The short answer
Microsoft Copilot Studio is a tool for building your own AI agents without writing code. You choose where the agent gets its knowledge, such as SharePoint or your website, you decide what it is allowed to do, and you deploy it in Teams, in Microsoft 365 Copilot or as a chat window on your site.
Until late 2023 Copilot Studio was called Power Virtual Agents. It belongs to Microsoft’s Power Platform, alongside Power Automate and Power Apps, and it is aimed at IT teams and citizen developers: staff without a programming background who want to automate processes in their organisation.
For what an AI agent is and how it works under the hood, read what are AI agents. This article is about Microsoft’s agent builder. Looking for the AI assistant for programmers? Read what is GitHub Copilot.
What you build in Copilot Studio
An agent in Copilot Studio is made of a handful of building blocks that you put together in a web interface.
Instructions
You describe in plain language what the agent does, who it serves and what tone it uses. Copilot Studio drafts a first setup from that.
Knowledge sources
You connect SharePoint sites, websites, individual files or Dataverse tables. The agent searches them and bases its answers on what it finds.
Topics
For fixed conversations, like a leave request or a password reset, you define step by step what the agent asks and answers.
Actions
Through Power Automate and more than a thousand ready-made connectors, the agent can do things: open a ticket, update a record, send an email.
Channels
You publish the same agent to Teams, Microsoft 365 Copilot, your website or other channels, and manage it in one place.
Besides agents that wait for a question, you can build autonomous agents. They start on their own when something happens, such as a new email in a shared inbox or a new file in a folder, and handle it without anyone talking to them.
How to build an AI agent in Copilot Studio
You build an agent that answers questions from your own documents in five steps.
Describe what it should do
On copilotstudio.microsoft.com you create a new agent and describe in plain language who it is for and which questions it answers. Copilot Studio uses that to fill in the name and the first instructions.
Connect the knowledge
Add the SharePoint site, web pages or files that hold the answers. Start with one source that is well maintained.
Give it actions, read-only first
Pick a connector or a Power Automate flow for what it needs to do. Only let it create or change anything once its answers are right.
Test with real questions
In the test pane, ask the questions colleagues actually ask, including the vague and the wrong ones. Check which source each answer comes from.
Publish and keep watching
Put the agent in Teams and Microsoft 365 Copilot, or as a chat window on your site. When the analytics show questions it cannot answer, add to its knowledge.
The clicking is quick. The time goes into the knowledge sources: outdated or contradictory documents produce outdated or contradictory answers, however well the agent is set up otherwise.
How much does Copilot Studio cost?
Microsoft charges by usage, in Copilot Credits. Every answer and action costs a number of credits depending on what the agent does: a fixed scripted reply costs less than an answer the AI composes from your documents. You buy credits in monthly packs or pay afterwards for what you use through Azure.
With a Microsoft 365 Copilot licence, staff can use some agents inside their own environment without separate credits. Agents on your website or for external users are always billed by usage. Prices and rules change often, so check Microsoft’s website for current figures.
Run the numbers up front. An agent that handles thousands of conversations a month is cheap per credit and can still cost more per year than you expect.
Where Copilot Studio is strong
You already run on Microsoft 365
Sign-in, permissions and security run through your existing Microsoft accounts. The agent only sees what the user is allowed to see.
A quick first version
An agent that answers questions about your staff handbook or IT procedures can be live in Teams in an afternoon.
No developer to maintain it
A functional admin can update text, knowledge sources and simple steps on their own.
Governance
IT decides centrally which connectors are allowed, who can publish agents and where the data stays.
When a custom AI agent fits better
Copilot Studio is a construction kit, and every kit has a shape. As long as your agent fits that shape, you move fast. Outside it, every step gets more expensive. If your data has to stay with a European supplier, you can build a custom agent on the models of Mistral, the company behind Le Chat.
| Copilot Studio | Custom AI agent | |
|---|---|---|
| Environment | Strong inside Microsoft 365, Teams and Dataverse. | Works with any system that has an API, inside or outside Microsoft. |
| Logic | Fixed steps and simple decisions in a visual editor. | Complex rules, calculations and edge cases, in code you can test. |
| Model | The models Microsoft offers, with limited say over how they are used. | You pick the model per task and switch when a better one arrives. |
| Cost | Cheap to start, pay per conversation as you grow. | Higher build cost, then only the cost of the model and hosting. |
| Audience | Staff inside your own organisation. | Customers too, as part of your own product or service. |
The better your process fits the shape of Copilot Studio, the faster and cheaper your agent goes live.
In practice we see three signs that custom is the better choice:
Your key systems live outside Microsoft
Your planning runs in your own application, your customer data sits in a CRM without a good connector, or you rely on an older system with its own kind of integration.
The agent has to calculate or decide
Drafting quotes with your pricing rules, assessing claims or scheduling with dozens of constraints. In a visual editor that quickly turns messy and hard to test.
Your customers will use the agent
For an agent on your website or in your app, per-conversation costs add up, and you want full control over how it behaves and what it logs.
For what goes into building your own agent, read building an AI agent. If you are unsure whether an agent is the right answer at all, AI agents: when to use them helps.
Conclusion: start in Copilot Studio when it fits
If you work in Microsoft 365 and want an agent for your own staff, try Copilot Studio first. You get something working quickly and learn what you actually need. When the agent has to reach systems outside Microsoft, run serious logic or become part of what you deliver to customers, a custom agent is often cheaper and more reliable over time.
Copilot Studio = fast inside Microsoft
Knowledge from SharePoint, actions through Power Automate, and the agent lives in Teams.
Custom = any integration, any rule
Your own systems, your own logic and the model that best suits the task.
What is Copilot Studio?
How do you build an AI agent in Copilot Studio?
What is the difference between Copilot and Copilot Studio?
How much does Copilot Studio cost?
Is Copilot Studio the same as Power Virtual Agents?
Can Copilot Studio connect to systems outside Microsoft?
When should you choose a custom AI agent?
Copilot Studio or custom: what fits your process?
We look at the process, the systems and the volume with you, and tell you honestly which route costs least and works best. If custom wins, our senior developers build the agent.