How to implement Microsoft Copilot Studio: licensing, preparation and a plan for your first agent

Microsoft Copilot is a ready-made assistant: it writes emails, summarises meetings and searches documents. Copilot Studio is for something else – you use it to build your own agent for a specific process. The agent answers questions based on selected documents and can carry out actions in systems, such as submitting a request or checking an order’s status.

Copilot Studio makes most sense in a company that already works on Microsoft 365, keeps its documents in SharePoint and uses Teams. The agent then uses the same accounts and permissions as employees. You will find typical first uses in the list below.

It is not always a good choice. If a process follows fixed rules and does not require language understanding, ordinary automation in Power Automate will be cheaper and more reliable. If the company works on Google Workspace, it is better to look for tools from that environment.

With Copilot Studio you do not pay a licence for each agent user. You pay in credits for the agents’ work: for answers, searches in documents and actions in systems. The prices in the table below are Microsoft list prices in US dollars, as of October 2026.

How many credits a single conversation uses depends on its complexity. An answer from a simple, pre-scripted scenario costs less than an answer generated from documents or an action carried out in a system. Microsoft provides a calculator where you enter the type of agent and the expected traffic.

Credits from a pack expire at the end of the month. If the company uses more than it bought, Microsoft may restrict the agents. That is why it is more sensible to pay for usage during the pilot and choose a pack after a few weeks of measurement.

Configuring the agent itself takes less time than you might think. Most of the work goes into preparing the process, the documents and the permissions. The list below lets you check whether your company is ready.

If one of the points is not met, that does not rule out the project. But it must go into the schedule as a separate task with a named owner.

Every agent in Copilot Studio has four layers. The table below shows them using the example of an agent that answers employees’ HR questions.

The order matters. Publishing in Teams is the last step, not the first. If you do not take care of knowledge and boundaries first, users will get a nice chat window on top of a poorly prepared process.

A pilot of the first agent usually takes 8–12 weeks. Most of that time goes into preparing knowledge and testing, not into work in Copilot Studio itself. The plan below assumes the twelve-week option.

Our case study of a knowledge assistant for production engineers, built in Copilot Studio and made available in Teams, shows what such an agent looks like in practice.

Before launch, prepare a set of test questions with model answers, e.g. 50–100 questions collected from the department that answers them today. The set should cover typical questions, difficult ones and ones the agent should not answer.

Copilot Studio has a built-in automatic evaluation of the agent against a set of questions, but the process owner has the final say. They decide whether the quality is good enough for launch. Repeat the same set after every major change to instructions or documents.

By default, a new agent requires sign-in with a company account in Microsoft Entra ID. However, the person building the agent can choose the no-sign-in option. Then anyone with the link can talk to the agent and ask about the documents connected to it.

These risks are limited with data protection policies in the Power Platform admin center. Microsoft’s documentation calls them data policies, and previously DLP policies (data loss prevention). They let you block publication without sign-in, selected knowledge sources and selected channels.

A separate decision concerns AI models. In Copilot Studio you can choose Anthropic models (Claude), but in the EU they are switched off by default. Once enabled, data is processed outside Microsoft’s EU Data Boundary, so it is worth making this decision together with your data protection officer.

Launching the agent is the start of the work, not the end of the project. Copilot Studio shows conversation statistics, and credit usage is visible in the Power Platform admin center. It is worth reviewing this data every week during the first few months.

The most valuable item is the list of questions the agent could not answer. It usually points to a missing or outdated document rather than an error in the agent. Correcting a document often achieves more than changing the instructions.

Remember the AI Act too. If the agent talks to customers, from 2 August 2026 it must tell them they are dealing with AI (Article 50).

These mistakes recur regardless of industry and company size. Almost none of them stem from the limitations of the tool itself.