The short answer
Copilot Studio is the low-code builder for business makers. Azure AI Foundry, which Microsoft now brands as Microsoft Foundry, is the pro-code platform for engineering teams. Copilot Studio gets a working agent live in hours with Microsoft 365 integration and central governance, priced tenant-wide through Copilot Credits at $200.00 per 25,000 credits per month. Foundry gives developers model choice, prompt versioning and evaluation gates, and the Agent Service itself carries no additional charge, though you pay for model tokens and every tool the agent invokes. Pick by who is building and how large the knowledge base is.
Last updated: July 2026. Sources: Microsoft Copilot Studio pricing page, Azure Foundry Agent Service pricing page, Microsoft technical community guidance.
This question comes up about a week into every Microsoft AI project. Someone builds a proof of concept in Copilot Studio, it works, and then a developer asks why they are not doing this properly in Foundry. Both tools genuinely can build an agent. They are aimed at different people solving different sized problems, and picking the wrong one is expensive in different ways.
Here is the practical split, including the naming change that makes half the documentation you find confusing.
Is Azure AI Foundry the same as Microsoft Foundry?
Yes, and the rename is worth knowing before you start reading docs. The platform was Azure AI Studio, then became Azure AI Foundry, and Microsoft has since been consolidating the branding toward Microsoft Foundry, with pricing pages now living under both names. The underlying platform is continuous. If a tutorial says Azure AI Studio, it is old but usually still directionally right. If it says Azure AI Foundry and you land on a Microsoft Foundry page, you are in the right place.
This kind of mid-cycle rename is now normal across the enterprise AI category, and it is the main reason procurement documents go stale so fast. Check the publication date on anything you use to make a decision.
What is the difference between Copilot Studio and Azure AI Foundry?
Copilot Studio is a managed, low-code environment aimed at business makers. You describe what the agent should do, attach knowledge, wire in Power Platform connectors, publish to Teams or a website, and govern it centrally alongside every other agent in the tenant. The Microsoft 365 integration is native and the guardrails are built in. A capable non-developer can ship something useful the same week.
Foundry is a pro-code platform aimed at engineering teams building AI as a product or a core capability. You get model choice across the catalog, fine-grained control over temperature and top-p, prompt versioning, evaluation gates, and the ability to scale retrieval far past what Copilot Studio's managed knowledge handles. You also get the obligations that come with a real platform: someone owns the deployment, the evals, the cost and the on-call.
| Dimension | Copilot Studio | Foundry (Azure AI Foundry) |
|---|---|---|
| Built for | Business makers, low code | Developers and ML engineers, pro code |
| Time to first agent | Hours | Days to weeks, depending on scope |
| Model control | Managed defaults | Model catalog, temperature, top-p, prompt versioning, evaluation gates |
| Knowledge scale | Comfortable for modest document sets, strains on large libraries | Built for large corpora such as technical manuals and regulatory libraries |
| Microsoft 365 integration | Native, plus Power Platform connectors | Available through tools and connections you configure |
| Pricing shape | Tenant-wide, $200.00 per 25,000 Copilot Credits per month, or pay as you go | No charge for the Agent Service itself; you pay model tokens plus every tool and knowledge connection invoked |
| Who operates it after launch | The maker, with central admin governance | An engineering team, permanently |
How much does Azure AI Foundry Agent Service cost?
Microsoft's pricing page states there is no additional charge to use Foundry Agent Service itself. That headline is true and it is also the most misread sentence in Microsoft's AI pricing. You pay for everything the agent touches: model token consumption through Foundry Models, and separate charges for tools and knowledge connections such as Azure AI Search, SharePoint grounding, Fabric, Bing grounding, Logic Apps and Functions.
So Foundry looks free and bills like infrastructure. Copilot Studio looks expensive up front and bills like a capacity plan. Which one is actually cheaper depends entirely on volume and on how chatty your agent is, and neither can be forecast honestly before you have real traffic. Instrument early, and set budget alerts on day one rather than in month three.
When should I use Copilot Studio instead of Foundry?
Use Copilot Studio when the agent is a business process wrapped in conversation, the knowledge base is modest, the audience is internal Microsoft 365 users, and the person who understands the process is not a developer. That combination is common and Copilot Studio handles it well. Central governance across many maker-built agents is a real strength, and it is the thing that keeps a hundred departmental agents from becoming a shadow IT problem.
Use Foundry when the agent is part of a product you ship, when retrieval quality on a large corpus is the hard problem, when you need to evaluate and version prompts like code, or when you need a model Copilot Studio does not expose. Regulated industries with large document libraries usually end up here whether they planned to or not.
Many enterprises run both, and that pattern is worth naming because it is the one Microsoft's own guidance points at: developers build specialized capability in Foundry, makers assemble the user-facing experience in Copilot Studio, and the two talk to each other. It works, and it also means two teams, two cost models and two governance surfaces.
Do I need either of these to search my company's content?
This is the question worth asking before the build starts, because a large share of Copilot Studio and Foundry projects begin as retrieval projects. Someone says people cannot find answers, someone else says we can build an agent for that, and six weeks later a team is tuning chunk sizes.
If the requirement is genuinely bespoke, custom actions, a workflow that ends in something being done, a domain-specific evaluation harness, then build it, and pick the tool from the table above. If the requirement is that an employee asks a question in plain language and gets a trustworthy answer from company content, that is a solved product category and building it yourself is the expensive path. You will spend the first month on connectors, the second on permissions, and the third discovering that permission changes need to propagate faster than your reindex schedule.
Structured data is the one place the build-versus-buy line moves, because questions like "what did revenue do last quarter by segment" want a system that can turn a plain-English question into a query against your warehouse rather than a document retriever. Unstructured knowledge, which is most of what people actually search for, does not need a bespoke platform.
What this means for your shortlist
Copilot Studio and Foundry are both good at what they are for. Neither is an enterprise search product, and neither reaches outside Microsoft's boundary without work you own. If your knowledge lives in Slack, Notion, Google Drive, Gmail, Jira and Confluence as much as it lives in SharePoint, that work is the project, not a detail of it. We break the licensing and cost side down further on the Copilot Studio alternative page and the Microsoft Copilot pricing breakdown.
InSearch takes the other route. One query across Drive, Slack, Notion, Confluence, Gmail, Jira and Salesforce, on day one, with no build phase. A written answer with inline citations to the exact source document, so people verify rather than trust. Item-level permissions inherited from each app and enforced live at query time, so a revoked permission takes effect immediately rather than at the next sync. One per-seat price you can budget before you start. Try a question in the box above.