Copilot Studio vs Microsoft 365 Copilot Agent Builder vs Foundry: where to build each enterprise agent
Microsoft offers three places to build agents and they overlap. A decision guide: what each one can and cannot do, how each is paid for, and five questions that place an agent in the right tool.

A team asks for “an agent that answers questions about our travel policy.” Another asks for “an agent that opens a purchase request in the ERP when a manager approves it.” A third wants “an assistant inside our customer portal, with our own model fine-tuned on claims data.” All three are agents, and at Microsoft they belong in three different places.
Microsoft sells three overlapping ways to build agents: Agent Builder inside Microsoft 365 Copilot, Copilot Studio on Power Platform, and Microsoft Foundry on Azure. Choosing well saves money and months; choosing badly means rebuilding the agent when it outgrows the tool, or paying for a platform the use case never needed.
What each one is for
Agent Builder is the quickest path. Any Microsoft 365 Copilot user can describe an agent in natural language, point it at SharePoint content and Copilot connectors, add instructions and skills, test it and share it. It builds declarative agents that run inside Microsoft 365 Copilot. Its limits are explicit in Microsoft’s documentation: it cannot author actions that call external services; agents cannot be used in Teams group or one-to-one chats; customer-managed keys are not supported. When an agent needs actions, connectors or workflows, Microsoft’s instruction is to copy it to Copilot Studio.
Copilot Studio is the low-code platform for agents that act. It adds connectors, custom connectors and MCP tools, agent flows and workflows, multiple channels including websites and customer-facing voice, three harnesses with different reasoning and billing, and the lifecycle machinery of Power Platform: environments, data policies, solutions and pipelines. It serves business makers and professional developers on the same platform.
Foundry is the pro-code platform: custom models and fine-tuning, model choice across providers, complex orchestration in code, and products that engineering teams build, test and ship like any other software. Microsoft also sells its own agent loop there, the Copilot harness in Foundry.
| Agent Builder | Copilot Studio | Foundry | |
|---|---|---|---|
| Who builds | Any Microsoft 365 Copilot user | Business makers and developers | Engineering teams |
| Actions in other systems | No, copy to Copilot Studio | Connectors, MCP, agent flows | Anything you code |
| Where it runs | Inside Microsoft 365 Copilot | Teams, Microsoft 365, web, apps, voice | Any application you build |
| Lifecycle | Sharing and admin approval | Environments, solutions, pipelines | Your software delivery process |
| How it is paid | Mostly within the Microsoft 365 Copilot license | Copilot Credits, some included for licensed users | Azure consumption per model and service |
How the money differs
The cost difference is larger than the feature difference. An Agent Builder agent’s generative answers are not charged as long as they do not use tenant graph grounding, according to Microsoft’s rate card notes. A Copilot Studio agent consumes Copilot Credits per answer, action and flow step, with many employee-facing uses included for Microsoft 365 Copilot users acting with their own identity, but not autonomous or customer-facing use. Foundry bills Azure consumption for models, search and compute, plus whatever you build around them. Our Copilot Studio pricing guide has the full rate card.
Five questions that place an agent
1. Does it only answer, or does it act? Answering from SharePoint and connected content: start in Agent Builder. Creating records, sending requests, updating systems: Copilot Studio or Foundry.
2. Who uses it, and where? Employees inside Microsoft 365 favor Agent Builder or Copilot Studio. Customers on a website, a portal or a phone line rule out Agent Builder and make Copilot Studio’s external channels or a Foundry application the candidates.
3. Who will maintain it? If a business team owns the process and will keep changing it, a low-code tool they can edit is worth more than elegant code they cannot touch. If an engineering team owns it, Foundry fits their tools and reviews.
4. How much control over the model and the loop do you need? Bringing your own or a fine-tuned model, custom orchestration, unusual latency or evaluation requirements push toward Foundry. Copilot Studio offers model choice and harness choice, but within its platform.
5. What does failure cost? We argued that the platform should follow the risk tier: the more an agent can change in business systems, the more lifecycle control it needs. A personal helper can live in Agent Builder. An agent that commits transactions needs environments, testing, approvals and monitoring, which Copilot Studio and Foundry provide in different ways.
Common mistakes
Building in Copilot Studio what Agent Builder would do. A policy Q&A agent for employees who already have Microsoft 365 Copilot rarely needs connectors, flows or a credit budget.
Building in Agent Builder what will need actions next quarter. If the roadmap includes “and then it files the request,” plan the move to Copilot Studio now, before users depend on the first version.
Building in Foundry to avoid governance. Code does not exempt an agent from identity, data and approval controls; it only means you have to build them yourself.
Choosing per company instead of per agent. Most enterprises will use all three. The useful standard is a short rule that sends each new agent to the right one, which is the first thing we help teams write. For the bigger picture of what Copilot Studio contains, see our Copilot Studio guide.