Choose Copilot Studio when a governed, low-code business-agent experience is the main requirement. Choose Microsoft Foundry when the application needs developer-controlled AI integration, model evaluation and an Azure operating model. Combine them only where a clear specialist boundary justifies the additional complexity. Azure AI Foundry is now Microsoft Foundry; the older name remains in this comparison because many teams still use it when evaluating the platform.

An enterprise agent is not merely a chat interface. It retrieves information, calls tools and may influence business decisions. The platform choice therefore affects who owns its behavior, how access is enforced, what is tested and how cost is controlled. A low-code demonstration can hide a privileged connection; a custom SDK implementation can hide missing review controls. Evaluate the whole service rather than treating low-code and code-first as a quality ranking.

Key takeaways

• Copilot Studio is not the same product or license as Microsoft 365 Copilot.
• Microsoft Foundry includes multiple model and agent development paths with feature-specific availability.
• Neither platform automatically makes private data retrieval permission-aware.
• Agent actions need explicit authorization, validation and appropriate review.
• Estimate credits, tokens, tools, retrieval and operating costs separately.
• A native agent connection described as preview must not be presented as generally available production functionality.

What is Microsoft Copilot Studio?

Copilot Studio provides a graphical, low-code environment for building and managing agents and workflows. Makers can define behavior, connect approved knowledge and tools, test interactions and publish to supported channels. Professional developers can contribute APIs and custom integration. Current documentation also covers evolving agent and workflow experiences, so verify the exact authoring mode and maturity rather than assuming every demonstrated capability is part of an established production path. See the Copilot Studio overview.

For example, an internal IT assistant might answer approved support questions and create a ticket after confirming the required fields. The agent’s connection must have only the permissions needed for that task. A policy document containing an instruction to disclose credentials is still untrusted content; it must not change the tool’s authority. The user should see whether the ticket was actually created or only drafted.

What is Azure AI Foundry, now Microsoft Foundry?

Microsoft’s AI platform terminology has evolved from Azure AI Studio through Azure AI Foundry to Microsoft Foundry. Use the current documentation and project type when designing a new solution; older hub-based or classic portal examples may have different resources and APIs. The Microsoft Foundry overview describes a platform grouping models, agents, tools and governance. It also explicitly notes that some capabilities are preview.

A Foundry application can use model endpoints directly, an appropriate managed agent path or hosted custom agent code. The application still needs a user interface, identity boundary, business API and operational ownership. A model catalog does not imply that every model, tool or customization method is available in every region or agent mode. Distinguish the platform resource, selected model, agent runtime and additional retrieval or integration services in both architecture and budget.

Feature comparison: 22 enterprise dimensions

The practical trade-off is responsibility. A managed authoring environment can reduce interface and orchestration work, but governance and testing remain necessary. A custom application can provide more control over data and behavior, but the team owns more of the engineering lifecycle. Neither is universally more secure, cheaper or more scalable; those outcomes depend on the selected configuration and service requirements.

Three reference architectures and their trust boundaries

A. Copilot Studio enterprise agent

Separate conversational identity from connection identity. If a tool uses an application account, enforce the user’s permitted operation at the service boundary. Restrict knowledge sources, record provenance and check sensitive-data behavior. A user must not receive a document simply because the agent’s backend account can read it. Validate handoff and outage behavior before enabling a critical process.

B. Foundry enterprise AI application

The application API establishes tenant and user context, applies rate limits and validates inputs. Retrieval applies the appropriate access filters before evidence is returned to the model. Tools expose narrow operations rather than broad database or administrative access. Treat the model’s proposed action as untrusted until schema, policy and authorization checks pass. Retain a correlation ID across the application, agent and downstream system.

C. Hybrid business agent and specialist service

Microsoft’s connected-agent overview identifies Foundry agent connections as public preview in the documented experience. Another Foundry connected-agent guide specifies a preview GitHub Copilot harness, the current Foundry portal, an enabled Activity protocol and appropriate access. Do not combine those pages into a claim that every Copilot Studio agent can connect to every Foundry project in production.

A hybrid should preserve authorization, conversation scope and tool boundaries across the handoff. Verify whether identity is delegated or an application credential is used; do not assume one is equivalent to the other. A documented, approved API integration can be assessed as an alternative, but it has its own engineering and support requirements. If one platform already meets the use case, adding another agent may only increase latency, failure points and cost.

Pricing: compare total workload cost, not starting prices

Copilot Studio uses Copilot Credits with relevant pay-as-you-go, capacity and pre-purchase arrangements. Applicable Microsoft 365 Copilot rights are a separate licensing question, not a universal substitute for all Copilot Studio usage. Consult the current Copilot Studio licensing guidance and product terms for the actual scenario. Do not assume a fixed credit cost per conversation; operations and features can consume usage differently.

Foundry cost depends on the selected model and service path. The Foundry Agent Service pricing overview currently distinguishes native prompt/workflow agents from hosted agents: it describes no additional charge for creating or running the native agent itself, while underlying model and tool charges remain, and hosted agents use container-compute billing. “No additional agent charge” does not mean the solution is free.

Include input and output tokens, selected retrieval/search, tools, storage, telemetry, networking and application hosting where applicable. Model rates and availability differ. The Microsoft Foundry pricing page directs buyers to selected service pricing and the calculator; displayed estimates are not a contract quote. Verify region, currency, agreement and feature status immediately before procurement.

Copilot Studio platform cost = sum(billable credits by operation x applicable rate)
  subject to purchased entitlement, capacity and current terms

Foundry platform cost = selected model inference + applicable agent hosting
  + retrieval/search + tools + storage + network + monitoring

Solution TCO = platform cost + application hosting + integration
  + evaluation + human review + security operations + maintenance
# Avoid double counting a service already included in another line.

Use measured pilot traces to estimate operation counts and token distributions. A short answer can still involve multiple retrievals and tool calls. Long conversation history and repeated retries can increase cost. Caching, smaller suitable models and narrower tool paths may help, but must preserve authorization, freshness and task quality. Budget alerts and quotas should be combined with an application response for exhausted capacity; an alert alone does not stop spending or provide a graceful user experience.

Ten enterprise use cases and platform starting points

These are hypothetical architectures, not client implementations. Select a platform after testing the hardest requirement: the restricted document, privileged action, unsupported connector or unusual failure. A successful general chat test does not establish suitability for any of these workflows. Define the denominator of a success metric and include abstentions, escalations and rejected outputs.

Security, retrieval and responsible operation

Establish Entra identities and least privilege, then verify environment or resource access independently from business authorization. Power Platform data policies and Azure network controls address different layers; neither automatically filters every retrieved record for the end user. Secrets belong in approved credential mechanisms, not prompts or repository configuration. Test cross-user and cross-tenant access deliberately.

Prompt injection can arrive in a document, tool response or user message. Treat each as untrusted input and keep action authority outside generated text. Restrict tools to narrow schemas, validate arguments and require approval where consequences justify it. Model refusal or content filtering is one layer, not the entire security design. Redact sensitive data from telemetry and establish retention and support access.

Neither platform automatically guarantees GDPR, HIPAA, SOC 2 or other organizational compliance. Confirm service availability, data boundaries, contracts and customer responsibilities for the jurisdiction and use case. Foundry’s agent service documentation describes multiple development paths and feature-specific capabilities. Verify the selected path rather than extrapolating a platform-level statement to every runtime and model.

Which should your organization choose?

Decision sequence: if the main need is a supported business-agent experience with governed makers, assess Copilot Studio first. If custom application and model engineering dominate, assess Foundry. If both have genuinely distinct responsibilities, evaluate a hybrid with a documented interface. If the process is deterministic, an ordinary workflow or API may be simpler than an agent. Do not add autonomous planning where a reviewed form and rule solve the problem.

Eight steps from pilot to production

1. Define a bounded task, user and success measure.
2. Assess authorized sources, integrations and consequences of error.
3. Select the smallest architecture that meets those needs.
4. Establish identities, permissions, network boundaries and review authority.
5. Build a controlled proof with representative permitted data.
6. Evaluate task quality, unsafe actions, access control, latency and cost.
7. Release through versioned configuration and approval with rollback.
8. Monitor outcomes, exceptions and spend; re-evaluate changes before scaling.

A release dataset should include unanswered questions, malicious source instructions, revoked access, unavailable tools and ambiguous requests. Test the full user journey rather than only isolated model responses. Maintain a fallback or human handoff for unsupported tasks. Record model, instructions, retrieval settings and tool versions so a regression can be investigated.

Frequently asked questions

Technical references

• Copilot Studio overview
• Microsoft Foundry overview
• connected-agent overview
• Foundry connected-agent guide
• Copilot Studio licensing guidance
• Foundry Agent Service pricing overview
• Microsoft Foundry pricing page
• agent service documentation

Technical source review: 1 November 2026. Illustrative architectures, assumptions and examples are not client implementations or measured production results. Deployment-specific verification is required.