Copilot Explained: Which Microsoft Copilot Is Right for You?

ARTIFICIAL INTELLIGENCE

Copilot is a family name, not one magic app

The word has become a little too convenient. It appears in a browser, a Windows laptop description, Microsoft 365, GitHub, security tools and business software—and those products do not have the same job, the same data access or even the same buyer. One name, several cabins on the boat.

At the simplest level, Microsoft Copilot is a general-purpose conversational assistant. Microsoft 365 Copilot brings that style of assistance into work apps and, where the account permits it, into work context. GitHub Copilot is built for software development. Copilot+ PC describes a class of Windows hardware with an on-device AI capability; it is not a second chat subscription hiding inside your laptop.

That map does most of the heavy lifting. A person writing a proposal, a student learning Python and a developer fixing a production bug may all say “I use Copilot”, while describing three different products. Buying the wrong one is not a moral failure; it is just an expensive way to discover that labels are not explanations.

ARTIFICIAL INTELLIGENCE

Microsoft Copilot: the everyday starting point

For a normal user, this is the Copilot most people mean. You can ask it to explain a concept, compare options, outline a presentation, rewrite a paragraph, turn notes into a checklist or help you think through a problem. The exchange feels like a conversation, but the useful unit is still the task: a clear question, enough context and a result you will actually inspect.

It can also be useful for small, unglamorous jobs. “Turn this meeting note into three decisions and five follow-ups” is better than “Be creative.” “Rewrite this message for a Hong Kong customer; keep the dates and do not invent a discount” gives the assistant somewhere to stand. The computer does not need another motivational speech.

1. Start with the outcomeSay what you need at the end: a comparison table, a short email, a study explanation, a list of risks or a draft agenda.
2. Give it the materialPaste the relevant text or attach the file where the current product and account support it. Do not assume it knows a document you have not supplied.
3. Name the audienceA manager, a secondary-school student and a customer in Osaka need different levels of detail and different vocabulary.
4. Set boundariesAsk it not to invent numbers, to flag uncertainty, to preserve names and dates, or to separate facts from suggestions.
5. Read the answer as a draftA fluent answer can still be wrong, out of date or slightly off in a way that matters. Your eyes remain part of the system.

The free consumer experience and paid Microsoft 365 experiences should not be treated as identical simply because the chat box looks similar. Microsoft describes Copilot as an AI assistant that uses large language models, while the exact tools, limits and access depend on the product, account and subscription attached to it. Check the current Microsoft account page rather than relying on a screenshot from last year.

For a student or a working professional who mainly wants explanations, drafts and idea-shaping, this general version is often the sensible place to begin. Try the actual tasks first. A label on a plan page cannot tell you whether the assistant fits the way you work.

ARTIFICIAL INTELLIGENCE

Microsoft 365 Copilot: where the work context changes everything

This is the version people often imagine when they hear “Copilot at work”. In Word it can help shape or rewrite a document; in Excel it can help explore data; in PowerPoint it can help turn a brief into a presentation; in Outlook and Teams it can help with messages, meetings and follow-up work. The exact buttons and capabilities change as Microsoft ships updates, so treat the app names as a map, not a frozen contract.

The important distinction is the data boundary. Microsoft’s current architecture documentation says Copilot can use information in the user’s Microsoft 365 context—such as email, chats and documents—but only where that user already has permission. It does not grant a person a new right to read a manager’s private folder. If your organisation’s permissions are a mess, the assistant does not make the mess disappear; it may simply make the consequences easier to notice.

There are also two phrases that are easy to muddle: Microsoft 365 Copilot and Copilot Chat. Availability depends on the account type, licence and administrator settings. Some organisations may provide chat access without every in-app capability; others may restrict features while they work through data governance. Ask your administrator what is enabled, and do not infer access from a colleague’s screen.

For a company, the preparation is not only “buy the licence”. Clean up permissions, agree what information can be used, decide who reviews generated documents and write a short internal rule for confidential material. The boring work is the useful work here.

ARTIFICIAL INTELLIGENCE

GitHub Copilot: an assistant for software, not a junior engineer in a box

GitHub describes its Copilot as an AI assistant that can suggest code as you type, answer questions about a codebase, review changes and work on tasks assigned to it. That makes it a different kind of tool: the surrounding context is a repository, a file, a test, an issue or a pull request—not a general “tell me something interesting” chat.

For a new developer, the explanation feature can be more valuable than the autocomplete. Ask what a function assumes, where a value comes from, why a test fails or which file appears to own a behaviour. For an experienced developer, the time savings may show up in scaffolding, repetitive transformations, test cases, documentation and the first pass through unfamiliar code.

GitHub’s current documentation also describes agent-style features that can research a repository, make changes and prepare a pull request for review. “For review” is doing important work in that sentence. A branch and a pull request are useful guardrails; they are not proof that the code is correct, secure or welcome in production.

Students should not use it as a way to skip learning the language they are meant to learn. Ask it to explain a solution, offer a smaller hint, create test cases or compare two approaches. If it writes the whole assignment and you cannot explain the result, the short-term win has quietly become a long-term bill.

ARTIFICIAL INTELLIGENCE

Copilot+ PC: a hardware label, not another subscription

This is where product naming gets cheeky. Copilot+ PC appears in laptop specifications, while Microsoft Copilot appears as software. The “plus” is not a premium chat tier. It describes hardware designed to handle certain AI workloads on the device, alongside the usual Windows and cloud capabilities.

Microsoft’s original 2024 announcement defined Copilot+ PCs around neural processing performance of more than 40 trillion operations per second and highlighted features such as Recall, Cocreator and Live Captions. Since then, device models, Windows versions, regional availability and feature requirements have moved on. Use the current specifications for the exact model you are considering; do not buy a laptop on a two-year-old feature list.

Why does local processing matter? It can affect responsiveness, battery use, privacy design and which experiences are possible without sending every step to a remote service. But “runs partly on the device” is not the same as “everything stays on the device”. Read the feature’s current documentation and the manufacturer’s privacy information.

In short: if you need a chat assistant, look at the software and account. If you need a new laptop, evaluate the whole machine. The plus sign is helpful information, but it is not a shopping personality.

ARTIFICIAL INTELLIGENCE

There are specialist Copilots too—choose by workflow

A security analyst may need an assistant connected to alerts, incidents, threat intelligence and investigation tools. A sales or service team may need help inside customer records, cases and business processes. That is why specialist Copilots exist: the useful answer depends on the underlying system, not just on the language model’s ability to write a paragraph.

1. Name the system firstAre you working in Microsoft 365, GitHub, a security platform, Dynamics 365 or simply a blank chat? The system determines what context is available.
2. Name the roleA student, developer, finance analyst and security responder have different approval duties and different costs for a wrong answer.
3. Name the actionDo you need a draft, a diagnosis, a code change, a customer summary or an incident response suggestion? “AI” is not an action.
4. Name the review gateDecide who must approve the result before it reaches a customer, a production server, a regulator or an internal decision.

For an individual, this usually means you do not need to study every product name. If your work does not happen in a security console or a customer-management system, those specialist offerings are probably not your starting point. If you are responsible for buying them, begin with the workflow and the information boundary, then compare the product.

Marketing pages naturally make the products sound close. The practical question is less glamorous: where does the assistant sit, what can it read, what can it change, and who checks the result? Ask those four questions and a lot of fog clears.

ARTIFICIAL INTELLIGENCE

One ordinary workday, four different answers

Notice what is not happening: one universal assistant is not swallowing the company whole. Each tool has a lane. That separation is healthier than treating every “Copilot” button as a portal to the same invisible brain.

Students can use the same mental model. A general assistant may explain a topic, a work-context assistant may help organise a group project’s documents, GitHub Copilot may help with code, and a Copilot+ PC may offer device features. The product is secondary to the learning task. If the goal is understanding, ask for reasoning, examples and questions—not just the final answer.

ARTIFICIAL INTELLIGENCE

A prompt that works is usually less clever than you think

People sometimes blame the assistant for a vague request, then write an even vaguer follow-up. “Make this better” can mean shorter, friendlier, more persuasive, more accurate, more local or less embarrassing. Pick one. Or say all six, in plain language.

1. Role and situation“You are helping a Hong Kong project manager prepare a client update.” This gives the response a point of view without pretending the machine is a person.
2. Material and sourcePaste the text, attach the workbook or point it to the authorised work context. Say what is missing instead of inviting it to guess.
3. Output shapeAsk for a 120-word email, a two-column comparison, a six-step checklist, a code diff or three study questions.
4. BoundariesKeep the original numbers, do not invent sources, use Traditional Chinese, flag assumptions and leave unknown fields blank.
5. Review questionsEnd with “What should I verify?” or “Which parts rely on an assumption?” It nudges the interaction away from polished nonsense.

Three prompts to try

For a general assistant: “Explain this lease clause in plain Cantonese-style Traditional Chinese. Give me the practical meaning, then list the points I should ask a lawyer about. Do not tell me it is safe just because it sounds normal.”

For Microsoft 365 Copilot: “Using the project notes and emails I can access, draft a one-page status update with completed work, open decisions, owners and dates. Separate confirmed facts from items that need checking.”

For GitHub Copilot: “Before editing, explain how authentication currently flows through these files, identify the smallest safe change and propose tests. If the repository context is incomplete, say exactly what you cannot see.”

The phrasing is not magic. It just puts the work on the table. A useful prompt is a small brief, not a spell.

ARTIFICIAL INTELLIGENCE

Privacy, permissions and the bit people skip

For Microsoft 365 work, the permission model matters. Microsoft says Copilot accesses content the individual user is authorised to access and does not grant access to material that user cannot already open. That is reassuring, but it does not excuse sloppy permissions. If too many people can read a folder, an assistant that helps them find it does not create the underlying problem.

Do not let the word “Copilot” flatten those distinctions. A consumer chat, a company-controlled Microsoft 365 account and a repository assistant are different operating environments. Read the current privacy and admin documentation for the product you actually use, not the one with the most memorable launch video.

For teams, make a one-page rule: approved tools, prohibited data, required review, retention questions and a person to contact when something looks wrong. It does not need to be a 47-page policy nobody opens. It needs to be clear enough that someone can use it on a busy Tuesday afternoon.

ARTIFICIAL INTELLIGENCE

The limitations are ordinary—and still important

1. It may invent or blur factsAsk for sources, open the sources and compare the wording with the original. This matters especially for law, medicine, finance, news and company claims.
2. It may miss contextA summary can leave out a caveat, a sarcastic remark, a missing attachment or the one sentence that changed the decision.
3. It may write code that passes one testA green test is not a security audit, a performance review or proof that the design belongs in production.
4. It may struggle with messy dataCheck date formats, hidden rows, formulas, duplicated records, units and filters before trusting a spreadsheet explanation.
5. Its product behaviour changesRollouts, models, quotas, menus and regional access can change. A tutorial is a snapshot, not a promise carved in stone.

A good review is specific. Instead of asking “Is this correct?”, ask “Which dates came from the document?”, “Show me the rows used for that calculation”, “What assumption creates this conclusion?” or “What would make this code fail?” You are giving the answer somewhere to be tested.

There is a human temptation to admire the fluent paragraph and stop reading. Resist it. The smooth bit is often the least informative bit.

ARTIFICIAL INTELLIGENCE

How to choose: let the task pick the product

For a working professional, the best first test is not “write me a book”. It is one recurring task you understand well enough to judge: weekly status notes, a meeting follow-up, an Excel explanation, a bug triage summary or a code review checklist. Measure minutes saved and corrections required. A tool that saves ten minutes but creates a 30-minute fact-checking chore has not saved time; it has moved the invoice.

For a student, choose the tool that helps you explain the material back. Ask for hints, a simpler example, a counterexample or a mini-quiz. If you cannot answer a similar question without the assistant, you have borrowed the result rather than learned the skill.

For a small business, start with permissions and a narrow workflow. One well-defined use case, a clear reviewer and a short feedback loop beats a company-wide announcement full of shiny nouns.

ARTIFICIAL INTELLIGENCE

A practical first week with Copilot

1. Day 1 — name the jobChoose one repeated task and write down the current steps, the time it takes and the mistakes that would be costly.
2. Day 2 — use a general promptAsk Microsoft Copilot for a draft or explanation. Keep the task low-stakes and record what you had to correct.
3. Day 3 — try the native contextIf you work in Microsoft 365 or GitHub, try the matching assistant with a file, meeting, issue or repository task you are allowed to use.
4. Day 4 — stress-test the answerChange an input, add a missing constraint and ask what the assistant cannot know. Look for confident errors.
5. Day 5 — set a review ruleWrite the two or three checks you will always make before sending, merging, publishing or submitting the result.
6. Day 6 — compare the timeCount the first-draft time and the checking time separately. Do not call a task faster until both numbers are honest.
7. Day 7 — decide or stopKeep the workflow if it genuinely helps. Change product if the context is wrong. Stop using it if the review cost outweighs the benefit.

This approach sounds almost disappointingly sensible. That is why it works. You are testing a work habit, not auditioning a chatbot for a science-fiction role.

Start with the version your account already offers, learn where the data comes from, and only then consider a paid tier, a company deployment or a new machine. Copilot is most useful when it is close to the work—and when a real person remains close to the answer.