Microsoft Copilot Studio

Copilot Studio is Microsoft’s tool for building custom AI agents. Instead of using Copilot as a general assistant, you build a named agent that answers from your own documents and takes actions in your own systems, then publish it to Teams, a website or another channel.

A flow follows steps written in advance and does the same thing every time. An agent reasons about a request and decides what to do next, which suits messy inputs like an email or a question typed into Teams. Most businesses end up using both, with the agent calling flows for the parts that have to be exact.

There are two parts: a per-user licence for the people who build and manage agents, and a tenant licence for the organisation itself. Usage is billed in Copilot Credits, which replaced the older per-message model on 1 September 2025.

Copilot Studio enforces purchased capacity monthly, and unused Copilot Credits do not carry over to the next month. If usage goes past what you have bought, Microsoft can deny service until more capacity is added. That is why we size capacity against real usage and monitor it, rather than setting it once and forgetting it.

Not necessarily. Copilot Studio can be licensed on its own for building agents. Whether you also want Microsoft 365 Copilot depends on whether you want the assistant working inside Word, Excel and Outlook too. We work through both in a readiness assessment rather than assuming you need everything.

Two ways, and you want both. Fix the permissions and oversharing in the content first, which is a Microsoft Purview job, then scope the agent’s grounding to only the sources it needs. An agent inherits the access of the person asking it, so tidy permissions do most of the work.

Your team does. We build agents in your tenant, document how they work, and hand over a register that names an owner and a review date for each one. We are not trying to make you dependent on us to change a prompt.

A contained first agent is usually a matter of weeks rather than months, provided the content it needs to read is in reasonable shape. Where projects blow out is almost always the content and permissions work underneath, not the agent build itself.