Explainer
How does Copilot work? Without the jargon.
Copilot is not a search engine and not a chatbot with internet access. It's a language model that first searches your files and only then writes an answer. That difference explains almost everything: what it does well, and where it goes wrong.
What happens when you ask something
01 · The mechanics
There are three steps between your question and the answer on your screen. Once you know them, it's immediately clear why some questions work and others don't.
Step one: your question goes to the Microsoft 365 search layer. It looks through the files, mail and conversations you have access to and pulls out the passages that probably matter. This is plain search, with the same permissions you already had, no more and no less.
Step two: those retrieved passages go to the language model together with your question. So the model doesn't see your whole SharePoint, but a selection. If the search layer misses the right document, it doesn't exist as far as the model is concerned, and Copilot won't say 'I couldn't find it', it will answer based on what it did have.
Step three: the model writes the answer, predicting word by word what probably comes next. That's why the tone is always equally confident: confidence is a property of the writing style, not of the content. An answer with citations can be checked; an answer without them is a well-written guess.
From those three steps follows the single most practical rule: your question has to help the search layer. 'Summarise the project plan' works worse than 'summarise the March project plan for Noordwijk'. Not because the model is dumber for the first question, but because the search layer hands it the wrong passages.
What this means
Copilot rewards those who give context and punishes those who keep it brief. That's learnable, and it's exactly what the second module covers: task, context, form.
Where Copilot lives
02 · The places
Teams
Meetings and conversations
Summarising a meeting you missed, extracting action points, or reducing a long thread to what was decided. Does require the recording or transcript to have been on.
Outlook
Mail and calendar
Reducing a full inbox to what needs attention, or drafting a reply based on the thread underneath. Rewriting your own draft works better than starting from nothing.
Word
Writing and rewriting
A first draft based on existing documents, or making existing text shorter, more formal or more accessible. Pointing at your own sources is the key here.
Excel
Analysis without formula knowledge
Asking questions of a table in plain language, and generating formulas you can then read back. Only works on ranges formatted as actual tables.
PowerPoint
From document to story
Turning an existing document into slides, or shortening a deck. The structure usually comes out well; the design always needs manual work.
Copilot Chat
The standalone assistant
Asking questions about your whole work environment outside any specific file. This is where the gap between a vague and a precise question is largest.
What Copilot doesn't do
03 · The limits
It decides nothing
Copilot delivers a draft, not a decision. The moment you forward the output unread, you made the error, not the model. That sounds harsh and it is the entire point.
It doesn't remember you
Outside a single conversation, Copilot knows nothing about your preferences, your tone of voice or your previous task. Whatever you need every time belongs in a block of text you reuse.
It doesn't tidy your data
If three versions of the same policy document coexist, Copilot picks one, and not necessarily the valid one. Findability is a precondition, not an outcome.
Questions about how Copilot works
04 · Frequently asked
What exactly is Microsoft 365 Copilot?
An AI assistant that sits inside your Office apps and can reach your own files, mail and calendar. That last part is the difference from a standalone chatbot: Copilot can base an answer on documents you have access to, rather than only on what the model once learned.
Where does Copilot get its answers?
From two sources. The language model supplies the phrasing and general knowledge; the Microsoft 365 search layer supplies the facts from your environment. For a question about a document it first retrieves the relevant passages and only then writes the answer.
Can Copilot reach files I'm not allowed to see?
No. Copilot inherits your permissions: what you can't open, it won't use. The reverse is a genuine risk, though. If a folder is open to everyone by mistake, Copilot makes that painfully visible, it finds what you never would have.
Why does Copilot sometimes make things up?
Because the model predicts what probably comes next rather than looking up what is true. If the search layer finds nothing usable, the model fills the gap with something plausible. The hard part is that the tone doesn't change: an invented answer sounds just as confident as a correct one.
Does Copilot learn from our company data?
No. Your data isn't used to train the underlying models; it stays within the boundaries of your Microsoft 365 tenant. What Copilot does do is retrieve it per question to base an answer on. That is different from remembering.
What do I need to get started?
A Microsoft 365 licence with Copilot added, and data that's in order. The second is underestimated: Copilot is only as good as the findability of your documents. A messy SharePoint produces messy answers, and no training fixes that.
Read on
- From rollout to real use
Why Copilot rollouts rarely fail on technology.
- Copilot in Dutch
What works and what doesn't once you work in Dutch.
- Copilot or ChatGPT?
The difference is context and governance, not who writes better.
- Prompts that actually work
Five patterns, with what goes wrong when you drop each one.
- The eight modules
Exactly what's in the training, per module and per minute.
Knowing how it works is step one.
Step two is doing it every day. That's the gap most licences disappear into, and it's exactly what we measure before training anything.