Copilot Studio vs Azure AI Foundry: Which Agent Platform, What It Costs, and When to Bring Us In

Both platforms reach the same frontier models: Copilot Studio uses Microsoft's hosted models and can call models you deploy in Foundry through its bring-your-own-model option for prompts, while Foundry exposes the Azure OpenAI catalog and other model families directly. The real questions are who builds it, where its data lives, which channel the users are in, who governs it, and how the meter runs.

A Copilot Studio agent is built by a power user or Power Platform developer inside the tenant's existing environments, data policies and Microsoft 365 permissions. A Foundry agent is built by a developer with an Azure subscription, a landing zone and a pipeline. The first reaches a useful version in days; the second takes weeks and carries operations for life.

Copilot Studio is a maker tool. An agent is instructions, knowledge (SharePoint, files, websites, Dataverse, Microsoft Graph), topics for guided conversations, and tools: connectors, prompts, agent flows, MCP servers and other agents. It publishes to Teams, Microsoft 365 Copilot, the web and custom channels.

Licensing has three parts. The maker license, the Microsoft Copilot Studio User License, is $0.00. Capacity is a tenant subscription, Microsoft Copilot Studio, sold as packs of 25,000 Copilot Credits per month (Microsoft's Copilot Studio product page, opened 2026-10-11); the pack is $2,400 per year on an annual commitment, or $200 per month (Microsoft list price, October 2026 price list). The third is the pay-as-you-go meter, a per-credit rate billed through an Azure subscription linked to the environment, alone or as overflow above the packs; Microsoft states there is no feature difference between the pack and the meter, only how you pay. That meter is Microsoft's metered charge and it is always yours.

Microsoft Learn's Copilot Studio billing table (checked 2026-10-11) rates a generative answer at 2 credits, an agent action or flow call at 5, and tenant Graph grounding at 10. A help desk agent answering 2,000 questions a month with a generative answer and one connector action each uses about 14,000 credits, inside one pack; add Graph grounding to every answer and it is 34,000. Run your numbers in the Copilot Studio estimator before buying a pack.

One exemption matters. Microsoft's product page states that for users licensed for Microsoft 365 Copilot, $30.00 per user per month on an annual term ($360.00 per year, Microsoft list price, October 2026 price list), usage of agents published to Microsoft 365 Copilot is included in their license. If most users hold that license, an employee-facing agent may need no capacity pack; the Copilot licensing guide covers the seat side.

Foundry is a developer platform, and Foundry Agent Service is its managed runtime for agents. Microsoft's documentation describes it this way: pick the model from the catalog, attach knowledge tools (Azure AI Search, Fabric, SharePoint, Bing grounding) and action tools (Azure Functions, OpenAPI, Logic Apps, MCP servers), give the agent a managed identity, and run it inside your own virtual network when needed. An agent can be published to Microsoft 365 Copilot and Teams from the Foundry portal, which provisions an Azure Bot Service resource and an Entra application, or through the Microsoft 365 Agents Toolkit.

There is no license to quote. Microsoft's Foundry Agent Service pricing page states that there is no additional charge for creating or running Foundry-native agents; you pay for model token consumption through Foundry Models, for the tools and knowledge connections the agent uses, and for hosted agent compute if you bring your own framework. All of it is Azure consumption on your subscription, metered by Microsoft, and always the customer's charge. Our Foundry work is priced for the build; the Azure bill is separate and sized with you before the first deployment.

That is the real difference: a Copilot Studio pack is a fixed monthly number, while a Foundry agent's cost moves with tokens per request, requests per day and the model you chose. Budget it like any Azure workload, with a secure AI landing zone (from $9,500 per project, 4 weeks) for keys, networking and logging.

Copilot Studio wins when the agent is for your own employees, the knowledge already sits in SharePoint, Dataverse or Microsoft Graph, the channel is Teams or Microsoft 365 Copilot, the actions are existing or custom connectors, and the people who will maintain it also maintain your Power Automate flows. Help desk, HR policy, sales enablement and phishing triage agents are Copilot Studio builds in our projects.

Foundry wins when you need a specific or fine-tuned model, when the agent is part of your product or a customer-facing site, when it must run inside a private network or a regulated Azure environment, when you need code-level control over tool calls and evaluation, or when volume is high enough that a per-token bill on a smaller model costs less than Copilot Credits.

The two platforms are designed to stack. Copilot Studio's bring-your-own-model option lets a prompt inside a Copilot Studio agent call a model you deployed or fine-tuned in Foundry, so a maker-built agent in Teams uses your domain model for one step while the rest stays low-code. In the other direction, a Foundry agent exposed through MCP becomes a tool or connected agent that a Copilot Studio agent hands work to. Identity, channel and data policies stay in Copilot Studio; the hard reasoning and the code live in Foundry.

This is our default when a Microsoft 365 Copilot tenant has one problem low-code cannot solve alone, such as contract analysis or a pricing engine. Both meters run, credits on one side and tokens on the other, so estimate them separately.

Whichever platform you choose, three controls come first.

Microsoft Purview: sensitivity labels and oversharing cleanup before any agent reads SharePoint, so retrieval cannot surface a document the user could not open themselves; our Purview before Copilot article gives the order.

Power Platform data policies: Microsoft's documentation (April 2026) describes how data policies cascade to every environment in the tenant, classify connectors, and set a violating connection to a disabled state; it also notes that the Copilot Studio controls are evolving into their own governance rules. Give agents their own environment with a policy listing only the connectors they may use.

Microsoft Entra Agent ID: Microsoft's documentation (April 2026) describes it as the identity framework that gives agents their own directory identities, with Conditional Access, Identity Protection, lifecycle management and audit logs applied to them, on Microsoft and non-Microsoft platforms. Our agent governance article explains the inventory problem; the AI Agent Inventory and Lifecycle Governance Implementation ($4,950 per project, 3 weeks) builds it, and the Copilot Studio Agent Security Review ($2,950, 2 weeks) tests one agent the way an attacker would.

Microsoft announced Microsoft Agent 365 on 18 November 2025 as the control plane for AI agents: a registry of every agent in the tenant, access control, visualization of what agents touch, interoperability across Microsoft, open-source and third-party agents, and security through Defender, Entra and Purview. It is not a third build surface. Copilot Studio and Foundry are where an agent is made; Agent 365 is how an estate of them is inventoried, owned, monitored and retired. The Agent 365 subscription is $15 per user per month on an annual commitment, $180 per year (Microsoft list price, October 2026 price list, listed since 1 May 2026).

The questions it answers become manual investigations as soon as a second team builds agents: which agents exist, who owns each one, which knowledge sources and actions they use, which are exposed to broad audiences, what happens when an owner leaves, and which still depend on a retired flow or site. Bring it in early when the risk is sprawl: several builder teams, several environments, regulated processes, or actions that change business records.

Until then, keep an operating record for every production agent, whichever platform built it: owner, sponsor, purpose, approved audience, data classification, knowledge sources, connected systems, action permissions, monitoring plan, review cadence and retirement criteria. If writing that down feels excessive, the agent is not ready for production; it is closer to an app than to a document. Our AI Agent Inventory and Lifecycle Governance Implementation ($4,950 per project, 3 weeks) builds the record and the review cadence.

The first agent should prove trust, not coverage. A workable candidate is IT access-request triage in Teams: it answers access-policy questions, collects the required details, checks the request against approved patterns, creates or updates the ticket, and escalates exceptions to a human approver. Knowledge is limited to the access policies and the service catalog, identity comes from Entra ID, Power Automate handles only ticket creation, status lookup and routing, and every access change needs a human approval.

• Week 1: pick the one workflow, define the audience, list the approved knowledge sources, audit their permissions, and decide which actions are allowed.
• Week 2: build in Copilot Studio in a dedicated environment, not the default one; connect only the approved sources; implement one or two low-risk actions; apply the data policy and solution-based deployment.
• Week 3: test with real users and attack the predictable failures: prompt injection, ambiguous requests, attempts at restricted content, stale policy text and action misuse. The Copilot Studio Agent Security Review ($2,950 per project, 2 weeks) does this against your agent.
• Week 4: publish to a pilot security group, read the telemetry, tune the responses, complete the operating record, and decide to expand, hold or retire.

Measure more than deflection: completed requests, escalation accuracy, abandoned conversations, wrong-answer reports, policy gaps and repeated clarifying questions. Do not start with a company-wide knowledge agent over every SharePoint site; that turns unknown content hygiene into searchable incidents.

Is Copilot Studio included with Microsoft 365 Copilot?

Building an agent needs the Copilot Studio user license, listed at $0.00. Running it needs capacity, except that Microsoft's guidance states agents published to Microsoft 365 Copilot are included for users who hold that license, subject to fair use. Unlicensed users and other channels consume credits.

Do we need Azure for Copilot Studio?

Only for pay-as-you-go billing, which links an Azure subscription to the environment. Prepaid packs alone need no Azure subscription. The pay-as-you-go credits setup is included with IT Partner licensing.

Which one is cheaper?

For an employee-facing agent on Microsoft 365 data, usually Copilot Studio, and it may be covered by Microsoft 365 Copilot seats. For high-volume or customer-facing agents, a Foundry agent on a smaller model can cost less per interaction. Estimate both.

Is Agent 365 a replacement for Copilot Studio?

No. Copilot Studio builds an agent; Agent 365 is the control plane that registers, secures and monitors agents from Copilot Studio, Foundry and third parties. A first agent needs Copilot Studio and an operating record; Agent 365 earns its place when several teams are building.

Which agent should we build first?

A narrow, measurable workflow with low-risk actions and clearly owned knowledge: IT access-request triage, HR policy questions or service desk intake. Prove trust with a pilot group before any agent that reads broad SharePoint content.

• Microsoft Copilot Studio — Microsoft product page, opened 2026-10-11 (packs of 25,000 Copilot Credits per month; pay-as-you-go meter; agent usage included for Microsoft 365 Copilot users)
• Microsoft Copilot Studio licensing guidance — Microsoft, opened 2026-10-11
• Copilot Studio billing rates and management — Microsoft Learn, checked via search 2026-10-11 (2, 5 and 10 credit rates)
• Microsoft Agent 365: The control plane for AI agents — Microsoft AI at Work blog, 18 November 2025, opened 2026-10-11
• Microsoft Learn, "What is Microsoft Entra Agent ID?", ms.date 14 April 2026, and "Data policies" (Power Platform), ms.date 7 April 2026, read from the MicrosoftDocs GitHub sources in October 2026
• Microsoft Azure pricing, "Foundry Agent Service pricing", October 2026 (no additional charge for Foundry-native agents; tokens, tools and hosted compute billed as Azure consumption)
• Microsoft, US commercial price list, October 2026 (Microsoft 365 Copilot list price)