All insights

Copilot Agents · 6 min read

What Is a Copilot Agent? A Plain-English Definition

By James Wilkinson 3 August 2026

A plain-English definition of Microsoft Copilot agents: what they are made of, the jobs they do well, where they run and how firms go from one agent to a team.

TL;DR
  • A Copilot agent is Copilot given one defined job, with plain-English instructions and named knowledge sources such as a SharePoint site.
  • Agents answer from your own content, run inside your own Microsoft 365 tenant and only show people what they can already see.
  • A first agent takes about 15 minutes to build in plain English. The Agent Journey maps the five stages from there to a team of agents.

A Copilot agent is Microsoft 365 Copilot given one defined job. It has a name, instructions written in plain English and named knowledge sources, such as a SharePoint site, and it answers questions about that job from your organisation’s own content rather than the open internet. Agents run inside your own Microsoft 365 tenant, inherit the permissions you already have and work with agreed human review points. They assist staff. They do not replace them.

That is the whole definition. The rest of this article unpacks it: what an agent is made of, what it is not, what the first one takes to build and how firms get from one agent to a team of them.

The anatomy of an agent

Every Copilot agent is three things.

A name and a job. The useful agents have named jobs. An agent that answers HR policy questions. An agent that drafts replies in the company tone of voice. An agent that summarises incoming post for the office team. Not an assistant for everything, but a colleague’s worth of help with one thing.

Instructions. Written in plain English: what the agent is for, who it serves, how it should answer and what it should refuse or hand to a person. Instructions shape behaviour. They are not a security control, which is why the permissions layer below matters more than clever wording.

Knowledge sources. The SharePoint sites, document libraries or files the agent is allowed to draw on. This is called grounding: the agent answers from that content and can point to the document an answer came from, so you can check it. A grounded agent on a tidy, current site gives answers people trust. The same agent pointed at an old file dump does not.

What an agent is not

An agent is not Copilot itself. Microsoft 365 Copilot is the broad assistant across Outlook, Word, Excel and Teams. An agent is that capability narrowed until it becomes dependable for one job.

An agent is not an automation flow. If a process is rule-based and should run silently in the background, Power Automate is usually the better tool. The choice between them is a big enough question that we keep a separate guide to Copilot Studio vs Power Automate.

And an agent is not a member of staff. It prepares, finds, drafts and chases. Judgement, advice and sign-off stay with people, at review points agreed before the agent goes live.

Your first agent takes about 15 minutes

The fastest way to understand agents is to build one. In Copilot Chat, the agent builder lets you describe an agent in plain English: tell it the job, point it at a SharePoint site and it writes the instructions for you. This kind of agent is called a declarative agent, which means instructions plus knowledge and no code.

From a standing start to a working agent takes about 15 minutes, and it is how we close our training sessions: every attendee leaves with a real agent built and running in their own tenant. Watching your own agent answer from your own documents changes what people think agents are.

When an agent needs to act rather than just answer, such as raising a ticket or starting an approval flow, it moves into Copilot Studio territory. The agent builder vs Copilot Studio guide covers that line in detail.

The jobs that work

Agents earn their keep on jobs that repeat. The pattern across sectors is consistent: internal knowledge, guided intake, first drafts and structured chasing. In an accountancy practice that looks like document intake, client onboarding, engagement letter drafting and records chasing, which is exactly what the AI agents for accountancy firms page walks through job by job.

The common thread is that each agent has one job, one audience and one set of knowledge sources, plus a named owner who keeps it current.

Where agents run and what they can see

Agents run inside your own Microsoft 365 tenant. Your content is not used to train the underlying models, it does not leave the Microsoft 365 service boundary and an agent can only show someone what they are already allowed to open, because it inherits your existing permissions. That is the short answer, and the full security explainer walks through where the data goes and what your admins control.

From one agent to a team of agents

A first agent is a capability, not a destination. The Agent Journey maps five stages: Use an agent, Build your first agent, Team agents, Copilot Studio and A team of agents. At every stage there are two doors: learn to do it yourselves through training, or have it built for you and handed over documented.

If you are trying to place your own team on that map, a short call is usually enough.

Related reading

More on copilot agents

Copilot Agents Copilot Agent Builder vs Copilot Studio: Which Do You Need? The plain-English difference between the agent builder in Copilot Chat and Copilot Studio, and the one question that decides between them. Copilot Agents Microsoft's Agent Governance Toolkit: Three Questions to Ask of Any Agent Microsoft's open-source Agent Governance Toolkit enforces agent rules in code, not prompts. Why that matters for firms that will never install it. Copilot Agents Are Copilot Agents Secure? Where Your Data Goes The long answer on Copilot agent security: where answers come from, what stays inside the Microsoft 365 service boundary and what your admins control. Capability map The Agent Journey Five stages from using your first Copilot agent to running a team of agents, with two doors at every stage. Agents Microsoft Copilot agents Plain-English guidance on where Microsoft Copilot agents fit, how to govern them and when to build. Agent builds AI agents for accountancy firms AI agents for UK accountancy firms, built in Copilot Studio: document intake, onboarding and records chasing, in your own Microsoft tenant, from £5,000. Service Microsoft 365 Agents Training A plain-English introduction to agents, use cases and responsible adoption. Service Copilot Studio Agents Custom agents grounded in the right knowledge and built for defined jobs. Next step Book a free consultation Share where you are now and what you want Microsoft 365 to help with next.

Common questions

Questions about Copilot agents

What is the difference between Copilot and a Copilot agent?
Microsoft 365 Copilot is the general assistant that works across Outlook, Word, Excel and Teams and answers almost anything. A Copilot agent is Copilot narrowed to one defined job, such as answering HR policy questions, with its own name, instructions and knowledge sources. The narrowing is the point. Because the agent only draws on agreed content, its answers are more consistent and easier to trust for that job.
Do you need to be technical to build a Copilot agent?
No. A declarative agent is described in plain English in the agent builder inside Copilot Chat and grounded on a SharePoint site in about 15 minutes. Anyone who can write a clear email can do it. Copilot Studio agents, which can take actions, need maker-level skills that teams learn through training or bring in.
Are Copilot agents secure?
Agents run inside your own Microsoft 365 tenant and inherit your existing permissions, so people only ever see content they can already open. Your data is not used to train the underlying models and it does not leave the Microsoft 365 service boundary. No third-party AI tools are involved.
Do Copilot agents replace staff?
No. Agents take on defined jobs such as answering policy questions, preparing drafts or chasing missing information, and they run with agreed human review points. The aim is fewer repeated interruptions and less manual admin, not fewer people.