A regional manager asks in Teams: Why did margin fall last month, and which accounts need attention?

A trustworthy answer may need a model, approved measures from Microsoft Fabric, policy evidence from enterprise documents, the manager's identity and permissions, and a controlled workflow if the manager wants someone to act. That is why Microsoft's agentic AI offering is a stack rather than a single product.

The simplest useful definition is this: an enterprise agent combines five capabilities (an experience, a runtime, model intelligence, business context and tools) with controls around all of them. Microsoft provides products for each responsibility.

The whole system

One agent. Five connected capabilities. Governance around all of them.

Radial map of the Microsoft agentic AI stack. A central agent runtime connects model intelligence, business context and tools. Copilot Studio, Microsoft Foundry and Agent Framework connect to the runtime as three build paths, experiences sit above, and governance services surround the system.
A production agent needs a place to run, a model to reason, context to ground the answer, permitted tools to act, and a user experience. Identity, data protection, threat defense and lifecycle management surround the complete system.

One Request, Five Capabilities and a Control Layer

  • Experience: Teams, Microsoft 365 Copilot, Dynamics 365 or a custom application is where a person encounters the agent.
  • Build and orchestration: Copilot Studio, Foundry Agent Service and code built with Microsoft Agent Framework provide different levels of abstraction and control.
  • Model intelligence: Foundry Models supplies model choice. Azure OpenAI models are an important family inside that broader catalog.
  • Business context: Microsoft IQ, Fabric, Azure AI Search, SharePoint and other enterprise sources ground the agent in information it could not know from the model alone.
  • Actions: Power Automate, Dataverse, Dynamics 365, APIs and MCP tools let an agent retrieve information or initiate controlled work.
  • Control: Entra, Purview, Defender, Agent 365, evaluation and application-level policies answer different questions about identity, data, risk, quality and accountability.

The products are related, but they are not interchangeable. A development SDK, an Azure platform, a low-code agent product and a governance service belong at different layers.

Three Practical Ways to Build

The right starting point depends on who owns the solution and how much behavior must be engineered. If the choice also involves Claude, ChatGPT or Codex, Copilot, Claude, and Codex on a Microsoft Stack explains which layer each product occupies.

Builder choice

Choose by ownership and complexity, not by logo.

Two-axis builder-choice map. Copilot Studio sits near configured and business-owned work. Foundry prompt agents sit in the middle. Agent Framework sits near engineered and engineering-owned work. A hybrid area connects the paths.
Copilot Studio, a Foundry prompt agent and a code-based agent are three implementation paths. Microsoft Foundry also supplies platform capabilities underneath code-first and hybrid solutions.

Copilot Studio

Copilot Studio is a low-code authoring and managed runtime environment. It is a strong starting point for business-facing agents that need Microsoft 365 channels, Power Platform connectors, knowledge sources, agent flows and administrator-managed publishing.

Choose it when the workflow can be expressed through configured instructions, topics, tools and approvals without turning the agent into a custom software product.

Foundry Prompt Agent

A prompt agent in Foundry is configured with instructions, a model and tools. Foundry operates the runtime, endpoint, identity and scaling. It suits teams that want Azure control and production operations without maintaining custom runtime code.

Code-Based Agent

Microsoft Agent Framework is an open-source SDK for custom agents and workflows. It is the code-first path for explicit state, graph-based workflows, long-running work, human approval, custom middleware and multi-agent coordination.

Agent Framework is the successor to Microsoft's Semantic Kernel and AutoGen agent work. A code-based agent can run in Foundry as a hosted agent, or run elsewhere and call Foundry models and tools.

A hybrid is also valid. Copilot Studio can provide the employee-facing experience while a specialist agent in Foundry handles engineered logic.

What Microsoft Foundry Provides

Microsoft Foundry, formerly Azure AI Foundry, is Microsoft's Azure platform for building and operating AI applications. Foundry Agent Service is the managed agent runtime inside that platform.

Inside Foundry

Foundry turns agent development into a production loop.

Continuous Microsoft Foundry production loop. Teams choose models, connect tools, evaluate behavior, publish a managed endpoint, and trace production activity back into evaluation. Prompt and hosted agents enter the central Agent Service runtime.
Foundry connects model access, agent runtime, tools, evaluation, deployment and observability. Production traces feed the next evaluation and release cycle.

The important components are:

  • Foundry Models: access to OpenAI, Microsoft, Anthropic, Meta and other model families through the Foundry model catalog.
  • Agent Service: managed runtime for prompt agents and hosted code agents, with sessions, identity, endpoints and scaling.
  • Tools and knowledge: built-in tools, custom functions, MCP servers, APIs and reusable knowledge sources.
  • Evaluation and observability: quality tests, safety evaluation, traces, tool-call inspection, latency and failure analysis.
  • Platform controls: Azure RBAC, Entra identity, network isolation, policy and content-safety controls.

Foundry is useful when an agent must be versioned, evaluated, traced and operated like software. Evaluating Production AI Agents with Microsoft Foundry explains what those evaluations can and cannot prove.

Microsoft IQ Supplies the Context

A model can reason, but it does not automatically know your customers, measures, policies, meetings or current market conditions. Microsoft IQ is Microsoft's family of context capabilities intended to make that organizational knowledge reusable across agents and copilots. Each capability is reached through its own integration.

Enterprise context

Different questions need different kinds of knowledge.

Four separate context streams reach an agent. Work IQ provides people and collaboration context. Fabric IQ provides business entities and measures. Foundry IQ provides policies and authoritative documents. Web IQ provides fresh external information. The agent assembles what it needs into a grounded answer.
Microsoft IQ names four kinds of context, each with its own route and permissions. The agent connects only the ones the question needs and assembles the result itself.
  • Work IQ supplies context about people, meetings, collaboration and work activity.
  • Fabric IQ supplies business entities, properties, relationships, measures and current operational state.
  • Foundry IQ supplies knowledge bases over policies, documents and institutional knowledge, built on Azure AI Search. It enforces user permissions at query time only for supported, correctly configured sources.
  • Web IQ supplies current external information from the web. It is currently limited access.

This distinction matters. A policy document and an approved margin measure are both context, but they require different retrieval methods, permissions and evidence. Agent Context Through Microsoft IQ goes deeper into each capability and its access routes.

Where Microsoft Fabric Fits

Fabric is the data and analytics foundation of this stack. OneLake unifies the data estate, data engineering produces governed data products, and Power BI semantic models define measures and hierarchies.

Fabric IQ adds business entities and relationships so agents can work in the language of the business rather than raw table names. Adopting Microsoft Fabric for Business and AI covers where to start, and Medallion Architecture for AI-Ready Data in Microsoft Fabric shows how that meaning is prepared.

Fabric grounding

Raw data becomes business context that several kinds of agents can use.

Fabric grounding cutaway. ERP, CRM, files and event data flow into OneLake. Curated data products, semantic models and an ontology add business meaning. Fabric IQ then supplies context to a Fabric data agent, an operations agent, or an external Foundry or Copilot Studio agent.
Fabric is not simply storage for AI. The valuable layer is the governed meaning built over the data: approved measures, entities, relationships, permissions and evidence.

Fabric supports several agent patterns:

  • A Fabric data agent answers natural-language questions over selected Fabric sources. It is designed for governed analytical Q&A and queries data with the caller's effective permissions. Talk to Your Data with Agents in Power BI and Microsoft Fabric compares it with the other analytical agent routes.
  • A Fabric operations agent monitors real-time data or an ontology, detects conditions, recommends actions and can run configured actions when policy allows. Monitoring Fabric Data Pipelines with the Operations Agent shows one bounded example.
  • A Foundry or Copilot Studio agent can use Fabric IQ as a tool when the wider workflow also needs documents, external systems or custom orchestration.
  • An MCP-compatible agent can connect to supported Fabric IQ endpoints when the scenario calls for a different agent client.

Fabric IQ and ontology capabilities include preview features. Buyers should confirm region, capacity, identity flow, supported data sources and production support before committing an architecture.

The Fabric items that ground agents are also built and released like software, often with help from coding agents. Agent-Assisted Development in Microsoft Fabric and Agentic Development for Power BI Reports and Semantic Models explain how those changes are made; DevOps and Release Automation for Microsoft Fabric explains how they reach production.

Governance Happens at the Point of Use

Governance is not a single control added after development. It is a sequence of controls applied as an agent receives context and proposes action. Governing Agents and Data in the Microsoft Agentic AI Stack maps each control in depth.

Controlled action

A tool call should cross identity, data, policy and approval boundaries.

Governed action sequence. A user request passes an Entra identity check and a data-scope check based on source permissions, with Purview policies where supported, before reaching an enterprise agent. A proposed tool call then passes an allow, approve or block check. An approved workflow executes and an audit rail records the full event.
Identity proves who is acting. Data controls limit what can be retrieved or disclosed. Tool policy and human approval limit action. Audit and evaluation make the outcome inspectable.

The services have different jobs:

  • Microsoft Entra Agent ID gives agents a manageable identity and supports authentication, authorization and Conditional Access.
  • Microsoft Purview applies classification, information protection, DLP, audit and compliance controls to supported agent activity.
  • Microsoft Defender adds posture management, threat detection and investigation.
  • Microsoft Agent 365 provides organization-wide inventory, lifecycle management and administrative oversight.
  • Foundry evaluation and observability measure quality, regressions, latency, tool use and failures. These are operational controls, not substitutes for access control.
  • Application policy and human approval determine whether a specific tool call is allowed, blocked or held for review.

What Deployments Look Like

The same stack supports different levels of autonomy. A useful design decision is whether the agent should answer, recommend or execute.

Deployment patterns

Answer, recommend and execute require different controls.

Three Microsoft agent deployment storylines. A Fabric semantic model grounds an analytics answer in Power BI. Dynamics lead information supports a qualification recommendation for seller review. A Teams service request passes policy and approval before Power Automate and Dataverse execute an update.
Increasing autonomy changes the control point. Answers need evidence. Recommendations need accountable review. Actions need explicit tool policy, approval and audit.

Governed Analytics Over Fabric

A business user asks a question in Fabric, Power BI, Copilot Studio or a custom application. The agent delegates the analytical question to a Fabric data agent or Fabric IQ tool, which evaluates it against approved data and permissions. The response should carry the measure, scope, source and relevant caveats.

Sales Qualification in Dynamics 365

Microsoft's Sales Qualification Agent is a first-party Dynamics 365 example. It uses CRM and Dataverse context, research and specialist agents to help qualify leads. The seller remains responsible for commercial judgment and the next customer action.

Service Workflow in Teams

A Copilot Studio agent collects a request in Teams. A deterministic Power Automate flow validates required fields and obtains approval. Dataverse or another system is updated only after the policy gate passes.

What a Buyer Should Decide First

Before choosing products, answer six questions:

  1. Should the agent answer, recommend or act?
  2. Which measures, entities, documents and live signals must ground it?
  3. Is the solution primarily configured, engineered or hybrid?
  4. Which tools can it call, and which actions require human approval?
  5. Who owns evaluation, monitoring, incident response and retirement?
  6. Which preview capabilities are acceptable for the workload and risk level?

Microsoft provides a broad agent stack. The architecture becomes simpler when each product has one clear responsibility and the design begins with the business decision, the trusted context and the permitted action.

Discuss your Microsoft data and agent architecture with Reliancy.

Sources and Product Documentation

All articles