Microsoft Copilot Studio: How to Build Custom AI Agents (2026 Guide)
Best for organizations building B2E agents quickly without deep developer resources. For complex RAG pipelines or custom ML, move to Azure AI Foundry.
Microsoft Copilot Studio is the platform 230,000+ organizations are using to build custom AI agents, and the gap between it and Power Automate is simpler than you think. Power Automate executes rules. Copilot Studio reasons. That distinction determines which tool to reach for on any given project. This guide covers what Copilot Studio actually is in 2026, how the Copilot Credits pricing model works, what you can build, how connectors and MCP connect your data, and a five-step deployment process that gets a working agent into Teams or SharePoint without writing a line of code.
Adoption figure: Microsoft FY2026 Q2 Letter to Shareholders (vendor-sourced).
What Is Microsoft Copilot Studio?
Microsoft Copilot Studio is a SaaS low-code agent platform. As of 2025 it is no longer just a chatbot builder. It builds three distinct agent types:
Conversational agents handle chat-based question answering and guided workflows. Think IT helpdesk bots that surface answers from a knowledge base, or onboarding assistants that walk new hires through HR policies step by step.
Autonomous agents run in the background without requiring a human to initiate each step. They execute multi-step workflows, manage approval chains, and only escalate when exceptions fall outside defined parameters; the standard cases run without human involvement.
Computer-Using Agents (CUAs) use vision and reasoning to interact directly with Windows desktop applications when no API is available. This is preview functionality as of September 2025, useful for legacy tooling that predates modern API design.
What ties all three together is generative orchestration. Rather than following a hardcoded decision tree, the AI decides which plugins to call, which follow-up questions to ask, and how to format output based on the context of each interaction. The default model since October 2025 is GPT-4.1. GPT-5 is available in preview but is not the default.
Platform Milestones: 2025-2026
Sources: Microsoft Learn Copilot Studio documentation, Microsoft FY2026 Q2 Letter to Shareholders.
Copilot Studio Pricing: Copilot Credits Explained
Copilot Studio pricing changed in September 2025. The old "messages" model is gone. Everything now runs on Copilot Credits.
What a Credit measures: Credits measure the time and compute an agent needs to retrieve information, respond to prompts, and use actions or custom skills. The amount consumed per interaction depends on task complexity: a simple FAQ retrieval costs fewer credits than an autonomous workflow that calls three external APIs and writes to a database.
Pay-as-you-go: An Azure meter option is available with no upfront commitment. You pay only for credits consumed. Useful for pilots or unpredictable usage.
M365 Copilot licensed users: If your users are already licensed for Microsoft 365 Copilot, building and running internal agents in Teams and SharePoint is zero-rated; no credits consumed for those users on those internal channels.
When you need the standalone license: The $200/month pack or pay-as-you-go is required for external channels (public websites, apps, WhatsApp), non-licensed user access, and agent usage outside the M365 zero-rated channels.
Model Cost Context: Claude in Copilot Studio
Claude Sonnet 4.5 and Claude Opus 4.5 are available in Copilot Studio for custom agents, labeled as experimental. If you choose one of these models, be aware of the credit cost difference at Microsoft Foundry rates:
- Claude Sonnet 4.5: approximately 1.5-2x the credit cost of GPT-5.1 (per million tokens, Foundry rates)
- Claude Opus 4.5: approximately 7-10x the credit cost of GPT-5.1 (per million tokens, Foundry rates)
Context required: The 7-10x Opus figure is the Studio/Foundry token cost ratio for Claude Opus versus GPT-5.1. It is not the GitHub Copilot premium request multiplier, which is a separate 3x figure in GitHub's billing model. These are two different products with different billing mechanisms. Do not conflate them.
Sources: Microsoft Copilot Studio pricing page, Practical365 industry analysis (Foundry rates).
What You Can Build: Agent Types in Detail
The three agent types are not just categories. Here is what each looks like deployed in production.
Conversational: IT Helpdesk Bot
Load 300 knowledge articles into Studio. The agent handles common support requests (password resets, VPN setup, software access) instantly and without opening a ticket. Employees get answers in Teams without waiting in a queue. Only questions the agent cannot confidently answer get escalated. Most IT teams can deploy this in days. The ROI is immediate and measurable from week one.
Autonomous: Expense and Reimbursement Workflow
An employee submits an expense report. The autonomous agent triggers automatically, validates each line item against HR policy rules, routes approved items to the relevant SaaS accounting tool, and escalates only exceptions: missing receipts, categories exceeding policy thresholds. Standard cases require zero human review. This is what "background agent" means in practice: the full workflow loop runs without a human in the chain for routine submissions.
Customer Service: Real-Time Order Status
The agent pulls live order data from a CRM or ERP system, resolves status inquiries instantly, and closes tickets without handoff. The agent is only as useful as the data it can reach, which is where connectors and MCP become critical.
Procurement: Vendor Quote Analysis
A procurement agent reviews incoming vendor quotes, compares them against historical purchasing data stored in Dataverse or a connected ERP, recommends preferred suppliers, and drafts purchase orders for human sign-off. This is a multi-step autonomous workflow spanning data retrieval, analysis, and document generation: not a single action, but a full process.
Multi-Agent Systems
Agents can collaborate. A research agent gathers data, passes structured output to an analysis agent, which produces a summary formatted by a document agent for distribution. Copilot Studio supports multi-agent orchestration for workflows that exceed what a single agent handles cleanly.
Source: Microsoft Learn Copilot Studio documentation.
Connectors, MCP, and Data Sources
An agent is only useful if it can reach the right data. Copilot Studio connects to data through three mechanisms: pre-built connectors, MCP servers, and Dataverse.
Pre-Built Connectors
1,000+ connectors are available out of the box. The list covers most enterprise tooling: Salesforce, ServiceNow, SAP, Adobe, Dynamics 365, SQL, Oracle, Workday, Zendesk, Jira, Box, Dropbox, Google Drive, and hundreds more. For most organizations, the connector you need already exists. Connector-based tools support per-action Data Loss Prevention (DLP) blocking; admins can allow or block specific connector actions at a granular level.
Model Context Protocol (MCP)
MCP support was added in October 2025. MCP servers enable dynamic, real-time connections to files, database records, and non-Microsoft APIs that do not have a pre-built connector. If a connector does not exist, MCP is the path to custom integrations.
MCP governance gap (read before enabling): When you allow an MCP server in Copilot Studio, you allow access to the full tool surface on that server. There is no per-tool DLP blocking for MCP connections. Contrast this with connector-based tools, which support per-action DLP at a granular level. If your governance requirements demand per-action control over what an agent can do with a data source, use connectors rather than MCP for that connection.
Dataverse
Dataverse is the structured data backbone underneath Copilot Studio. It stores agent configurations and logs conversation transcripts. Transcript retention defaults to 30 days; this is not permanent storage. The retention period can be extended, and transcripts can be exported to Azure Data Lake for longer-term analysis. Do not assume transcripts are available indefinitely in the default configuration.
Authentication
Copilot Studio supports Microsoft Entra ID SSO with SAML and OIDC. The SSO Consent Card (July 2025) allows users to authenticate external actions inside the chat window without leaving to a browser redirect. Entra agent identities (GA December 2025) enable automatic machine-to-machine authentication for multi-agent systems.
Source: Microsoft Learn Copilot Studio documentation.
Access & Prerequisites: Getting Into Copilot Studio
Before you can build anything, you need to actually reach the Studio dashboard. This is where many admins and developers get stuck -- particularly in enterprise tenants with strict governance policies. If you navigate to copilotstudio.microsoft.com and see a "Start a trial" or sign-up prompt instead of the authoring canvas, one of the following is blocking you.
Four Valid Paths Into Copilot Studio
Microsoft documents four configurations that grant access (Microsoft Learn: Copilot Studio Licensing). You need exactly one:
- Microsoft 365 Copilot license -- assigned to your user account ($30/user/month). Includes Studio access for internal agents at no extra cost.
- Copilot Studio User License -- free, but requires your organization to already own a tenant-level Copilot Studio prepaid Credit pack subscription.
- Copilot Studio authors role -- assigned via an Entra ID security group in the Power Platform admin center.
- Copilot Studio trial -- individual sign-up for testing. You can build and test agents but cannot publish them.
Common Blockers (Why You See "Start a Trial")
copilotstudio.microsoft.com with your regular M365 account -- the one that has Copilot in Word, Teams, and Outlook. Use your admin account only for governance tasks in the M365 Admin Center.admin.microsoft.com → Users → Active Users → [your account] → Licenses and apps → Microsoft 365 Copilot. If the checkbox is unchecked, that is your problem.admin.powerplatform.microsoft.com → Manage → Environments. If empty, a Global Admin or Power Platform Admin must create a default environment before Studio will load.admin.microsoft.com → Settings → Integrated Apps → locate Copilot Studio. If it is blocked or your group is not in the allowed list, the admin must update this setting.allowedToSignUpEmailBasedSubscriptions policy in Microsoft Entra ID (configured via Microsoft Graph PowerShell). When this is set to false, the trial fallback path is dead -- even when your license should grant access but is not being recognized correctly, the system cannot create the trial enrollment that would let you in. Check: Run Get-MgPolicyAuthorizationPolicy in Graph PowerShell and look for AllowedToSignUpEmailBasedSubscriptions = False. Your admin can resolve this by directly assigning Studio access rather than relying on self-service.Scope distinction: Your M365 Copilot license includes Copilot Studio for internal agents only (Teams, Copilot Chat, SharePoint). Internal agent interactions are zero-rated -- no Copilot Credits consumed. Publishing to external channels (websites, mobile apps, WhatsApp, social) requires a standalone Copilot Studio subscription ($200/month per 25,000 Copilot Credits or pay-as-you-go via Azure meter). (Microsoft Learn)
Diagnostic Commands
If you have admin access, these PowerShell commands identify which specific blocker is active in your tenant:
# Check if self-service signups are blocked
Import-Module Microsoft.Graph.Identity.SignIns
Connect-MgGraph -Scopes "Policy.Read.All"
Get-MgPolicyAuthorizationPolicy | Select-Object AllowedToSignUpEmailBasedSubscriptions
# Check self-service purchase policies for all products
Get-MSCommerceProductPolicies -PolicyId AllowSelfServicePurchase
# Verify a user's license includes M365 Copilot
Get-MgUser -UserId "user@yourdomain.com" -Property DisplayName,AssignedLicenses |
Select-Object DisplayName,AssignedLicensesRequires Microsoft.Graph and MSCommerce PowerShell modules. Minimum permissions: Policy.Read.All, User.Read.All. Sources: Microsoft Learn: Assign licenses, Microsoft Learn: Block self-service sign-ups.
How to Build Your First Agent (5-Step Deployment)
This is the deployment sequence that works for most first-time Studio projects. Steps 1 and 2 build the agent. Steps 3 and 4 make it safe to deploy. Step 5 gets it in front of users and measures what happens.
Publishing Channels
Copilot Studio supports the following deployment targets: Microsoft Teams, SharePoint, Outlook, Microsoft 365 Copilot Chat (internal channels, zero-rated for M365 licensed users), public websites, mobile apps (iOS and Android), custom web portals, and WhatsApp (added July 2025). The client SDK enables embedding in Android, iOS, and Windows native apps.
Source: Microsoft Learn Copilot Studio documentation.
Copilot Studio vs Power Automate: When to Use Which
The clearest way to frame the difference: Power Automate knows what to do. Copilot Studio decides what to do.
Power Automate is a static rules engine. You define: if A happens, trigger B. If B produces result C, do D. The logic is linear, deterministic, and explicit. It is excellent for predictable, repeatable automation where every step and condition is known at design time.
Copilot Studio uses generative AI orchestration. The AI dynamically decides which plugins to call, what follow-up questions to ask, and how to format output, based on the context of each specific interaction. It handles non-linear, adaptive logic where the right path through a workflow is not known until runtime.
They are not competitors; they complement each other. Copilot Studio can trigger Power Automate flows for traditional background process steps inside a larger agent workflow. The agent handles the reasoning and user interaction; Power Automate handles the deterministic backend.
Copilot Studio vs Azure AI Foundry
Use Copilot Studio when the use case is well-scoped, the knowledge base is under approximately 500 documents, rapid deployment matters, and the team has low-code skills.
Move to Azure AI Foundry when the logic requires conditional branching and custom error handling, the retrieval requirement is complex RAG at enterprise scale, the team needs access to 1,600+ models (GPT, Llama, Mistral, or bring-your-own), or the workflow requires custom ML pipelines and full SDK/CLI control.
| Capability | Copilot Studio | Power Automate | Azure AI Foundry |
|---|---|---|---|
| Logic type | AI orchestration (adaptive) | Rules-based (deterministic) | Code-first (full control) |
| Target user | Low-code builders, IT admins | Process owners | Developers, ML engineers |
| AI reasoning | Yes Yes | No No | Yes Yes |
| Connectors | 1,000+ pre-built | 1,000+ pre-built | Custom via SDK/API |
| Best for | Agents, chat, knowledge retrieval | Linear workflow automation | Complex RAG, custom ML |
| Time to deploy | Days | Hours | Weeks |
| Knowledge base | ~500 docs optimal | N/A | Enterprise-scale |
| Model selection | GPT-4.1 default; GPT-5 preview; Claude experimental | N/A | 1,600+ models (GPT, Llama, Mistral, BYOM) |
Sources: Microsoft Learn Copilot Studio documentation, Microsoft Copilot Studio pricing page.
Frequently Asked Questions
Copilot for M365 is a productivity layer for individual users inside Office apps; it summarizes meetings, drafts emails, and generates content inside Word and PowerPoint. Copilot Studio is a platform for organizations to build custom AI agents. One is a consumer-facing productivity tool; the other is a developer and admin platform for building AI systems that serve an organization's users or customers.
$200 per month for 25,000 Copilot Credits (prepaid, tenant-wide license). A pay-as-you-go Azure meter option is also available with no upfront commitment. M365 Copilot licensed users get internal agent use in Teams and SharePoint zero-rated; no credits consumed for those users on those channels. External channels, public websites, and non-licensed users require the standalone license or pay-as-you-go.
Power Automate runs rules-based linear workflows: if A happens, do B. Copilot Studio runs AI-orchestrated non-linear workflows: the agent reasons about what to do, decides which tools to call, asks follow-up questions when needed, and adapts its output to context. Copilot Studio can trigger Power Automate flows for traditional background processes; the tools complement each other inside a single solution.
Yes. Claude Sonnet 4.5 and Claude Opus 4.5 are available in Copilot Studio for custom agents, labeled as experimental. Claude Opus costs approximately 7-10x more in Copilot Credits than GPT-5.1, measured at Microsoft Foundry rates per million tokens. Claude Sonnet costs approximately 1.5-2x more. This is the Studio token cost ratio; it is not the same as the GitHub Copilot premium request multiplier (3x), which applies to a separate product and a different billing model entirely.
Video Resources
Video resources will be added after publication. Check back for curated Copilot Studio tutorials covering agent authoring, connector setup, and governance configuration.