

Agents Can Reach Work IQ in Three Main Ways

The Work IQ APIs use standard routes, so the outer agent does not have to be Microsoft 365 Copilot.

Agent-to-Agent (A2A) is designed for delegation. Another agent sends a natural-language task, and Work IQ retrieves and reasons over the user's Microsoft 365 context before returning a result. This fits a Foundry or custom agent that wants a workplace specialist.

REST provides a conversational request-and-response interface for applications and service-hosted orchestrators. The current surface returns text and supports multiturn conversation, but it has no action or file-generation skills and is not intended for long-running work.

Model Context Protocol (MCP) gives an LLM-based client discoverable tools. Work IQ MCP uses ten generic tools that act on resource paths rather than one tool per operation: entity tools for reads and changes, Copilot tools for natural-language answers, and schema tools for discovering paths at runtime. A local command-line option can give coding assistants the signed-in developer's workplace context.

The choice is about control. Delegate through A2A when Work IQ should own the reasoning, use REST for a conversational answer, and use MCP when the outer agent should choose explicit resources and operations.

Work IQ Preserves User Context but Does Not Validate the Answer

The Work IQ API uses delegated Microsoft Entra authentication. Calls run as the signed-in user, on-behalf-of flows are supported, and application-only authentication is not. Microsoft 365 permissions, sensitivity labels, and compliance policies apply to each request.

Work IQ MCP adds a request-time policy layer. Entra verifies the identity, four broad OAuth permissions gate what an application can request, the user's own permissions limit which resources it can reach, and tenant policy can allow or deny each path and operation. Mutations such as sending email are blocked by default until an administrator enables them.

That boundary answers "may this caller retrieve this evidence?" It does not answer "did the model interpret the evidence correctly?" Microsoft also notes that generated responses may be inaccurate.

The Work IQ APIs are generally available with usage-based billing. Specific host integrations, such as Work IQ tools in Foundry and Copilot Studio, include preview capabilities, so verify the exact host, region, licensing, and identity flow.

Fabric IQ Describes the State and Meaning of the Business

Work IQ follows collaboration. Fabric IQ gives an agent a governed view of the business itself.

Microsoft describes Fabric IQ through three layers. OneLake provides the unified data foundation. Power BI semantic models provide accepted measures, dimensions, hierarchies, and relationships. Ontology adds business entities, properties, relationships, rules, and source bindings for operational and cross-domain reasoning. The wider workload also groups Graph, Planning, data agents, and operations agents.

Fabric IQ adds value when an agent must interpret Customer, Shipment, Margin, or Service Breach consistently rather than infer their meaning from table names. The Fabric IQ workload and ontology are in preview, while services such as semantic models have their own maturity.

Medallion Architecture for AI-Ready Data in Microsoft Fabric explains how that governed meaning is prepared, and Talk to Your Data with Agents in Power BI and Microsoft Fabric compares the analytical agent routes and their answer controls.

Foundry IQ Turns Distributed Content Into Reusable Knowledge

Foundry IQ is the managed knowledge layer for institutional content: policies, manuals, contracts, product documentation, and other sources that several agents need to retrieve consistently. It has three principal parts:

  • Knowledge sources connect indexed or remote content.
  • A knowledge base groups one or more sources and defines retrieval behavior.
  • Agentic retrieval decomposes a complex question, selects sources, runs and reranks searches, and returns evidence or a synthesized result with references.

Azure AI Search provides the underlying retrieval infrastructure. A knowledge base can be called from Foundry Agent Service, Microsoft Agent Framework, a custom application through supported APIs or SDKs, or a supported MCP host.

Instead of every agent building its own document index, several agents can share one curated knowledge base. Foundry IQ can enforce permissions at query time by synchronizing access control lists and honoring Purview sensitivity labels for supported sources, and by running queries under the caller's Entra identity. Permission behavior therefore depends on each source and its configuration.

Maturity is mixed. Knowledge-base APIs and MCP have generally available coverage, while some portal functions, source types, query-planning features, and security capabilities remain in preview.

Web IQ Adds Current External Evidence

Web IQ supplies current public information from web pages, news, images, video, and related sources. It returns ranked, citation-ready material through REST, MCP, or SDK interfaces. It is currently limited access, offered to select enterprise customers rather than something every tenant can enable.

It is useful when the outside world can change a decision: a recent acquisition announcement for an account agent, disruption news for a supply-chain agent, or a new regulator notice for a policy agent.

Public information may be incomplete, contradictory, or newer than approved policy. It is evidence to evaluate, not the organization's system of record.

A Useful Agent Assembles Evidence Before It Recommends Action

The four domains become valuable when they are combined around one decision rather than connected indiscriminately. In the renewal briefing, Work IQ retrieves recent meetings and messages, Fabric IQ supplies account measures and open service issues, Foundry IQ retrieves the contract and renewal guidance, and Web IQ adds public developments.

Illustrative agent pattern

Context becomes useful when each source has a defined job.

Illustrative customer renewal pattern. Work IQ contributes recent collaboration, Fabric IQ contributes account measures and service state, Foundry IQ contributes contract and policy guidance, and Web IQ contributes current public developments. An agent assembles a cited briefing, which passes a human or application review gate before any separate follow-up action.
This is an illustrative architecture, not a bundled Microsoft workflow. Context supports the recommendation; it does not authorize a customer communication or system update.

The same separation of responsibilities applies elsewhere:

  • Support triage: Work IQ supplies the active conversation and owners, Fabric IQ supplies customer and product state, and Foundry IQ supplies support procedures. A separate case tool updates the record after validation.
  • Operational response: A Fabric operations agent detects a condition, Foundry IQ retrieves the procedure, and Work IQ identifies the responsible team. Notification or remediation stays behind an explicit action policy.
  • Long-running research: A2A delegates workplace research, Foundry IQ supplies authoritative sources, a Work IQ Workspace keeps intermediate material, and Tools create the final artifact after review.

These are architectural examples. Verify availability, connections, permissions, and data movement for the actual tenant and agent host.

Context Does Not Replace Evaluation or Governance

Microsoft IQ can improve grounding while leaving four decisions with the application owner:

  • Authority: define which source owns each claim, such as the semantic model for a metric and the contract repository for a legal term, and what wins when they disagree.
  • Provenance: record the source, caller, retrieval scope, result time, and citations. If context is stale or denied, say the answer was not verified rather than filling the gap from model memory.
  • Validation: check required evidence, accepted definitions, unsupported claims, and disclosure scope before a consequential answer reaches a user.
  • Action authority: a retrieved recommendation does not grant permission to send an email, update a forecast, or change a case. Tool policy, approval, and audit still apply.

Governing Agents and Data in the Microsoft Agentic AI Stack covers those controls in depth.

Choose the Smallest Context Surface That Fits

Start with the decision the agent must support and the evidence it needs:

  • Use Work IQ for the signed-in user's workplace activity, collaboration, Microsoft 365 content, tools, or persistent working state.
  • Use Fabric IQ for governed measures, entities, relationships, plans, and operational signals.
  • Use Foundry IQ for curated, reusable retrieval across authoritative institutional sources.
  • Use Web IQ when current external evidence is material and access is available.
  • Choose the narrowest route (delegated answer, agent-ready context, explicit tools, or workspace) and verify its identity, permissions, policy, availability, and data movement.

Reliancy can help define the context contract for an enterprise agent and test whether its evidence, permissions, and validation match the decision it supports. Discuss a Microsoft IQ agent-context review with Reliancy.

Sources and Product Documentation

All articles