0.1What is Microsoft Copilot?
The essentials in 30 seconds
- Copilot = an AI assistant installed where you work: the same type of language model as ChatGPT, but inside Word, Excel, PowerPoint, Outlook, Teams and Windows — with no copy-pasting to a separate tab.
- Microsoft 365 Copilot knows your context: via the Microsoft Graph, it connects your files, emails and meetings — that access is what turns a generic assistant into your professional copilot.
- A copilot, not a pilot: it drafts the material, you remain the responsible author — and since it can be wrong with confidence (hallucination), every important output must be verified.
Threaded mini-case — Hicham discovers his copilot in Algiers
Hicham, administrative assistant in Algiers, spends his days moving between Word, Excel and Outlook. A colleague mentions “Copilot” and he imagines yet another ChatGPT. He applies the guide's method: he opens Copilot where he already works, asks a first work-related question, refines by iteration instead of starting over, then verifies each precise fact elsewhere. He notes the key difference on a cheat sheet: the free chatbot does not know his files, while Microsoft 365 Copilot sees what he sees — and he remains the author of whatever comes out.
Standalone case for this section: it returns in the lesson, the diagram, the script and the 48h challenge of section 0.1.
1. An AI assistant installed where you work
Microsoft Copilot is the AI assistant built into the Microsoft ecosystem: the same type of language model as ChatGPT, but installed RIGHT where you work — inside Word, Excel, PowerPoint, Outlook, Teams and Windows. You do not have to copy-paste your work into a separate tab: the AI comes to you. The fundamental difference from a general-purpose chatbot: Microsoft 365 Copilot knows your work context (your files, your emails, your meetings) via the Microsoft Graph. It is that secure access to your real data that turns a generic assistant into YOUR professional copilot. Hicham remembers the picture: no more back-and-forth to an external site, the assistant lives inside his tools.
2. A language model, with its limits
Under the hood, Copilot relies on a language model (LLM): an AI trained on huge text corpora, able to generate relevant text — but without “understanding” or guaranteeing accuracy. It can be wrong with confidence: that is called a hallucination, the No. 1 risk of generative AI. The grounding mechanism reduces that risk by anchoring the answer in your real data (files, emails) rather than in the model's memory alone. But it does not eliminate it: every important output must be verified. Hicham sets himself the rule: a figure, a date, a quote — check before sharing.
3. A copilot, not a pilot
The principle to keep in mind throughout the guide: Copilot is a copilot, not a pilot. It proposes the material; you remain the responsible author of whatever comes out. The conversation is iterative: you refine, clarify, ask again — the first answer is a draft, not a deliverable. Copilot does not replace your work: it replaces tasks (first draft, summary, formatting), not your judgement. The measured effect is a shift: less raw production, more review, steering and decision-making — hence the importance of learning to use it well. Hicham displays it above his screen: “Copilot proposes, Hicham disposes.”
Did you know? A copilot never holds the controls alone
Copilot proposes, you dispose: always keep your hands on the controls.
🔎 Concrete example — 🔎 Concrete example — The same question, two worlds
Hicham asks “Where does the leave file stand?” to two versions of Copilot.
Hicham's question: “Where does the leave file stand?”
Free Copilot: “I don't have access to your documents. Can you give me more details?”
Microsoft 365 Copilot: summary cross-referencing emails, tracking file and meeting report, with clickable sources.Reading: the free version does not know Hicham's context; the 365 version sees what he sees — files, emails, meetings — and cites its sources.
Lesson: it is access to YOUR data that makes the difference, not the fluency of the answer.
Section vocabulary
- Copilot
- Microsoft's AI assistant, available in Windows, the web and Microsoft 365 applications.
Simple example: Hicham calls it from Word in Algiers instead of opening a separate tab.
What it is not: a simple external chatbot with no link to his work tools.
Reflex: go to the AI where you work, not the other way round.
- Language model (LLM)
- AI trained on huge text corpora, able to generate relevant text — but without “understanding” or guaranteeing accuracy.
Simple example: Hicham gets a fluent first draft, which he will have to verify.
What it is not: an exact database that never makes mistakes.
Reflex: faced with a precise fact, verify elsewhere before sharing.
- Microsoft Graph
- The map of your work data (emails, files, meetings, contacts): what makes Microsoft 365 Copilot contextual.
Simple example: Hicham receives a summary that cross-references his tracking file and his emails.
What it is not: a shared folder or a web search engine.
Reflex: ask Copilot 365 what it sees in your own data.
- Hallucination
- A wrong answer phrased with confidence — the No. 1 risk of generative AI: every output must be verified.
Simple example: Hicham checks a confidently quoted date before sending it.
What it is not: a simple typo that is easy to spot.
Reflex: the more confident the tone, the more you verify the facts.
- Grounding
- The mechanism that “anchors” the answer in your real data (files, emails) rather than in the model's memory alone.
Simple example: Hicham's answer cites its clickable sources instead of inventing.
What it is not: a guarantee of total accuracy that removes the need to verify.
Reflex: demand sources and open them before validating.
In practice — In practice — First conversation with Copilot
- Open Copilot (Windows app, Edge or the web) and sign in with your account, like Hicham in Algiers.
- Ask a general, testable work question: “Explain the difference between gross margin and net margin, with a numeric example”.
- Refine by chaining: “Redo the example with a restaurant” — notice that Copilot keeps the thread of the conversation.
- Test its limits: ask for a very recent or precise fact and VERIFY the answer elsewhere — critical thinking is the first reflex.
Ready-to-say script — explaining Copilot to a colleague (30 s)
“Copilot is an AI assistant installed in our tools — Word, Excel, Outlook, Teams. The 365 version also knows our context: our files, our emails, our meetings. But beware: the first answer is a draft, the AI can be wrong with confidence. Like Hicham, I refine by iteration and I verify before sharing.”
Say while showing the engine + context + author diagram; stress “copilot, not pilot”.
Copy-paste template — my first testable conversation
Prompt 1 (general work question):
Explain the difference between gross margin and net margin, with a numeric example.
Prompt 2 (iteration — keeps the thread):
Redo the example with a restaurant.
Prompt 3 (limit test — verify elsewhere):
What is the latest official text setting public holidays? Cite your sources.Free version only
Without a 365 licence: run the whole conversation with general questions (definitions, numeric examples) — no internal documents, and systematic verification elsewhere.
Equipped with Microsoft 365 Copilot
Hicham, signed in with his work account in Algiers, adds an anchored prompt: “Where does the leave file stand according to my emails and my tracking file?” — then he opens each cited source.
Personal account
On a personal Microsoft account: same conversation and iteration habits, but only on non-professional content — never company data.
Pair in Algiers
Hicham dictates the prompt, his colleague reads the answer aloud: together, you spot claims to verify more easily — each notes one fact to check.
✓ Best practices
- Treat the first answer as a draft: refine with successive instructions rather than starting over.
- Always specify the context of your request: the more Copilot knows, the better the answer.
- Keep the verification reflex for every precise fact, figure or quote — especially outside your grounded data.
✗ Common pitfalls
- Taking the first answer at face value without verifying or refining it.
- Confusing Microsoft 365 Copilot (contextual) with the free chatbot (general-purpose): they do not have access to the same things.
- Believing that an AI answering “with confidence” must be right — hallucination is fluent and convincing.
Key takeaways
- Copilot = an LLM built into your Microsoft tools, with access (in the 365 version) to YOUR work context.
- The conversation is iterative: refine, clarify, ask again — the first answer is a draft.
- Generative AI can hallucinate: every important output must be verified.
- You remain the responsible author: Copilot proposes, you validate.
48h challenge — apply tomorrow
Run the conversation template like Hicham (general question + iteration + limit test), note what Copilot retains of the thread, and verify elsewhere one precise fact from its answer.
- I can explain what distinguishes Microsoft 365 Copilot from a general-purpose chatbot.
- I have run a conversation with iteration and seen that the thread is kept.
- I have verified elsewhere a precise fact instead of taking it at face value.
Frequently asked questions
Copilot, ChatGPT, Claude, Gemini: which to choose?
For work in the Microsoft ecosystem, Microsoft 365 Copilot has a decisive advantage: secure access to your files/emails/meetings. For general tasks, all four are broadly comparable — integration makes the difference.
Will Copilot replace my work?
It replaces TASKS (first draft, summary, formatting), not judgement. Studies mainly show a shift: less raw production, more review, steering and decision-making — hence the importance of this guide.
What distinguishes Microsoft 365 Copilot from a general-purpose chatbot?