How to build an agent in Copilot
How to Create an Agent in Copilot (and the GCSE Prompt Method Behind It)
I am not an engineer. I run marketing. I mention that because my first Copilot agent took about twenty minutes to put together, and the hardest part was not the technology. It was deciding what to actually ask it to do.
Yes, Copilot is on a lot of people’s agendas, but have they actually sat down and made an agent yet?
If your team answers the same question every week, this is for you.
If your SharePoint holds the answers but nobody can find them, this is for you.
If you bought Copilot licences and the enthusiasm quietly faded around month two, this is for you.
If someone has asked you what a Copilot agent even is, start with the next section.
We recorded a webinar covering prompt engineering and agent building inside Copilot, start to finish. It is here if you would rather watch than read: https://www.youtube.com/watch?v=Jc4y_q6Dp0w
What is a Copilot agent?
A Copilot agent is a version of Microsoft 365 Copilot that you have given a specific job, a fixed set of instructions and a fixed set of knowledge sources. Instead of starting from a blank box each time, your team opens the agent and gets an answer built the same way, from the same approved information, for whoever is asking.
It is the difference between telling a new starter to write a quote and handing them your quoting process.
What do Copilot agents actually do?
The useful ones are narrow. A few that work well in real businesses:
A HR agent pointed at your handbook and policy library, so holiday and sickness questions stop landing in one person’s inbox.
An onboarding agent that knows your new starter process, your tools and who to ask about what.
A bid agent pointed at your past tender responses, so nobody rewrites the same company overview for the fourth time this quarter.
A first-line IT agent pointed at your own internal guides, answering the printer and password questions before they become tickets.
None of those are clever. That is why they work.
Microsoft’s own figures show active agents across the Microsoft 365 ecosystem grew 15 times year over year, and 18 times in large enterprises. The businesses seeing that growth are not building one enormous agent. They are building small ones that do one job properly.
What is the GCSE prompt method?
GCSE stands for Goal, Context, Source and Expectations. It comes from Microsoft’s own prompting guidance, which sets out four parts to a good prompt: the goal, the context, the expectations and the source. A clear goal is the only part you strictly need. You add the others when you want something more specific back.
| Part | The question it answers | What it looks like |
|---|---|---|
| Goal | What do you want back? | Summarise the last two weeks of the Riverside project emails |
| Context | Why do you need it and who is it for? | It is for Monday’s call with their operations manager |
| Source | What should Copilot use? | The Riverside channel in Teams and the project plan in the Projects SharePoint site |
| Expectations | How should the answer look? | Five bullets, written for a non-technical reader, flag anything uncertain |
Stitched together, that becomes:
“Summarise the last two weeks of emails and chats about the Riverside project and pull out anything that affects the go-live date. It is for Monday’s call with their operations manager. Use the Riverside channel in Teams and the project plan in the Projects SharePoint site. Give me five bullets, written for a non-technical reader, and flag anything you are not confident about.”
Now compare that with “summarise the Riverside project”. One of those answers you can send. The other one you delete.
Why the GCSE method matters so much in Copilot
Copilot is wired into your business data. That connection is the reason you are paying for it, and the Source element is the thing that switches it on. Leave the source out and Copilot falls back on general knowledge, which produces exactly the bland, could-be-anyone output that makes people stop opening it.
The other three each earn their place. Goal stops the rambling. Context is what makes the tone right for a client rather than a colleague. Expectations are what save you the rewrite, because you asked for five bullets rather than receiving eleven paragraphs.
When an answer disappoints, the missing ingredient is usually obvious. No source produces generic output. No context produces the wrong tone. No expectations produce the wrong shape.
Why GCSE is the foundation of a good agent
This is the part most people skip past, and it is the whole reason the webinar covers prompting before it covers agents.
When you create an agent, you write a GCSE prompt once and lock it in for everybody.
The instructions field inside an agent is a permanent prompt. Every question anyone asks runs through it. A vague instruction does not cost you one weak answer; it costs you every answer that agent gives to every person who uses it, for as long as it exists. A well-written one does the reverse. Your team gets your output standard without ever learning to prompt.
That is the actual return on this. Prompting stops being a skill you have to train into forty people and becomes something you build into a tool once.
How to create an agent in Copilot, step by step
Agent Builder sits inside Microsoft 365 Copilot. It runs in the browser, needs no code, and can draw on SharePoint, Teams, Outlook and OneDrive.
1. Pick one job worth repeating
Choose a question your team answers over and over. Narrow beats ambitious every single time. If you cannot describe the agent’s job in one sentence, it is too big.
2. Open Agent Builder
In the Copilot app, go to Agents in the navigation pane and choose New agent. You can describe what you want in ordinary language and let Copilot build it, choose Skip to configure and fill it in yourself, or start from one of the templates. The natural-language route is genuinely fine to start with. Read back what it wrote for you before you save.
3. Write the instructions as a GCSE block
This step decides whether the agent is any good. Set out the goal, the context it works in, the sources it is allowed to use and the expectations for every answer. Tell it what to do when it does not know. An agent that says “I cannot find that in the handbook, please ask HR” is worth far more than one that guesses well.
4. Point it at the right knowledge
You can name dedicated knowledge sources, including SharePoint content and information from Copilot connectors. Be specific. One tidy SharePoint site beats five sites where two of them hold last year’s version.
5. Add starter prompts
Starter prompts are the example questions people see when they open the agent. They teach your team how to use it and save you from running a training session.
6. Test it properly before you share it
You can test an agent before deploying or sharing it. Ask it the awkward questions. Ask it something it should refuse. Ask it something where you know the underlying document is out of date, and see whether it tells you or covers for you.
7. Share it, then diarise a review
An agent is only ever as current as the documents behind it. Six months of policy drift will quietly turn a helpful agent into a liability.
How to train a Copilot agent so the answers get better
You do not retrain a Copilot agent in the machine learning sense. You improve two things.
The first is the instructions. When an answer comes back wrong, work out which part of GCSE was missing and add it. Most bad answers trace back to a source that was too broad or expectations nobody set.
The second is the knowledge behind it. Agents cannot be better than your documents. If three versions of the expenses policy live in three places, the agent will faithfully hand people the wrong one. Fixing the documents fixes the agent.
Ask the people using it what it got wrong this week. That feedback loop is worth more than any amount of clever prompt writing.
Check your permissions before you create an agent
Agents inherit the permissions of the data behind them. SharePoint agents, for example, automatically inherit SharePoint permissions. That is sensible design, and it is also where this goes wrong.
If a site is overshared, an agent does not cause the problem. It makes the problem efficient. Something technically visible but practically buried under four folders is suddenly one plain question away.
Do not run before you can walk, and do not automate before you are secure. Permissions reviewed, sensitivity labels applied, oversharing cleaned up, then agents. We went through ISO 27001 and ISO 9001 ourselves, and in-house Cyber Essentials assessors are on our technical team, so we tidy the data before pointing anything clever at it.
Agent Builder, SharePoint agents or Copilot Studio?
| SharePoint agent | Agent Builder | Copilot Studio | |
|---|---|---|---|
| Who builds it | Anyone with site access | Anyone in Copilot | Low-code builder or IT |
| Knowledge | One SharePoint site | SharePoint, Teams, Outlook, OneDrive | Business data plus external systems |
| Actions in other systems | No | No | Yes |
| Best for | Questions about one document library | Most first agents | Multi-step workflows and integrations |
If you need advanced capabilities such as actions that reach into external services, that is Copilot Studio territory. It lets you connect agents to business data and publish them across the channels your teams and customers use. For a first agent, start with the simplest option that solves the real problem.
What changes when you get this right
Your team stops spending its week answering questions that already have answers. New starters get a source of truth on day one instead of week six. And the licences you are already paying for become visible to whoever signs the invoice.
The research supports the pattern. Microsoft’s 2026 Work Trend Index found 58% of AI users say they are producing work they could not have completed a year earlier, while 86% treat what AI gives them as a starting point rather than a finished answer. Both of those are true at the same time. The tool widens what your people can do, and your people still own the result.
So, how much of your team’s week goes on questions somebody has already answered?
Copilot agent FAQs
What is the difference between a Copilot prompt and a Copilot agent?
A prompt is a single instruction you type each time. An agent stores that instruction permanently along with its knowledge sources, so everyone gets the same quality of answer without writing the prompt themselves.
Do I need a Copilot licence to use an agent?
Agent building sits inside the Microsoft 365 Copilot experience, and Microsoft has changed how agent usage is licensed and metered more than once. Check your current entitlement before you plan a rollout, because the answer in six months may not be the answer today.
Are Copilot agents free?
Creating a simple agent is part of the Copilot experience rather than a separate purchase, but running costs depend on your licensing and how you use the agent. Treat anyone giving you a flat yes or no with suspicion.
Can Copilot agents send emails?
Not the ones you build in Agent Builder. Those answer questions from the knowledge you give them. Agents that take actions in other systems, including sending things out, are built in Copilot Studio.
Do I need to be technical to create an agent in Copilot?
No. Agent Builder is a no-code tool, and you describe what you want in plain language. The skill that matters is writing clear instructions, which is closer to briefing a new colleague than to programming.
Why is my Copilot agent giving vague answers?
Almost always a missing source or missing expectations. Run the agent’s instructions back through GCSE and check it knows exactly which documents to use and exactly what shape the answer should take. Then check those documents are current.
How many agents should we start with?
One. Pick the question your team is most tired of answering, build for that, and let people use it for a month before you build a second.
Thinking of putting a Copilot strategy together for your business? Feel free to get in touch.