You don’t need a Copilot agent to fill a Word document from data in another file. Copilot in Word can reference an Excel workbook with a forward slash, fill the table in the open document and rewrite the text around it, with nothing to build or maintain. An agent earns its place when the same job has to run for other people, run without you, or follow the same instructions and knowledge every time.
Everybody wants an agent at the moment, and a request I get all the time is “Daniel, can Copilot or an agent fill out this Word doc for me? I’ve got the data over here, I just want it to fill out a table.” For that request, I’d skip the agent and go to the place where the work is happening, which is the Word document.
Copilot in Word vs a Copilot agent at a glance
| Copilot in Word | Copilot agent (Agent Builder) | |
|---|---|---|
| Where it works | Inside the document you have open | In Copilot Chat, wherever people @mention or open the agent |
| Setup | None | Name, description, instructions and knowledge to configure |
| Instructions | Your prompt, each time | Saved in the agent, the same for everyone who uses it |
| Data from other files | Reference a file, email or meeting with / | Knowledge sources set on the agent, plus anything you reference in the chat |
| Changes the document | Yes, edits the open file directly, or Chat only | Answers in the conversation unless you build document output into it |
| Who uses it | You, in this file | Anyone you share the agent with |
| Maintenance | Nothing to maintain | Instructions and knowledge need looking after as things change |
| Best for | One-off and occasional work in a file you’re already in | A repeatable job done the same way by several people |
The difference that decides it: where the work happens
In my demo, the work was a Word report with a results table full of “to populate” placeholders, and the figures sat in an Excel workbook spread across several worksheets. A forward slash in the Copilot pane let me reference the workbook, and one prompt asked Copilot to replace the “to populate” entries from it. I ran it on Claude Opus 5.5 from the model selector, which in Word offers Auto, reasoning levels and models from OpenAI and Claude.
Copilot read the workbook and filled the six result cells, editing the document live. A second prompt had it rewrite the executive summary from the new figures, and it replaced the whole summary. Start to finish took under three minutes, recorded with no cuts or time lapse, and left nothing behind for anyone to maintain.
An agent would have meant naming it, writing instructions, connecting the workbook as knowledge and then getting its answer back into the report. For a one-off report, I’d be spending more time on the agent than on the report.
Editing the file vs answering in chat
Copilot in Word opens with Edit mode on, so it changes the document itself. At the bottom of the pane you can switch to Chat only when you want answers without changes, which I’d use for checking a document before anyone edits it. Every change Copilot makes in Edit mode is reviewable and reversible, and in a shared document it shows a preview of its suggested changes in the chat first, so nothing lands in front of collaborators until you approve it.
An agent built in Agent Builder works from the conversation. It can create documents if you set it up to, which I showed with an agent that writes Word documents on your branded template, but that’s a separate build with its own template and instructions.
Repeatable for other people vs a prompt for you
The prompt I used in Word lives in my head. If a colleague needs the same report next quarter, they write their own, and they might phrase it differently or forget the summary. An agent saves the instructions and knowledge once and gives everyone the same starting point, with conversation starters for people who aren’t sure what to ask. Agents can now carry skills as well, one for each task, which I cover in adding skills to Copilot agents in Agent Builder.
You pay for that consistency in setup and maintenance. When the workbook moves, the template changes or the process gets a new step, someone has to update the agent.
The four questions I ask before building an agent
My default is Copilot in the app, and I only build an agent when the answers push me there:
- Is the work happening in a file I already have open? If yes, start with Copilot in that app: Word, Excel, PowerPoint or Outlook.
- Is the information I need sitting in another file I can point Copilot at? If yes, reference it with / and you probably don’t need an agent’s knowledge sources.
- Does this need to run without me, or for other people? If it needs to run for a team, build an agent and share it. If it needs to run with nobody there at all, on a schedule or a trigger, that’s beyond an Agent Builder agent: look at Copilot Studio, or Copilot Cowork for scheduled prompts.
- Will I do it often enough to justify building and looking after something? Monthly reporting for a whole team, yes. A one-off report, no.
If you answered yes to the first two and no to the last two, open Copilot in the app. Copilot has been in Word, Excel, PowerPoint and Outlook for ages, and plenty of people go straight past it looking for an agent. Pulling in your other work files relies on the work data that comes with a Microsoft 365 Copilot licence, so check that before rolling this approach out.
Frequently asked questions
How do I reference an Excel workbook from Copilot in Word? Type / in the Copilot pane and start typing the file name, then pick it from the list. The same menu lets you reference emails and meetings, and you can select items from it without typing.
Can I stop Copilot in Word from changing my document? Yes. Select Chat only in the Copilot pane and it responds in the chat without editing the file. Edit mode is the default, and changes made in it can be reviewed and reversed.
When is a Copilot agent the better choice? When several people need to do the same job the same way, with the same instructions and knowledge each time. Each Agent Builder agent can hold up to eight skills, so one agent can cover several related tasks for a role.
