Copilot Studio vs Azure AI Foundry: When to Keep an Agent, When to Move It

A team lead asks one question in a planning meeting and the room splits three ways. Do we build this agent in Copilot Studio, or move it to Azure AI Foundry? Both are official Microsoft build paths in 2026. Both land in the same tenant-wide agent inventory. The wrong pick costs a rebuild once the project outgrows it, and now it costs money every month it runs on the wrong platform.

This is the decision guide. It answers when to keep an agent in Copilot Studio, when to move it to Foundry, and when to retire it.

The two platforms do different jobs

Microsoft’s Cloud Adoption Framework frames the build options as service tiers. Copilot Studio is SaaS. Azure AI Foundry is PaaS. Self-hosted SDK agents are IaaS. The first two are managed by Microsoft, and they suit different agents.

Copilot Studio is the front door. It is a low-code, fully managed platform inside the Power Platform. Makers build agent experiences without touching infrastructure, hosting, or governance. It connects through prebuilt and custom connectors and extends Microsoft 365 Copilot with enterprise data. It is strong for fast, business-facing copilots and lightweight automation.

Azure AI Foundry Agent Service is the engine room. It is a managed platform for building, deploying, and scaling production agents. It runs your containerized code with a dedicated endpoint, its own Microsoft Entra Agent ID, autoscaling, and built-in observability. It grounds agents in enterprise knowledge through Foundry IQ across SharePoint, Fabric, and Bing, connects to a large catalog of tools through Foundry Toolboxes, and gives every agent managed memory. It is built for the custom, multi-step, mission-critical agents that Copilot Studio cannot fully control.

Both paths can publish an agent back to Microsoft Teams and Microsoft 365 Copilot, so end users reach the same agent in the same place regardless of where it runs.

The cost models diverge, and that matters

The platforms bill in different ways, and the difference drives the decision.

Copilot Studio agents meter Copilot Credits at $0.01 each, billed by task complexity, with cost that climbs on grounding, reasoning, and tool calls. (For the full rate card and why one agent can cost 100x another, see why your Copilot bill went variable.) Foundry-native agents carry no per-credit charge for prompts and workflows. You pay Azure consumption for model tokens, tools, and hosted compute, and hosted agents scale to zero when idle. For a high-volume, reasoning-heavy agent, the Foundry model usually costs less and stays predictable under Azure Cost Management. For a low-volume scripted bot, Copilot Studio often stays cheaper.

Cost is not the only axis. Foundry adds private networking, a first-class agent identity, content safety, tracing, and evaluation that Copilot Studio does not expose. For a regulated or data-sensitive workload, that governance is the reason to move, before cost even enters the conversation.

What a “custom engine agent” is

When you move an agent to Foundry and surface it in Microsoft 365 Copilot, it becomes a custom engine agent. A declarative agent uses Copilot’s own orchestration and models. A custom engine agent brings its own orchestration, its own model choice, and its own integrations, with full control over the workflow. You can publish a Foundry agent to Microsoft 365 directly, which provisions the bot and Entra identity for you, or connect it through a proxy app built with the Microsoft 365 Agents Toolkit for deeper customization. Both routes connect Foundry Agent Service to Copilot and Teams through a bot layer, so users interact with the Foundry agent inside Microsoft 365.

The decision framework

Run each agent through four outcomes:

Keep in Copilot Studio. Low volume, scripted or lightly generative, cheap to run, and business-owned. Migration payback would take too long. Leave it where it is and keep it as the front door.

Move to Azure AI Foundry. High volume, reasoning-heavy, tool-calling, or mission-critical. These agents drive the Copilot Credits bill and need real cost control, governance, and observability. Moving them lowers and stabilizes cost and anchors the workload to Azure.

Move for governance, regardless of cost. Any agent touching regulated data, needing data residency, private networking, or a first-class agent identity. Foundry gives you the controls a compliance review asks for.

Retire. Duplicate or unused agents. They recover spend at zero migration effort. Every estate has them.

Most enterprises run all four in one estate. The work is deciding which agent is which, with the real cost and usage numbers in front of you. The Copilot to Azure AI Foundry migration guide covers how the move runs once you have made the call.

FAQ

Should I move all my agents from Copilot Studio to Foundry? No. Keep lightweight, low-cost agents in Copilot Studio and move the heavy, reasoning-driven, or regulated agents to Foundry. The right answer is per agent, not per estate.

Is Azure AI Foundry cheaper than Copilot Studio? For high-volume reasoning agents, usually yes, because Foundry-native prompts and workflows carry no per-credit charge and hosted agents scale to zero when idle. For small scripted bots, Copilot Studio often wins. Model both before deciding.

What is the difference between a declarative agent and a custom engine agent? A declarative agent uses Copilot’s built-in orchestration and models. A custom engine agent brings its own orchestration, model, and integrations, and can run on Azure AI Foundry while still surfacing inside Microsoft 365 Copilot and Teams.

Do users notice when an agent moves to Foundry? No, when it is done right. The agent republishes to Teams and Microsoft 365 Copilot through a stable endpoint, so the experience stays the same.

---

Not sure which of your agents to keep, move, or retire? Onepane runs a two-day Agent Readiness Assessment that classifies every agent with real cost and usage data, then hands you a before-and-after model and a migration plan. Microsoft funding covers most of the work. Model your ROI or book a demo.