Copilot Studio

Microsoft's low-code/no-code SaaS platform for building AI agents. Formerly Power Virtual Agents. The right tool when business users need to own the agent, when deployment to Teams or M365 Copilot is the goal, or when Power Automate workflows are the action layer.

When not to use it: when you need custom RAG, complex orchestration, custom model selection, or non-Microsoft infrastructure.

Copilot Studio vs Microsoft Foundry

DimensionCopilot StudioMicrosoft Foundry
Who builds itBusiness users, power usersDevelopers, ML engineers
InterfaceVisual canvas, natural languageSDK, code, portal
InfrastructureSaaS - Microsoft managedAzure PaaS - customer managed
ModelsMicrosoft-selectedAll 1,900+ Foundry Models
Custom orchestrationNoYes (via MAF)
RAGVia Foundry IQ or DataverseFull control
DistributionTeams, M365, web, Dynamics, mobileVia Copilot Studio publish pipeline
GovernanceDLP, Entra, environment isolationRBAC, networking, ACS
PricingPer messagePer model inference + agent run

Architecture


Knowledge Sources

Connect agents to enterprise knowledge without custom RAG. Copilot Studio handles chunking, indexing, and retrieval:

SourceWhat it indexes
SharePointDocuments, libraries, pages, lists
DataverseBusiness application data, tables
WebsitesPublic pages, intranet portals
Uploaded filesPDFs, Word, Excel, PowerPoint
Foundry IQEnterprise knowledge bases built in Microsoft Foundry
Fabric IQFabric semantic models, OneLake data
Work IQMicrosoft 365 data (email, Teams, calendar) - preview

Using Foundry IQ as a knowledge source is the recommended path for grounding Copilot Studio agents in structured enterprise data. Build the knowledge base in Foundry (with permission sync and layout-aware chunking), then wire it into Copilot Studio.


Actions

Agents do more than answer - they act across systems:

Action typeExamples
Power Automate flowsCreate service tickets, send approvals, update CRM
HTTP / REST API callsAny system with an API
Pre-built connectors (1,000+)Salesforce, SAP, ServiceNow, Jira, Workday
Agent flowsMulti-step automated workflows
Computer-using agents (GA May 2026)Interact with websites and desktop UIs directly

Computer-using agents are the standout addition in 2026. Agents interact with actual UI - browser automation, desktop application control - without API access. Useful for legacy systems that lack APIs.


Agent Types

TypeWhen to use
ConversationalMulti-turn Q&A over knowledge sources - IT help desk, HR FAQs
Task agentExecute a specific process end-to-end - expense submission, onboarding
Autonomous agentRun on a schedule, monitor conditions, take action without user trigger
Computer-using agentAutomate UI-driven processes on systems without APIs
Voice agentReal-time voice interactions

Multi-Agent

Copilot Studio agents can orchestrate other agents:

  • Agent-to-Agent (A2A) - delegate to specialist agents based on intent
  • M365 Agents SDK orchestration - integrate with Teams-native agents
  • Foundry-built agents as specialists - build the complex logic in Foundry, expose it as a Copilot Studio skill

Integration with Microsoft Foundry

Two integration patterns:

1. Publish Foundry agents to Copilot Studio channels Build the agent in Foundry (full control, custom RAG, complex orchestration) then publish it to Teams and M365 Copilot through the Copilot Studio distribution pipeline.

2. Use Foundry IQ as knowledge source in Copilot Studio Business users build their agent in Copilot Studio. Ground it in Foundry IQ knowledge bases that have enterprise ACL sync, layout-aware chunking, and permission-trimmed retrieval.

Foundry IQ in Copilot Studio


Distribution Channels

ChannelStatus
Microsoft TeamsGA
Microsoft 365 CopilotGA
Website (chat widget)GA
Dynamics 365GA
Mobile apps (Bot Framework)GA
Facebook MessengerGA
WhatsAppGA

Governance

CapabilityWhat it enforces
Entra ID authenticationEvery agent interaction is authenticated - no anonymous access unless explicitly allowed
DLP policiesPrevent sensitive data flowing through connectors - enforced across all agents in the tenant
Environment isolationDev, test, production environments with separate data and settings
Audit logsAll agent interactions logged to Power Platform admin centre
Sensitivity labelsInherited from source data through knowledge sources
Admin centre controlsTenant admins can enable/disable features, set quotas, view all agents

Cost

Copilot Studio uses message-based pricing:

ComponentBilling
Standard messagesIncluded in M365 Copilot license or billed per message
Generative AI messagesHigher rate (LLM inference cost embedded)
Premium connectorsPer connector usage above included allowance
Power Automate flowsSeparate Power Platform licensing

Copilot Studio pricing | Copilot Studio portal


Getting Started

ResourceLink
Create your first agentQuickstart
Add knowledge sourcesKnowledge docs
Publish to TeamsPublish guide
AI capabilities overviewAI capabilities
Copilot Studio vs Foundry decisionDecision guide

Key Takeaways

  • Copilot Studio is the distribution layer, not the orchestration layer. Build complex agents in Foundry/MAF, then publish them through Copilot Studio to reach Teams and M365 users.
  • Computer-using agents (GA May 2026) are the standout capability for legacy system automation - interact with any UI without needing an API.
  • Use Foundry IQ as the knowledge source, not raw SharePoint indexing, when you need permission-trimmed retrieval with layout-aware chunking.
  • DLP policies are enforced across all agents in the tenant through the admin centre - governance is not opt-in per agent, it's tenant-wide.
  • Per-message pricing means conversational agents can get expensive at scale. Estimate average messages per session before committing.