Copilot Studio vs Azure AI Foundry, Agent Builder or Power Automate

Copilot Studio vs Azure AI Foundry (now Microsoft Foundry), Agent Builder and Power Automate: who builds, where it runs, how much code, data, governance and cost.

Azure + Power PlatformChoose a tool12 min read

Checked against Microsoft Learn 22 Oct 2026 · 15 sourcesPosted by the Maker DeskHow we write guides

Copilot Studio and Azure AI Foundry, now named Microsoft Foundry, both build AI agents, but for different builders. Copilot Studio is a low-code service that makers use to build agents and publish them to Teams, Microsoft Copilot and websites, paid for in Copilot Credits. Foundry is an Azure platform where developers pick from thousands of models, write code if they want to, and pay Azure for each model call. Agent Builder is the quick option inside Microsoft Copilot, and Power Automate is for fixed steps that need no AI judgement at all.

Note

Checked against Microsoft Learn on 22 October 2026, and against Microsoft's Copilot Studio pricing page.

Which Microsoft tool to build with, in three questions. First: does the job need AI judgement, or are the steps fixed? If the steps are fixed, use a Power Automate cloud flow, licensed per user or per process. If it needs judgement: is it an agent for you or a small team that answers from Microsoft 365 content you can already see? If yes, use Agent Builder in Microsoft Copilot, included with a Microsoft Copilot licence. If no: do developers need their own code, a model of their choice, or Azure networking? If yes, use Microsoft Foundry, with prompt or hosted agents, paid per model call on Azure. If no, use Copilot Studio: low-code agents with connectors and channels, paid in Copilot Credits. A note: the answers combine; a Copilot Studio agent can call a Foundry agent or model, and a flow can call a Copilot Studio agent. — Three questions that pick Power Automate, Agent Builder, Copilot Studio or Microsoft Foundry.

The names first

Half the confusion is naming, and the old names are still everywhere:

Microsoft Foundry was called Azure AI Foundry, and before that Azure AI Studio. Microsoft Learn now uses the new name. Some Copilot Studio screens and connectors still say Azure AI Foundry, such as the Connect to Azure AI Foundry step and the Azure AI Foundry Agent Service connector. They mean the same product.
Agent Builder is also called Copilot Studio lite. It lives inside the Microsoft 365 Copilot app and Teams, not in Copilot Studio.
Microsoft Copilot is the 2026 name for Microsoft 365 Copilot, the assistant people chat with. It's what agents from all three builders can appear in.
Agent flows in Copilot Studio are deterministic flows billed in Copilot Credits. They're close to Power Automate cloud flows but aren't the same thing; cloud flow, agent flow or desktop flow compares them.

Side by side

	Power Automate	Agent Builder	Copilot Studio	Microsoft Foundry
Who builds	Makers	Information workers	Makers and developers, low-code	Developers, from no-code prompt agents in the portal to full code
How much code	None needed	None	None needed; REST APIs, custom connectors and MCP servers for the edges	Optional: prompt agents need none; hosted agents run your own code and frameworks
Thinks or follows steps	Follows the steps you design	Answers questions	Both: topics and flows for fixed paths, generative orchestration and reasoning for the rest	Reasons with the model and tools you choose
Where it runs	In the cloud, on a trigger or a schedule	Microsoft 365 Copilot app and Teams	Teams, Microsoft Copilot, websites, apps, and more channels by harness	Your own apps through an endpoint, and Microsoft Copilot and Teams once published
Data and grounding	Power Platform connectors to read and write data; no grounding	SharePoint, Copilot connectors and the web, within each user's existing permissions	SharePoint, websites, files, Dataverse, Copilot connectors, Azure AI Search, plus connectors as tools	Foundry tools and knowledge, Azure AI Search, MCP and OpenAPI tools, Azure Functions and Logic Apps
Models	None needed	Microsoft Copilot's own	Copilot Studio's list, or your Foundry model for prompts	10,000+ models from Microsoft, OpenAI, Anthropic, Meta and others
Governance	Power Platform admin center, environments, data policies	Microsoft 365 admin center	Power Platform admin center: environments, data policies, solutions for ALM	Azure: Microsoft Entra ID, role-based access, content filters, network isolation, Azure Policy
Cost model	Per user (Premium) or per process (Process licence)	Included with a Microsoft Copilot licence; otherwise Copilot Credits	Copilot Credits, paid by the tenant	Azure consumption: model tokens, tool use, and container compute for hosted agents

Microsoft's own architecture guidance frames the first choice as where the agent is hosted, not which designer you prefer. The host decides the orchestration, the models available and the running cost; the authoring tool is secondary.

When to use each

Power Automate: when the steps are fixed

If you can draw the process as a flowchart, with no question that needs reading or judging, you don't need an agent. A cloud flow does the same thing every time, and a flow with no AI steps spends no Copilot Credits.

Users with only a Microsoft 365 licence can use standard connectors. Premium connectors need Power Automate Premium for the user, or a Process licence assigned to the flow, which gives it 250,000 actions a day whoever runs it.

Many processes have one fuzzy step in a fixed flow, such as reading an email and deciding how urgent it is. Keep the flow, and let it call a Copilot Studio agent for that one step.

Agent Builder: answers for you and your team

Agent Builder suits a project FAQ, a product-manual helper or a team onboarding agent: knowledge your colleagues can already open, for a small audience. You describe the agent in natural language in the Microsoft 365 Copilot app on the web (microsoft365.com/chat) or in Teams on desktop and web. It isn't available on mobile.

What it can't do matters as much:

It can't call other systems. There are no connectors, actions or workflows.
Agents built in it can't be used in Teams group chats or one-to-one chats.
There's no versioning and no deployment pipeline: development, test and production environments are a Copilot Studio feature.

When you outgrow it, Copy to Copilot Studio carries the agent's instructions and configuration across. It's included with a Microsoft Copilot licence, and free for agents that ground only on the web.

Copilot Studio: agents for a wider audience

Copilot Studio is the default for an agent that a department, the whole organisation or customers will use, especially when it has to act: create a ticket, update a record, start an approval. It brings Power Platform's connectors, agent flows, event triggers for autonomous agents, external channels such as a website or phone line, and solutions to move agents between environments.

On the GitHub Copilot harness, you build an agent on one page, its Build tab. What the page shows, as of October 2026:

Build, Preview, Evaluate and Monitor: tabs to configure the agent, test it in a chat, run test sets, and review how it does once people use it.
Instructions: the agent's job, tone and limits, written in plain language.
The configuration panel: Model, Triggers, Skills, Tools, Knowledge, Connected agents and Memory.
Publish: deploys the agent to its channels.

The Build tab of a Purchase Request Agent in Copilot Studio, built on the GitHub Copilot harness. Along the top are the agent's name; the Build, Preview, Evaluate and Monitor tabs, with Monitor greyed out; save, share, feedback and more icons; and a Publish button. On the left are the agent's title and its Instructions: collect the item, cost, cost centre and reason one question at a time; check the request against the Purchasing Policy and the team's remaining budget; explain who approves it; create the purchase request record; and email the employee a summary. The instructions also say never to approve a request and to send requests over $5,000 to Finance. On the right is the configuration panel. Model is set to Claude Opus 4.7. Triggers lists Chat with the agent in Teams and Chat with the agent in Microsoft Copilot. Skills lists Budget Check. Tools lists Purchase requests in Dataverse and Send email in Outlook. Knowledge lists Purchasing Policy and Approved Suppliers in SharePoint. Connected agents is empty, and Memory is On. — Recreated by LowCodeStacks from Microsoft Learn screenshots (CC BY 4.0): the Build tab of a Copilot Studio agent on the GitHub Copilot harness, interface as of October 2026, with made-up names and instructions.

The tenant pays in Copilot Credits: in October 2026, Microsoft's pricing page lists $200 a month for a pack of 25,000 credits, with pay-as-you-go through Azure as the alternative. People with a Microsoft Copilot licence can build and use internal agents on the Copilot chat and standard harnesses at no extra cost. What is Copilot Studio explains harnesses and channels, and Copilot Studio licensing and Copilot Credits the rates.

Microsoft Foundry: when developers need control

Choose Foundry when a developer team owns the agent and needs things a low-code service doesn't give:

Its own code. A hosted agent runs your container or code (Microsoft Agent Framework, LangGraph, the OpenAI Agents SDK and others) on Foundry-managed compute with its own Microsoft Entra identity.
Model choice and quota. Deploy the model you want from the catalog, pick the deployment type, and manage its quota and content filters yourself.
Azure networking and controls. The standard setup supports private networking, role-based access and Azure Policy, for workloads that must stay inside your Azure estate.
An agent inside your own product, called through an endpoint or SDK from Python, C#, JavaScript or Java.

You pay Azure for what runs: model input and output tokens for every agent call, code interpreter per session, file search by the vector storage used, and container compute for hosted agents. Microsoft's Agent Service FAQ lists no extra charge for the service on top of these. Microsoft Foundry from zero sets up the resource, a model and the right roles. If you learned Azure OpenAI first, Azure OpenAI to Microsoft Foundry maps the renamed parts.

Warning

The no-code prompt agent in the Foundry portal can look like a shortcut around Copilot Studio. It still lives in Azure: someone needs an Azure subscription, Foundry roles and cost management, and publishing it to Teams also needs permission to create an Azure Bot Service resource.

Use them together

The tools aren't rivals so much as layers, and Microsoft documents several ways to combine them:

Pattern	How
A Foundry model inside a Copilot Studio agent	Connect your Foundry deployment to a prompt tool; the agent's main model still comes from Copilot Studio's list. See a Foundry model in Copilot Studio prompts
A Foundry agent as a specialist	Add it as a connected agent: in preview on the standard harness, and supported on the GitHub Copilot harness. On the standard harness it must be built in the new Foundry portal and have its Activity protocol endpoint turned on
An Azure AI Search index as knowledge	Add an Azure AI Search index as a knowledge source (standard harness only)
A flow around an agent	A Power Automate flow calls the agent for the judgement step; the agent calls agent flows or workflows as tools
Start small, then grow	Build in Agent Builder, then Copy to Copilot Studio when it needs actions or a wider audience
A Foundry agent where people work	Publish it from Foundry straight to Microsoft Copilot and Teams; organisation-wide publishing needs approval in the Microsoft 365 admin center

The hybrid is Microsoft's own suggestion: Copilot Studio's low-code front end over Foundry's models handles harder cases without a full pro-code build.

Run short, time-boxed prototypes before committing. Microsoft suggests one to two weeks per option, then stick with low-code if it meets the functional and security requirements.
Decide where the agent is hosted before you choose the designer.
Keep deterministic steps in flows or code, where they cost no model calls and give the same answer every time.

Don't

Don't build an agent for a process a flow can do.
Don't choose Foundry only for a model: Copilot Studio prompts can already use your Foundry deployment.
Don't leave an Agent Builder agent serving a whole department; it has no environments or deployment pipeline.

Sources

Was this helpful?

Comments

No comments yet. Share what worked for you.